Sierra: The AI Company Paid Only When the Job Gets Done

An episode of Dan's AI Intel

Sierra is paid only when its AI agent finishes the job, and that one change is moving software money into the budget that pays people.

Published · By Dan Walter

Executive summary

Sierra does not sell software in the way the last twenty years taught us to buy it. It sells finished work — a customer's problem solved, a cancellation saved, a loan application completed — and it is paid a pre-negotiated rate only when that work is done. When its agent has to hand a customer to a person, Sierra charges nothing. That one change moves the money out of the IT budget and into the budget that pays people, and it is why a company founded in February 2023 reported more than $400 million in annual recurring revenue in October 2026, four times the $100 million it announced eleven months earlier.

The deeper story is what this does to everyone else. A software company paid per seat earns less every time its own AI works, because a better agent means fewer human users. HubSpot showed the cost of escaping that trap in May 2026: it moved its AI agents to pay-per-resolution and its shares closed down 19% the next day. Across software, outcome pricing is spreading quickly as one part of hybrid deals, but as a company's main model it is still rare — roughly one company in twenty. It wins first where an outcome can be counted, and customer service is the place where it can.

Sierra's lessons for building an AI-native company are concrete: narrow the job before adding AI, give agents goals and guardrails rather than scripts, treat evaluation as a product, and expect to throw code away as models improve. There is no public record of Sierra rolling its automation back. The companies that went too far and retreated were Klarna and Commonwealth Bank of Australia — and Sierra's free hand-off to a human is the pricing term built to keep it from following them.

Context: why a customer-service company is a window on the AI revolution

Customer service is where AI agents first stopped being demos and started doing paid work at scale. Bret Taylor, Sierra's chief executive, puts annual spending on customer service at roughly $400 billion. Whoever automates that work, and however they charge for it, sets the template for every other white-collar function agents reach next.

Sierra has appeared on this show twice before, both times as a supporting character. In our benchmark-trap episode, number 42, released in August, Sierra's τ-bench was the test that exposed how reliability collapses when an agent must get the same task right several times in a row. In our real-time AI episode, number 69, released last month, Sierra was one of the phone-agent companies whose evidence was vendor case studies rather than controlled trials. Both points still stand. What has changed is the scale: Sierra has doubled its revenue since May, extended its benchmarks to voice, and moved from answering questions to chasing outcomes that take weeks.

The stakes are bigger than one company. Between 12 January and 23 February 2026 the S&P Software & Services Index lost 25% of its value, as investors priced in agents that act for people instead of people clicking through software. Sierra is the clearest working example of what replaces the seat.

Sierra sells finished work, and only bills when it is finished

Sierra's business model fits in one sentence: it charges a pre-negotiated rate for each completed outcome, charges nothing when the agent escalates to a person, and absorbs the cost of the AI models itself. Taylor described it plainly in March 2026: "if the AI agent resolves the case, no human intervention, there's a pre-negotiated rate for that. If we do have to escalate to a person, that's free." On model costs he was blunter: "reducing your token utilization for the same outcomes is your problem, not your customer's."

The outcomes are defined case by case with each customer before launch. Sierra's own pricing note, published in December 2024, lists a resolved support conversation, a saved cancellation, an upsell and a cross-sell. It also concedes the model "may not always be the best fit": a greeter that only routes calls is billed per conversation instead. Sierra does not publish its rates; third-party estimates of one to two and a half dollars per resolution have never been confirmed by the company.

The same success has opposite effects on the two kinds of vendor. When a seat vendor's agent solves a problem, its customer needs fewer human agents and, at renewal, fewer seats: the better its AI, the smaller its bill. When Sierra's agent solves the problem, that resolution is the bill. Between the two sit per-conversation and credit pricing, which track activity but still charge the buyer for failures; Sierra's 2024 note calls unused seats "shelfware". Under per-resolution pricing the vendor eats every failed attempt and every wasted token.

The product is what makes that bargain affordable. Sierra builds a branded agent that does a job end to end — refreshing a car radio, refinancing a mortgage, filing a claim — across chat, voice and email in more than 30 languages, and that product changes how the industry is seen: a company no longer buys software for its staff to use, it buys the finished work, priced per result. Underneath sits what Sierra calls Agent OS: a "constellation of models", in Taylor's words, from OpenAI and Anthropic plus Sierra's own fine-tuned layers. Agents are given "goals and guardrails, not the sequence of steps", and supervisor models send back any decision that breaks policy. The shift in how people see the industry is this: the buyer is no longer purchasing a tool for its staff. It is purchasing the work.

From building couches to $400 million in revenue

Sierra started small and hands-on. When it emerged in February 2024 with $110 million from Sequoia and Benchmark, it had about 30 staff and early customers including WeightWatchers, SiriusXM, Sonos and OluKai. Its pitch was that big companies did not want a kit of AI parts. "Most people just want to buy a couch," Clay Bavor, its co-founder, told Fortune: Sierra would build the agent for them.

The money followed the traction. In October 2024 Greenoaks led a $175 million round at a $4.5 billion valuation, when Sierra had passed about $20 million in annualised revenue, according to people familiar with it who spoke to Reuters. In September 2025 Greenoaks led again: $350 million at $10 billion. In May 2026 Tiger Global and Google's venture arm GV led $950 million at a $15.8 billion post-money valuation.

Exhibit — Sierra's recurring revenue has doubled roughly every 5½ months for two years. Source: Reuters via SiliconANGLE (Oct 2024, people familiar); TechCrunch (21 Nov 2025); Sierra blog 'Year two in review' (6 Feb 2026) and growth post (30 Oct 2026). Compiled by Dan's AI Intel.

Sierra reported $100 million of annual recurring revenue on 21 November 2025, seven quarters after launch. It passed $150 million in February 2026, took six months to go from $100 million to $200 million, and then five months to go from $200 million to more than $400 million — adding over $150 million in the third quarter alone. Counting from the roughly $20 million it had in October 2024, that is a doubling about every five and a half months for two years, at a pace that has barely varied. These are company figures, not audited revenue.

Investors have paid less for each dollar of that revenue as it grew. The October 2024 round valued Sierra at about 225 times its annualised revenue; the May 2026 round, at $15.8 billion on roughly $200 million, at about 79 times. Revenue rose about tenfold between the two rounds while the valuation rose 3.5-fold: the business has outgrown even its investors' enthusiasm.

Exhibit — Investors paid about 225 times Sierra's revenue in 2024 and about 79 times in 2026. Source: SiliconANGLE (28 Oct 2024: $4.5B round; revenue ~$20M per people familiar); CNBC (4 May 2026: $15.8B post-money); Sierra blog (revenue ~$200M). Compiled by Dan's AI Intel.

The customer base moved up-market as it grew. By February 2026 one in four Sierra customers had more than $10 billion in annual revenue and half had more than $1 billion, and Sierra said agents built on its platform reached more than 95% of US shoppers and half of US healthcare families. It opened in Tokyo in December 2025, with SoftBank's Vision Fund 2 investing, and announced an expansion into France the same month. Deployment times kept falling: a large healthcare company went live seven weeks after kick-off, which Sierra says would have been six without the winter holidays.

Where it has landed is a different business from the one it launched. Sierra now works with almost half of the Fortune 50 and, by its own count, one in three of the world's largest banks. Its newest agents no longer just answer questions: they originate loans, chase prior authorisations for medical procedures, and take payments inside the call. In October 2026 a US retailer with more than $40 billion in revenue went from first meeting to a live agent in 27 days. The couch-builder has become a factory.

Two Google product managers who had seen the seat trap from the inside

Bret Taylor and Clay Bavor both began as associate product managers at Google in its early years, and met there around 2005. Their careers since then explain why Sierra looks the way it does.

Taylor joined Google in 2003 and led the local-search work that became Google Maps, which he is credited with co-creating. He left in 2007 to co-found FriendFeed, which Facebook bought in 2009 for an estimated $50 million; he was Facebook's chief technology officer from 2010 to 2012. He then founded Quip, a document editor that Salesforce bought in 2016. At Salesforce he rose to chief product officer, then president and chief operating officer, led the purchase of Slack, and served as co-chief executive from November 2021 until January 2023. Along the way he chaired Twitter's board through Elon Musk's takeover in 2022. Since November 2023 he has chaired OpenAI's board.

That résumé matters for one reason above the rest. Salesforce is the largest company built on selling software per user, and Taylor ran it. Sierra's founding argument — that a vendor paid per seat earns less every time its AI succeeds — comes from someone who has sat on the other side of that conflict.

Bavor spent 18 years at Google, from 2005 to early 2023. He led product and design for Gmail, Docs and Drive, then started Google's virtual- and augmented-reality effort and Project Starline, a life-size video-calling system, and in 2021 became head of Google Labs, reporting to Sundar Pichai. His career has been about making strange new technology feel natural to ordinary people, which is the core design problem of an agent that answers a customer's phone call.

The company began at lunch. Shortly after Taylor left Salesforce, Bavor invited him to a Mediterranean restaurant in Palo Alto; Taylor assumed his old colleague was checking on him. They talked for hours and left having decided to found a company, named after the Sierra mountains to signal ambitions beyond Silicon Valley. The calibre of that pairing — and Taylor's seat at the head of OpenAI's board — is a large part of why four rounds of investors paid up so quickly.

Customers praise the numbers; reviewers question the dependence

What clients say, in public, is overwhelmingly positive, and almost all of it comes from Sierra's own customer pages. Airtable reports that Sierra's agent resolves 80% of its support conversations, doing the equivalent work of about 95 full-time agents, and that it went live on chat within four weeks. "Our customers don't just use the platform. They build mission-critical operations on it," says Amanda Harris, Airtable's vice president of customer support. Chime reports a resolution rate above 70%; SoFi reports a 33-point rise in its net promoter score. Those rates sit in the same band rivals report for themselves: Fin says its agents resolve on average 76% of support volume, and HubSpot's chief executive says its agent succeeds about 70% of the time.

The cluster is the point: across vendors, roughly three conversations in four now end without a person, and the companies that publish the figure are the ones with a reason to. Rocket Mortgage says clients who use both its AI chat and a human banker convert four times as often on refinancing, and clients who start with its digital assistant close at three times the rate of those who do not. Rocket's agent now holds more than 400,000 successful chat conversations and places more than a million outbound calls a month.

SiriusXM, whose 34 million subscribers talk to an agent called Harmony, replaced a keyword chatbot that led to dead ends. "Sierra is a game-changer for meeting a subscriber's current needs," says Moshe Pridan, its chief product and information officer. Investors are louder still. "Sierra is by all measures the winner in the 'customer experience' category," Peter Fenton of Benchmark, an early backer, told CNBC in May.

Three caveats belong beside those quotes. None of the figures is independently audited; they are vendor-published results from customers willing to be named. Review sites and rival vendors report recurring complaints that customers must go back to Sierra's engineers to change an agent, and that costs are hard to predict. And Sierra's approach depends heavily on its own forward-deployed engineers working inside each customer, which is expensive and is precisely what its newer self-build tools are meant to replace.

Seats are leaking fastest where the outcome can be counted

The move away from seat pricing is real, but it is happening first, and fastest, in customer service, because there an outcome is easy to count: either the customer's problem was solved without a person or it was not. Zendesk, by its own account the first in its industry, began charging per automated resolution in August 2024; its rate beyond a plan's allowance is $1.50 per resolution on committed volume and $2 pay-as-you-go. Intercom's Fin charges $0.99 per resolution. HubSpot switched its customer agent on 14 April 2026 from about $1 per conversation to 50 cents per resolved conversation, counting a resolution only if no human takes the case over within 72 hours. Salesforce announced a pay-per-resolution Help Agent at $2 per resolution in June 2026, according to pricing trackers. Within two years, every major help-desk vendor was selling AI on the same basis as Sierra.

The why is arithmetic. A seat counts a person; an agent that works means fewer people. Satya Nadella has described seats as "just entitlement to some consumption", and Sierra's 2024 note told buyers to ask their legacy vendor how much its seat bill would shrink. Under outcome pricing the seller's revenue rises when the agent succeeds, not when headcount does.

Two things made the switch possible now rather than a decade ago. Agents can finally finish a task without a person, so there is a discrete result to bill for; a chatbot that only deflected questions to a help page never produced one. And the cost of finishing that task has become small enough for a vendor to absorb. Taylor compares the shift to online advertising's move from paying per impression to paying per click: the closer the price sits to the value the buyer receives, the more efficient the market. Usage pricing, he argues, is a weak proxy, because the tokens a model burns correlate only loosely with whether the customer was helped.

How quickly depends on what you count. ICONIQ's surveys of AI builders found the share using some form of outcome-based pricing rose from 18% to 23% between January and July 2026, and usage-based pricing from 35% to 42%; the average company combines 1.7 models. Kyle Poyar's survey of more than 230 software companies, run in April and May 2026, found hybrid pricing — a platform fee plus usage or credits — rose from 25% to 37% of companies in a year. But pure outcome pricing as the main model remains rare: 5% of AI-monetising software companies in High Alpha's benchmark, and 3.8% of 80 agent companies in Orb's 2026 study. The largest vendors hold on to seats hardest; companies above $150 million in revenue are the most likely to keep seat pricing, at 29%.

So the answer to how fast is: very fast as an ingredient, slowly as the whole recipe. Outcome pricing is being added on top of seats and credits far faster than it is replacing them, and it spreads furthest where results are countable — resolutions, qualified leads, payments collected.

Outcome pricing moves software into the labour budget, and the fight to the definition

Outcome pricing means four things for the industry's future. First, software starts to be paid out of the budget for people. In Poyar's survey, 70% of companies say their AI revenue comes out of existing software budgets, but AI-native companies are paid more often from services budgets (35%) and headcount budgets (15%). Gartner estimated in July 2026 that up to $234 billion of application software spending, about a fifth of enterprise SaaS spend, is at risk from agents by 2030. "You are no longer buying software primarily for people; you are increasingly buying it for agents," says Gartner's George Brocklehurst.

Second, incumbents face a painful crossing. When HubSpot announced its switch, its shares closed down 19% at $197.35 the next day. Its chief executive, Yamini Rangan, called it "outcome-based pricing in its simplest form. Customers pay when the agent works." Its finance chief warned the changes "may extend sales cycles", and only about 9,000 of its nearly 300,000 customers were using the agent. The arithmetic of the switch shows why investors flinched: at about $1 a conversation, 100 conversations earned HubSpot about $100; at 50 cents per resolution and the roughly 70% success rate its chief executive cites, the same 100 earn about $35. A seat business must cannibalise itself to cross; an AI-native one starts on the far side. Taylor himself called the February selloff "rational, but... a bit overblown" and "an indictment of the sector more than... of an individual company". His test for who survives is how close a product sits to the system of record: the ledger or database a company cannot run without is durable, while the screens people click through on top of it are the part agents replace.

Exhibit — HubSpot's switch cut what 100 AI conversations earn it from about $100 to $35. Source: Boston Globe (8 May 2026): HubSpot's prices before and after 14 Apr 2026 and CEO Yamini Rangan's ~70% success rate. Arithmetic by Dan's AI Intel; assumes each success is billed as a resolution.

Third, the industry consolidates around those who can carry the risk. Salesforce agreed in June 2026 to buy Fin, the former Intercom, for about $3.6 billion; Salesforce's own Agentforce had reached $1.2 billion in annual recurring revenue.

Fourth, the fight moves to definitions. "The moment your service offers N types of value props or outcomes, you need N pricing models to go along with it," Box's chief executive Aaron Levie warned in 2024. Every vendor now needs a rule for what counts: HubSpot's 72 hours without a human, Zendesk's resolutions checked by a separate AI evaluation model. Attribution is the other trap. "Just because your tool was part of the workflow, it doesn't mean that your tool was the one that delivered the outcome," the product adviser Karan Peri argued in the same debate; a sale closed by an agent may owe as much to a discount or a salesperson. A free hand-off also creates a temptation to keep customers away from humans, which is the strongest argument critics make against the model.

Is this a good change? For buyers it largely is: they stop paying for shelfware and failed attempts, and the vendor's incentive points at their result. Sierra is doubling down rather than retreating. Its Horizon platform, launched in July 2026, extends per-outcome pricing from support tickets to loans refinanced, claims paid and sales closed. Whether rivals can match it depends less on their technology than on whether they can survive the revenue dip of leaving seats.

How Sierra runs: narrow jobs, guardrails, evals and disposable code

Sierra's lessons for building an AI-native company come down to seven practices, most of them stated by Sierra itself.

  • Narrow the job first. Sierra warns customers against "AI tourism" — starting from the technology — and starts from a defined job instead. "Narrowing the problem turns AI from a science project into an engineering problem," its February 2026 review says.
  • Goals and guardrails, not scripts. Agents get an objective and limits, and supervisor models check their reasoning before it reaches a customer.
  • Evaluation is a product. τ-bench, published by Sierra's researchers in 2024, was built to test agents doing real tool-and-customer tasks repeatedly; its successors now test voice, many languages and exact details.
  • Expect to throw work away. "So much of what we write, we plan to throw out later," Taylor says; Sierra built its own reasoning methods before frontier models absorbed them. "You have to have the best technology and have to be comfortable with throwing it out."
  • Agents build agents. Ghostwriter, introduced in March 2026, turns standard operating procedures, call transcripts and plain-English goals into deployable agents; it is how the 27-day retail deployment happened. Sierra's March post calls the model "prompts, not clicks".
  • Hire for the new work. In April 2026 Sierra dropped its two coding interviews and its algorithms interview. Candidates now plan a product, build it for two hours with whatever AI tools they choose, then defend the code and its path to production.
  • Organise around processes. "The atomic unit of productivity in AI is a process, not a person," Taylor says; companies, he adds, "ship our org charts".
  • Invent the roles the work needs. Sierra created the agent engineer in 2024: a software engineer who also chooses models and designs prompts, orchestration and integrations for systems that are not deterministic. "There's an enormous difference between building a demo and deploying an AI agent at scale," its post introducing the role says.
  • Keep the memory, rent the models. Sierra tells customers that the record of their customer interactions stays theirs as models change, so they can move to a better model without starting again.

Where does Sierra do this? Everywhere the work is repetitive, rule-bound and measurable: support, billing, subscriptions, mortgage origination, insurance claims and healthcare scheduling — in regulated industries first, because that is where the rules are written down. The lesson for others is that an AI-native company is defined less by how much it automates than by how precisely it defines the job it automates.

The overreach was elsewhere: Klarna and CBA, not Sierra, rolled back

There is no public record of Sierra having to pull its automation back. The closest thing to an overreach is a failure of configuration, not of strategy. In December 2025 a coordinated attacker tried to jailbreak more than a dozen agents running on Sierra. According to Sierra's statement, as reported in the trade press, its abuse detection stopped the attempts everywhere except Gap, where a guardrail had been misconfigured, and Gap's shopping agent was reportedly coaxed into off-topic and offensive conversation before the setting was fixed. The lesson is narrow but real: an agent is only as safe as its least careful deployment.

The second strain is human, not automated. Sierra's model has relied on its own engineers sitting inside customers to build and tune agents, and reviewers' most repeated complaint is having to wait on that team for changes. Ghostwriter is, in effect, Sierra pulling back from bespoke human labour, not from AI.

The companies that genuinely went too far were its customers' peers. Klarna promoted an AI assistant doing the work of 700 agents; in 2025 its chief executive admitted the cost-cutting push had gone too far and produced lower quality, and Klarna began moving people back into support. In Australia, Commonwealth Bank said in July 2025 it would cut 45 customer-service roles because its AI had reduced calls by 2,000 a week; the Finance Sector Union disputed the claim, and within weeks the bank reversed the cuts. Both failures share a design: automation counted as a win whether or not the customer's problem was solved. Sierra's pricing makes that impossible to hide. If the agent fails and hands over, Sierra earns nothing.

Where Sierra pushes the edge: weeks-long agents and agents that build agents

Sierra is pushing the boundaries in five specific ways.

It is stretching the task from minutes to months. Horizon agents pursue an outcome over days or weeks — following up on a loan, arranging a specialist referral that can take dozens of texts and calls — and each sale won or lost feeds back into the agent's decisions.

It is turning the agent-builder into a colleague. Since late September 2026 Ghostwriter can sit in Slack and Teams as a teammate: it reviews recent calls, proposes changes with the conversations behind them, runs the experiments, and reports when results are statistically significant. In one case it found that a "payment failure" was really a successful payment with a missing policy number.

It is building for the most regulated work first. In the last week of October 2026 Sierra launched a Compliance Console and hired a physician, Dr Mitesh Patel, as head of clinical — signals that healthcare, where an error is costly and the rules are written down, is the next frontier.

It is closing the transaction inside the conversation. With Stripe, a caller can pay mid-call while card details bypass the agent entirely; at the medical-billing company Cedar, 71.3% of calls in which a patient started a payment ended in one.

It is writing the rules for agents talking to agents. In October 2026 Sierra published a draft Personal Agent Protocol, Poppy, which lets a customer's own AI identify itself to a company and act with limited, customer-approved access. Meta, Shopify, Stripe and Walmart were among its launch partners, and 35 more design partners joined, among them OpenAI, Visa, Mastercard and Bank of America.

And it publishes the tests that expose its own limits.

Exhibit — Requiring a pass in noisy and background-speech audio too cuts scores by 44–55%. Source: Sierra Research, τ-Entity leaderboard, pooled across six languages (22 Oct 2026). Cuts shown in points and relative to the regular-audio score. Compiled by Dan's AI Intel.

Its τ-Entity benchmark asks a voice agent to capture one detail — a name, a phone number, an address — exactly. The best agent did it on 60.3% of ordinary calls and on only 34.0% when it had to succeed on regular, noisy and background-speech audio alike. Grok's and Gemini's voice agents tie on ordinary calls, at about 44%, but split to 24.0% and 19.5% once the audio gets hard. Requiring success in all three conditions cuts each leading agent's score by between 44% and 55%. Three-field calls succeeded 53% of the time despite 75% accuracy per field, the same compounding our benchmark-trap episode described. A company that publishes numbers this unflattering about the medium it sells is betting that measured honesty beats demo polish.

Who wins: workers, lawmakers and the labs

For the economy, the prize is the wage bill. America employed about 2.81 million customer-service representatives in 2024, at a median of $20.59 an hour, and the Bureau of Labor Statistics projects a 5% decline by 2034. Outcome pricing converts part of that wage bill into software revenue, priced against the cost of a person rather than the cost of a licence.

For politics, the collision has started. The Keep Call Centers in America Act, introduced in the US Senate in July 2025 by Ruben Gallego and Jim Justice and backed by the Communications Workers of America, would require companies to disclose at the start of a call that an AI is answering and to transfer a customer to a US-based human on request. It sits in committee, but its two demands — disclosure and a right to a person — are the terms on which agents will be allowed to scale. Australia has already seen the argument play out at Commonwealth Bank, where the Finance Sector Union's Julia Angrisano said workers "want a tech-savvy bank, but they expect to be part of the change, not replaced by it".

For the race between the labs, Sierra shows where value is captured. It runs on models from OpenAI and Anthropic and switches between them; because it charges a fixed rate per outcome, a cut in model prices widens Sierra's margin before it reaches any customer. Its chief executive also chairs OpenAI's board, a dual role that will draw more scrutiny as the application layer earns the outcome while model makers compete to be its inputs. Taylor says Sierra is "multiples larger than the next biggest" rival; Decagon, a venture-backed rival, was valued at $4.5 billion in January 2026. Sierra's own bet on where durable value sits is explicit: "the memory of your customer interactions", it says, "becomes a durable, expanding moat" — a moat that belongs to the application company and its customer, not to whichever lab supplies this month's best model.

Bottom line

Sierra's real invention is a unit of sale: software paid like a contractor, by the job, with the risk of failure on the seller. It works because customer service has a countable outcome, and it grew to more than $400 million in revenue because that outcome draws on the budget for people rather than the budget for software. The hinges to watch are three: whether outcome definitions hold up as agents take on weeks-long work; whether incumbents can survive the dip of leaving seats; and whether lawmakers' demand for disclosure and a human on request becomes law. If all three break Sierra's way, the seat will look as dated as the boxed software it replaced.

Sources

Transcript

Alex: So I'm staring at my help-desk renewal, and there are three different kinds of "resolution" on the page, and only some of them cost money. Sierra: the AI company paid only when the job gets done. If its agent has to fetch a human, the bill is zero.

Sam: And it still crossed four hundred million dollars a year. That is a very strange way to get rich. Before we start, my week. Our tool budget locks on Friday, so Wednesday night I was reading Microsoft and Alphabet earnings like tea leaves.

Alex: That is the saddest sentence you have said on this show.

Sam: I know. Microsoft spent about forty-five billion dollars on capital equipment in a single quarter. Alphabet more than doubled its spending from a year ago. And Satya Nadella said, quote, "As agents take on higher-value work, consumption opens up business-funded demand."

Alex: Consumption. That's the polite word for a meter.

Sam: That's what I wrote on my budget sheet, right next to my vendors' names. If the agents do the work, somebody is going to count the work, and then somebody is going to send me the count.

Alex: Hm. That is today's episode, and you found it in your own spreadsheet.

Sam: Wednesday, eleven at night. Very glamorous.

Alex: Welcome back to Dan's AI Intel, where we read the footnotes so you don't have to, and try to help a curious person understand the AI revolution rather than just keep pace with it. What it means for the economy, for politics, for the race between the labs. I'm Alex.

Sam: I'm Sam. And today we're talking about a pricing sheet.

Alex: A pricing sheet worth more than four hundred million dollars a year. Here's the trigger. A customer-service company called Sierra, started by two former Google product managers, says it only gets paid when its AI agent finishes the job. Resolve the case, there's a rate. Hand it to a person, nothing.

Sam: And underneath it is a bigger question: who pays for software once software starts doing the work of people? We'll go through where Sierra came from, who is copying it, who got punished for trying, and what happens to the humans in the middle.

Alex: There's a company in here that did the careful thing and got hammered for it, and a pair that did the loud thing and had to reverse. Which is which is not the obvious answer.

Sam: And if you like this show, hit follow in whatever app you're in. It's free, it's one tap, and the next one lands the moment it's out.

Alex: Before the history, I want us to say where we start, because I think we start in different places.

Sam: We do.

Alex: My honest starting position is that this is the first software deal I've ever wanted to sign. The vendor eats the failure. The vendor eats the wasted compute. I have a hundred and forty or so software contracts, and in nearly every one of them I pay whether or not anything got done.

Sam: My starting position is that "done" is a word the seller gets to define, and I've been burned by sellers defining words. Every price I've been handed that was "simple" turned out to have a footnote doing all the work.

Alex: That's a bold claim from someone who never reads the footnote.

Sam: I read the footnote. It's how I know where they hide it.

Alex: Fine. So you think the footnote is the whole game.

Sam: I think the footnote is where the money lives. If I only pay for finished work, someone decides what finished means, and the same someone has every reason to say it a lot. Okay, but who pays for that?

Alex: The vendor, in my version. If the agent fails, they get nothing.

Sam: Or the customer, in mine, who gets quietly steered away from the human they actually need.

Alex: So we're arguing about the same free handoff. You think it's a trap, I think it's the point.

Sam: Let's see who's right. So what does Sierra actually sell? Because "AI agent" is what everyone sells now.

Alex: It builds an agent carrying a company's name that does a whole job from start to finish. Refresh a car radio. Refinance a mortgage. File a claim. It talks over chat, voice and email, in more than thirty languages.

Sam: And the person on the phone thinks they're talking to Sonos or Chime, not to Sierra.

Alex: That's the point. And the line I like is from Clay Bavor, one of the two founders. When they came out of stealth holding a hundred and ten million dollars from Sequoia and Benchmark, he told Fortune, "Most people just want to buy a couch."

Sam: A couch.

Alex: Meaning most big companies don't want a box of AI parts and a manual. They want the finished thing in the living room. So Sierra started with a small team building it for them. WeightWatchers, SiriusXM, Sonos, OluKai.

Sam: I love that the customer list reads like my weekend errands.

Alex: And the money model came with it. Here's Bret Taylor, the other founder and the CEO, in March. "If the AI agent resolves the case, no human intervention, there's a pre-negotiated rate for that. If we do have to escalate to a person, that's free."

Sam: Free. Say that slowly, because I run a budget. Free.

Alex: Free. And then he said the part that made me sit up. On model costs: "reducing your token utilization for the same outcomes is your problem, not your customer's."

Sam: So when the AI burns a pile of compute and still fails, Sierra eats the lot.

Alex: All of it. That's backwards from every cloud bill I've ever paid.

Sam: Hold on. "Outcome" is the word I'd circle in red. Who decides what counts?

Alex: Case by case, with each customer before launch. Their own pricing note lists a resolved support conversation, a saved cancellation, an upsell, a cross-sell.

Sam: Then I want the rates in writing, because "we'll tell you after" is how every price I distrust started. Do they publish them?

Alex: No. Third parties guess somewhere between one and two and a half dollars a resolution, and the company has never confirmed it. And they say the model "may not always be the best fit." An agent that only greets people and routes calls gets billed per conversation instead.

Sam: There's the footnote. The footnote already exists, and it's in their own note.

Alex: I'll give you that. They're honest about where it doesn't fit. To see why this matters, look at how the other deals work. Say I sell software per seat, and I add an AI agent that resolves a customer's problem. Their human support team needs fewer people. At renewal they need fewer seats.

Sam: So the better the AI, the smaller the vendor's bill.

Alex: The vendor is paid less every time its own product works. That's the trap. Taylor sat on the other side of that trap for years, and we'll get to it.

Sam: And the in-between pricing? Per conversation, per credit?

Alex: Those track activity. You pay for the attempt, so a failed attempt costs the same as a good one. Sierra's note calls unused seats "shelfware". Under per-resolution pricing, the vendor eats every failed try.

Sam: Hm. That is a nicer deal for me than I'd like to admit.

Alex: Say it again, I want to record it.

Sam: No.

Alex: What's underneath is what Sierra calls Agent OS. A "constellation of models", in Taylor's words, from OpenAI and Anthropic, plus Sierra's own fine-tuned layers. The agent gets "goals and guardrails, not the sequence of steps", and supervisor models send back any decision that breaks policy.

Sam: Goals instead of scripts. So it isn't a phone tree with a nicer voice.

Alex: A phone tree follows a flowchart. This is more like telling a new hire the goal and the rules, and a manager checking the tricky calls.

Sam: And that's why eating the failures is affordable. The agent is good enough that it's worth it.

Alex: That's the bet. And they place it on every single call.

Sam: Does the bet pay? You keep saying four hundred million and I want to see how it got there.

Alex: This is where I get nerdy, sorry in advance. Two years ago Sierra had about twenty million dollars of annualised revenue, according to people who spoke to Reuters. By last November it was a hundred million. Now it's more than four hundred.

Sam: That's not a curve, that's a wall. And before anyone gets excited, that's a number the company reports about itself, and I've been handed plenty of those in pitch meetings.

Alex: A doubling about every five and a half months, and the pace has barely varied. And you're right, these are company figures, not audited.

Sam: Noted, and I'll keep saying it.

Alex: The investors paid up fast too. The latest round, in May, put the company at fifteen point eight billion dollars, led by Tiger Global and Google's venture arm.

Sam: That sounds like a bubble number.

Alex: Funny thing is, against its revenue it got cheaper. In the October round two years ago investors paid about two hundred and twenty-five times Sierra's revenue. In May, about seventy-nine times.

Sam: So the business outran even its fans.

Alex: Outran the people who believed hardest, which I'd have called impossible, because in my own failed startup the story outran the business, and then the business ran out, and I can still tell you the exact shade of the boardroom carpet —

Sam: Alex.

Alex: Sorry, where was I going with this?

Sam: Carpet. Tangent. Please come back.

Alex: The customers got bigger as it grew, and now Sierra works with almost half of the Fortune 50.

Sam: Which makes me look hard at the vendors on my own sheet, because none of them has that customer list.

Alex: And the newest agents don't just answer questions. They originate loans, chase prior authorisations for medical procedures, take payments inside the call. One huge US retailer went from first meeting to a live agent in twenty-seven days.

Sam: I once watched our IT team argue for a month about a font.

Alex: Which is why I'm nervous and impressed at once. The couch-builder became a factory, and our storefront bot looks like a hobby next to it.

Sam: So who are these two? Because "two ex-Google product managers" is also the description of about half my LinkedIn.

Alex: These two have receipts. Taylor and Bavor both started as associate product managers at Google, and they met there. Taylor led the local-search work that became Google Maps.

Sam: Okay, that's a good receipt.

Alex: Then FriendFeed, which Facebook bought, then Facebook's CTO. Then Quip, a document editor that Salesforce bought, where he rose to president and chief operating officer and led the purchase of Slack.

Sam: And then co-CEO of Salesforce.

Alex: And he chaired Twitter's board through Musk's takeover, and he chairs OpenAI's board today.

Sam: Let me do Bavor, because it's the one I'd have forgotten. Eighteen years at Google. He led product and design for Gmail, Docs and Drive, then started the virtual-reality effort and Project Starline, a life-size video call, and ended as head of Google Labs reporting to Sundar Pichai.

Alex: So one of them ran maps and the other ran Gmail. Between them they're specialists in making strange technology feel normal.

Sam: Which is the whole design problem of an agent picking up a customer's phone call.

Alex: And the Salesforce bit matters most. Salesforce is the biggest company ever built on selling software per user. Taylor ran it. The founding argument of Sierra, that a seller paid per seat earns less every time its AI succeeds, comes from a man who sat on that side of the conflict.

Sam: He knows where the bodies are buried, and now he sells shovels to the other side.

Alex: Ha. And the origin story is lovely. A lunch at a Mediterranean restaurant in Palo Alto. Taylor assumed his old colleague was checking in on him after he left Salesforce.

Sam: People always assume lunch is a wellness check.

Alex: They talked for hours and walked out having decided to start a company. They named it after the Sierra mountains, to signal ambitions beyond Silicon Valley.

Sam: And that pairing, plus the OpenAI board seat, is why the checks showed up.

Alex: I can't prove that part, but investors paying up that fast says nobody needed a lot of convincing.

Sam: Why customer service first? Why not everything at once?

Alex: Because there you can count the outcome. Either the customer's problem got solved without a person, or it didn't. And Taylor puts the annual spending on customer service at roughly four hundred billion dollars.

Sam: That's a huge wage bill with a very clean scoreboard, and my CFO is going to love a clean scoreboard.

Alex: And look at the neighbours. Everyone in my help-desk aisle has moved.

Sam: Intercom went first, with the Fin thing, if I remember —

Alex: Zendesk, which says it was first, in August two years ago.

Sam: Zendesk, yes. I mix those two up. I once sat through a whole budget meeting with the wrong vendor's invoice open.

Alex: Zendesk charges roughly a dollar fifty per automated resolution on committed volume. Intercom's Fin is ninety-nine cents.

Sam: And HubSpot flipped its customer agent this spring, from about a dollar a conversation to fifty cents per resolved conversation, and a conversation only counts as resolved if no human takes over within seventy-two hours. Salesforce followed with a help agent that charges per resolution too.

Alex: Within two years every major help-desk vendor sells AI the way Sierra does, which is why my renewal has three kinds of "done" on it.

Sam: Why now, and not years ago?

Alex: Two reasons. I said in the agents episode that agents still don't really work, and for open-ended jobs I'd still say it. But for a narrow, countable job like a support case, the published resolution rates changed my mind: agents can finally finish a task without a person, so there's something to bill for. A chatbot that sent you to a help page produced no result. And the cost of finishing a task got small enough for the vendor to swallow.

Sam: In marketing we went from paying per impression to paying per click, and nobody went back. I'd guess he'd say the same here.

Alex: He does, almost word for word. Online advertising moved from paying per impression to paying per click. The closer the price sits to the value you receive, the more efficient the market. And he says paying by usage is a weak proxy, because the tokens a model burns only loosely track whether the customer got helped.

Sam: I can use that with my CFO.

Alex: Mine is better. A seat is like paying a plumber a monthly retainer whether or not anything leaks. Per resolution, you pay for each leak that's fixed.

Sam: And here's where I put on gloves. Because "a leak that's fixed" depends who inspects the leak.

Alex: Yes. Which is where I get nervous about my own week. Take mine. I've got a help-desk contract coming up for renewal, and it now bills per automated resolution. Zendesk's rate is what I just said. And in May they split the counting into assisted escalation, contained resolution and verified resolution, and only some of those count against my allowance.

Sam: Three kinds of "done."

Alex: Three. So the footnote has its own footnotes.

Sam: Remind me who stood on this stage ten minutes ago and said it's the first deal he wants to sign.

Alex: Okay, fair, I deserved that. And analysts say the cost is hard to forecast, because everything turns on how a resolution is counted.

Sam: You, of all people. You're the one who can't stand a bill you can't forecast.

Alex: I am, annoyingly. I'm holding two opposite feelings and I'd like it noted.

Sam: Noted in the minutes.

Alex: But here's where I dig in. A bill that moves with success is not the same as a bill that moves with nothing. When my seat count goes up I pay for people who might be idle. When a resolution count goes up, something got solved.

Sam: Unless it got solved by a discount code and a coin flip.

Alex: Hm. That's fair. We'll come back to that, it's the attribution problem.

Sam: Before we get to who got hurt, give me the honest speed. Is every software company about to charge per result?

Alex: No, and this is the number that stopped me. Pure outcome pricing for a company's main model is rare, roughly one company in twenty.

Sam: Then what's all the noise about?

Alex: Hybrid. A platform fee plus usage or credits plus a slice of outcome. A survey of software companies by Kyle Poyar found hybrid pricing went from a quarter of companies to over a third in a year. And ICONIQ found the share of AI builders using some outcome pricing rose over the first half of this year too.

Sam: So it's an ingredient. It isn't the recipe.

Alex: Very fast as an ingredient, slowly as the whole recipe, which is how I'll negotiate: a platform fee plus a slice of outcome. And the biggest vendors hold seats hardest. The companies above a hundred and fifty million in revenue are the most likely to keep them.

Sam: Because they've got the most seat revenue to lose. I know two of my vendors who'll cling to seats until the last renewal.

Alex: Which brings us to the sad story.

Sam: This one's in my tools line, so I have feelings. HubSpot. A day after the pricing change, the shares closed down nineteen percent.

Alex: In a day, for being honest about the price.

Sam: And the arithmetic tells you why. A hundred conversations used to earn about a hundred dollars. At fifty cents a resolution, and a success rate near seventy percent, the same hundred earn thirty-five.

Alex: Down to roughly a third.

Sam: And that's before the critics point out the same thread can be billed again if the issues land more than three days apart. Their definition of done has a window in it.

Alex: Which is your footnote.

Sam: It's the footnote, and I'll still say they put it in plain English. The finance chief warned the changes "may extend sales cycles."

Alex: Which I'd read as a warning to every vendor in my renewals: the first to switch gets punished.

Sam: And Yamini Rangan, the CEO, called it "outcome-based pricing in its simplest form. Customers pay when the agent works." It's a lovely sentence. It cost her a fifth of the company's value in a day, and my budget sheet had the share price on it for about ten minutes.

Alex: That's the crossing. A seat business has to eat itself to make the switch. An AI-native company like Sierra starts on the far side, and I'd rather my vendors were over there than stuck halfway.

Sam: And two of the vendors on my sheet are on the wrong side of that race.

Alex: Taylor was dry about it. He called the February selloff, when software stocks lost about a quarter of their value between mid-January and late February, "rational, but... a bit overblown," and "an indictment of the sector more than of an individual company."

Sam: He would say that. He sells the other thing.

Alex: He would. But his test for who survives is interesting. How close does a product sit to the system of record? The ledger or database a company can't run without is durable. The screens people click through on top of it are what agents replace.

Sam: That's my martech stack. Screens on top.

Alex: It's half of my contracts. And last week the CIO of Morgan Stanley's wealth division, Hina Shamsi, said the bank flipped the script on how it approached data to enable AI.

Sam: That's the data team I keep asking my CMO for. Same instinct as Taylor's test, from the other end. There's a point that scares me more than the share price. If software gets paid for results, whose money is it?

Alex: This is the big one. In Poyar's survey, most companies say their AI revenue comes out of existing software budgets. But the AI-native ones get paid more often from services budgets, and from headcount budgets.

Sam: Headcount. They're being paid out of my people line.

Alex: And Gartner estimated in July that up to two hundred and thirty-four billion dollars of application software spending is at risk from agents by the end of the decade. About a fifth of enterprise software spending.

Sam: George Brocklehurst at Gartner put it this way: "You are no longer buying software primarily for people; you are increasingly buying it for agents."

Alex: Which is a strange sentence to say out loud on a Monday.

Sam: I've got a big team, and a good part of it is junior copywriters. If a vendor says "that line is mine now, I'll bill you per finished job," that's not a software conversation. That's a conversation about who I employ.

Alex: And the vendor pitches it like a bargain.

Sam: And it might be one, which is what makes it awful. In the end whose money is it, mine or the copywriter's?

Alex: In my version, the vendor pays when it fails. In yours, a person pays when it succeeds.

Sam: Both are true. That's what I can't unsee.

Alex: Hm. That's a better way of putting the argument than mine. Let me not agree with it too fast.

Sam: You can disagree with it slowly.

Alex: While we're on money moving, the field is consolidating around people who can carry the risk. In June Salesforce agreed to buy Fin, the former Intercom, for about three point six billion dollars.

Sam: And a vendor who takes on the risk needs deep pockets. So the small players get squeezed out.

Alex: That's a fair point, and you said "squeezed out" in the voice you use for underdogs.

Sam: That's a very specific accusation, and I will be looking into it.

Alex: Look into it later. Because the risk-carrying only works if everyone agrees what the risk is.

Sam: Meaning, back to the footnote.

Alex: The definition fight, then. This is where your footnote lives. Aaron Levie, the CEO of Box, warned in 2024, "The moment your service offers N types of value props or outcomes, you need N pricing models to go along with it."

Sam: So every outcome needs its own rulebook.

Alex: And every vendor now needs a rule for what counts. HubSpot's is a window with no human takeover. Zendesk's resolutions are checked by a separate AI model.

Sam: So an AI grades the AI. Which feels like marking your own homework.

Alex: It's a separate model, but I take the point. The other trap is attribution. Karan Peri, a product adviser, said, "Just because your tool was part of the workflow, it doesn't mean that your tool was the one that delivered the outcome."

Sam: A sale an agent closed might owe as much to a discount, or a salesperson.

Alex: Right.

Sam: And the strongest argument I have against all of it is the free handoff. If I only get paid when the human isn't needed, I have every reason to keep you away from the human.

Alex: Which is the strongest argument critics make against the model, and it's right there in the pricing.

Sam: It's the one that makes me stop.

Alex: Hold that thought, because the next act is the best evidence on exactly that. Who actually kept customers from humans, and who got caught.

Sam: First, though, what do the customers say? Because the pitch is all numbers, so let me read you the numbers.

Alex: Go on.

Sam: Almost all of it comes from Sierra's own customer pages. Airtable says Sierra's agent resolves about eighty percent of its support conversations, the work of roughly ninety-five full-time agents. Chime and SoFi report glowing numbers too, and rivals like Fin quote about the same band.

Alex: So across vendors, most conversations now end without a person.

Sam: And that's what bugs me. Every one of those numbers was picked by someone who profits from you reading it. They publish the figure that flatters, and I'd never accept that from my own agency.

Alex: Fair. It's why the unflattering benchmark later in this episode matters.

Sam: That caveat sits next to every one of these numbers. Fine, tell me one that isn't a resolution rate.

Alex: Rocket Mortgage. Clients who use both its AI chat and a human banker convert at four times the rate on refinancing. Clients who start with the digital assistant close at a much higher rate than those who don't.

Sam: The human is still in that.

Alex: Yes, and that's the interesting one. The agent doesn't replace the banker, it warms the lead. Rocket's agent now places over a million outbound calls a month.

Sam: I get those calls at dinner and I hang up on them, so it had better be good.

Alex: SiriusXM has thirty-four million subscribers who talk to an agent called Harmony, which replaced a keyword chatbot that led to dead ends.

Sam: And the chief product and information officer, Moshe Pridan, said, quote, "Sierra is a game-changer for meeting a subscriber's current needs." Game-changer. I'm putting a dollar in the jar.

Alex: And Peter Fenton of Benchmark, who backed them early, told CNBC, "Sierra is by all measures the winner in the 'customer experience' category."

Sam: The early investor thinks it's a winner. Hold me.

Alex: Now the fair part. None of the figures is independently audited. Review sites and rival vendors report two recurring complaints: customers have to go back to Sierra's engineers to change an agent, and costs are hard to predict.

Sam: There's that word again. Hard to predict.

Alex: And the model leans on Sierra's own forward-deployed engineers sitting inside each customer. That's expensive, and it's exactly what their newer self-build tools are meant to replace.

Sam: I have a small version of this from my own life. My best friend Lauren owns a small salon in Decatur, and she's looking at an AI phone receptionist billed per call.

Alex: Per call.

Sam: One vendor, Quo, has one called Sona: twenty-five dollars a month, then a dollar for each call past the allowance. Another, Allo, advertises an unlimited receptionist from eighteen dollars a month. And the vendors themselves say a cheap one might only take a message instead of booking the appointment.

Alex: Which is the difference between a call and an outcome, and I'd bet the vendor's slides blur the two on every page.

Sam: A message isn't a booking. Lauren keeps a tool only if it pays for itself in a month, so she'll be the one asking "what counts as done," and she'll be asking it with scissors in her hand.

Alex: She's the best customer on earth.

Sam: She's the most dangerous customer on earth. I told her to ask what the tool does when it can't book, because that's where the bill is.

Alex: Free handoff.

Sam: I was about to say the same. Now the companies that really overreached, because it wasn't Sierra. There's no public record of Sierra having to pull its automation back.

Alex: Not even the Gap thing?

Sam: The Gap thing is the closest. Last December a coordinated attacker tried to jailbreak a whole batch of agents running on Sierra. Sierra's statement, as the trade press reported it, says its abuse detection stopped the attempts everywhere except Gap, where a guardrail was misconfigured. Gap's shopping agent was reportedly coaxed into off-topic and offensive conversation before the setting was fixed.

Alex: A configuration slip, not a strategy slip. And it's my nightmare, because an agent is only as safe as its least careful deployment, and at my company that would be a storefront bot some team spun up in a week.

Sam: That's true of every contractor I've ever hired, too.

Alex: Then the genuine overreach. Klarna promoted an AI assistant doing the work of seven hundred agents. Last year its chief executive admitted the cost-cutting push had gone too far and produced lower quality, and Klarna started moving people back into support.

Sam: And in Australia, Commonwealth Bank said in July last year it would cut forty-five customer-service roles because its AI had reduced calls by two thousand a week. The Finance Sector Union disputed the claim, and within weeks the bank reversed the cuts.

Alex: And here's what I'd steal for my board deck. Both share a design: automation counted like a win whether or not the customer's problem got solved. Klarna and the bank got to claim savings without proving the job was done.

Sam: And Sierra's pricing makes that impossible to hide. If the agent fails and hands over, Sierra earns nothing.

Alex: The scoreboard is on the vendor.

Sam: So that moves me. A little. I came in saying the free handoff is a trap, and here it's the thing that would have caught Klarna.

Alex: Careful. Say what you mean.

Sam: The pricing term is the only thing in this business that makes the vendor lose when the customer isn't helped. Everything else counts calls deflected. That's an honest improvement and I'll say so. I still don't want it near my junior copywriters' line until someone shows me the definition in writing.

Alex: That's a fair place to be. It's also, weirdly, where I've ended up after the Zendesk thing.

Sam: Okay, so the question I care about. Does Sierra write its own definition of "done," or does the customer? Because everything we've said hangs on that footnote.

Alex: The customer, with Sierra, before launch. And that's rule one in how they build: narrow the job. They warn against what they call "AI tourism", starting from the technology instead of a defined job. Their February review says, "Narrowing the problem turns AI from a science project into an engineering problem."

Sam: A science project is what my team calls the thing nobody finishes. But it also means "done" is easier to write down, doesn't it? A narrow job has a narrow footnote.

Alex: That's the cleverest thing you've said all episode. A narrow job is a countable job, and a countable job is a billable one. The pricing and the product design are the same decision.

Sam: And the second decision, goals and guardrails instead of scripts, is why the failure bill lands on them?

Alex: Right. Which is why they expect to throw work away. Taylor says, "So much of what we write, we plan to throw out later," because the frontier models keep absorbing what they built.

Sam: That's the opposite of your engineers, Alex. You defend code you wrote years ago like it's a pet.

Alex: That's slander, and it's also true. The piece I'm happiest about is Ghostwriter, which came out in March. It turns a company's standard operating procedures, call transcripts and plain-English goals into deployable agents, which is how that retail deployment got live so fast. Their phrase is "prompts, not clicks", and it's the kind of elegant idea I fall for.

Sam: You fall for it. I read it as Sierra pulling back from the complaint that you have to wait on their engineers for every change. So I'd ask a different question: when it writes the agent, who checks it?

Alex: Sierra treats evaluation as a product in its own right, and supervisor models check the agent's reasoning before it reaches a customer.

Sam: So who audits the auditor?

Alex: I don't know, and I'd put that question in the contract. It's also where the hiring story lives. In the spring Sierra dropped its coding interviews and its algorithms interview. Candidates now plan a product, build it in a couple of hours using whatever AI tools they like, then defend the code and its path to production.

Sam: Defend it. That's the part.

Alex: My daughter Lucia writes code at a payments start-up. She isn't afraid of the tool, she's afraid the first rung of the ladder is the one that gets automated, and she defends craft and code review. A test where you defend what the AI wrote at least tests the part she can do.

Sam: Hm. That's the first thing in this whole show that makes a young person look good.

Alex: And the last rule is for your side of the table. Keep the memory, rent the models. Sierra tells customers the record of their customer interactions stays theirs, so they can move to a better model without starting again.

Sam: That's lock-in with the lock taken off.

Alex: Which is the thing you hate most, so that earns them a point.

Sam: One point. I'm tallying. But my tally gets nervous when the job gets longer. If "done" is a footnote today, what is it when the job takes three weeks?

Alex: That's exactly where Sierra is going. Its Horizon agents chase an outcome over days or weeks, following up on a loan, arranging a specialist referral that can take dozens of texts and calls, and each sale won or lost feeds back into the agent's decisions.

Sam: Seventy-two hours doesn't cover that.

Alex: And that's why the Horizon platform, launched in July, extends per-outcome pricing from support tickets to loans refinanced, claims paid, sales closed. It's pricing something you can only count at the end of a long road.

Sam: That's the footnote's boss.

Alex: And the builder side keeps growing. Since late September, Ghostwriter can sit in Slack and Teams like a teammate. It reviews recent calls, proposes changes with the conversations behind them, runs the experiments, and reports when the results are statistically significant.

Sam: Which, for my team, means briefs that get tested instead of argued about. I'd take that.

Alex: And in one case it found that a "payment failure" was really a successful payment with a missing policy number. A typo, basically. I will be stealing that story for my next review meeting with the VP of Engineering, who thinks every incident is a design flaw.

Sam: Every report I've ever read called something a failure that was a typo. And the payments get better. With Stripe, a caller can pay mid-call while the card details bypass the agent entirely. At the medical-billing company Cedar, a bit over seventy percent of calls where a patient started a payment ended in one.

Alex: Bypass the agent. That's the first thing I'd have asked.

Sam: And then Poppy. In October Sierra published a draft Personal Agent Protocol, which lets a customer's own AI identify itself to a company and act with limited, customer-approved access. Meta, Shopify, Stripe and Walmart were launch partners, and thirty-five more design partners joined, including OpenAI, Visa, Mastercard and Bank of America.

Alex: So your agent calls their agent.

Sam: And the two of them agree on a price, which should keep me awake.

Alex: The nightmare and the dream in one sentence. And before we go further, I want to know what ordinary people think about handing over the final call. Healthcare is the test. At the end of October Sierra launched a Compliance Console and hired a physician, Dr Mitesh Patel, as head of clinical. That tells you where they think the next frontier is: where an error is costly and the rules are written down.

Sam: What do patients think? That's what I'd want to know.

Alex: My wife Elena is an emergency doctor, and she sees patients arriving with chatbot answers. And Pew published a survey last week. Forty-four percent of Americans are open to AI scheduling their appointments, but only with a human final review. Twenty-nine percent would let it act alone.

Sam: And claims?

Alex: For insurance claim decisions, almost half aren't open to AI at all, and just five percent would accept it alone. And most are uncomfortable with their own doctor using AI for diagnosis and treatment.

Sam: So the closer it gets to a decision about someone's body or money, the more people want a person at the end. I'd feel that about my mother's claim.

Alex: And Elena would say the same about a diagnosis. That's why I said at the start the free handoff is the point. Pew says it's what patients are asking for.

Sam: That makes me less cynical, not more. The pricing lines up with what patients want.

Alex: And the last edge is my favourite, because Sierra published a test that makes its own medium look bad. τ-Entity asks a voice agent to capture one detail exactly. A name, a phone number, an address.

Sam: Which is what every voice call does.

Alex: The best agent got it right on about sixty percent of ordinary calls. Make it succeed on regular, noisy and background-speech audio alike, and it falls to roughly a third. Grok's and Gemini's voice agents tie on ordinary calls, then drop to roughly a quarter and a fifth when the audio gets hard.

Sam: So the scores drop by about half the moment there's a dog barking.

Alex: For every leading agent, yes. And calls needing three fields only succeeded about half the time even though each field was mostly right. That compounding is what we covered in the benchmark-trap episode.

Sam: Picture a decent basketball player who has to sink three free throws in a row. Each shot is likely, and the streak still breaks far more often than you'd expect.

Alex: That's it. And a company that publishes its own bad grades is betting honesty beats demo polish. I'm instantly in love with that, and I know that's a flaw.

Sam: It's your flaw. Falling for elegant tech.

Alex: It isn't elegant, it's a bar chart that hurts.

Sam: Let's end on who wins, and I'll start with my side, workers. America employed about two point eight million customer-service representatives in 2024 at a median of about twenty dollars an hour, and the Bureau of Labor Statistics projects a five percent decline by the mid-2030s.

Alex: Picture my store support team's line on my budget turning into a vendor's invoice. That's what outcome pricing does: part of a wage bill becomes software revenue, priced against the cost of a person rather than the cost of a licence.

Sam: And politics has noticed. The Keep Call Centers in America Act, introduced in the Senate by Ruben Gallego and Jim Justice and backed by the Communications Workers of America, would make companies say at the start of a call that an AI is answering, and transfer you to a US-based human on request. It's sitting in committee.

Alex: Disclosure and a right to a person.

Sam: That's me, every time I've been stuck in a bank's bot. Gartner surveyed thousands of people this year. Eighty-seven percent said an option to reach a human is essential when a company uses AI for service, and only twenty-seven percent would try a chatbot again after a bad experience.

Alex: So the free handoff isn't only a pricing term, it's the legal ask too.

Sam: And in Australia the Finance Sector Union's Julia Angrisano said workers "want a tech-savvy bank, but they expect to be part of the change, not replaced by it."

Alex: That's the sentence I'd put in front of our storefront team, before we automate anything.

Sam: And the people who build the models? Do they win here, or does Sierra?

Alex: That's the race between the labs. Sierra runs on models from OpenAI and Anthropic and switches between them. Because it charges a fixed rate per outcome, a cut in model prices widens Sierra's margin before it reaches any customer.

Sam: So the labs compete to be an input, and I'm the one buying the output. I'd like a price drop passed on to me.

Alex: Don't hold your breath, the margin stays with the application company. Which is odd, because Taylor chairs OpenAI's board, and that dual role will draw scrutiny.

Sam: I noticed you didn't joke about that.

Alex: Because I can't tell whether it's clever or conflicted. Sierra says its moat is "the memory of your customer interactions", which belongs to the customer, not to whichever lab has the best model this month.

Sam: As the person who signs the checks, I'd want that memory exportable on day one, in writing.

Alex: Same. My nightmare is a hundred and forty contracts, each holding a slice of my customers' history that I can't move. If Sierra means it, that's a promise I'd pay extra for.

Sam: So where do you land, against your starting position?

Alex: I started at "first deal I've ever wanted to sign." I'm ending at "I'd sign it, but only with the definition and the handoff rate written into the contract." The Zendesk renewal did that, and the seventy-two-hour window did the rest. The footnote is the product.

Sam: I started at "it's a trap." I'm ending at "it's the first term in all of this that makes the vendor lose when the customer isn't helped," and Klarna is why. But I didn't move on the other half. The money comes out of the people line. I wouldn't put my copywriters' budget on a per-job bill, even a good one, until I can see who's counting and what they count.

Alex: So you're half moved, and I'm half moved.

Sam: And we still disagree about whether turning a wage bill into a software bill is a good thing.

Alex: I think it's fine if the handoff is honest. You think it's fine if the person on the other end still has a job.

Sam: Both of those have to be true, and neither one is up to the pricing sheet.

Alex: Three hinges I'll watch from my seat. Whether definitions hold up when jobs run for weeks. Whether my vendors survive leaving seats. And whether the right to a human on request becomes law.

Sam: And mine is simpler. If all three break Sierra's way, the seat ends up as dated as boxed software, and I have to find my copywriters a new line on the sheet before my CFO does.

Alex: Which I'd buy, if the contract says what done means.

Sam: If you take three things home. Sierra gets paid a set rate per finished job, charges nothing when the agent hands over, and has grown past four hundred million a year. Every help-desk vendor has followed, but pure outcome pricing is still rare, and HubSpot learned what crossing costs. And the real fight is over what counts as done.

Alex: And the free handoff is the term that keeps vendors honest, as long as somebody reads the definition.

Sam: Thank you for listening. I hope you came away understanding a bit more of where this is heading, because it's a very short horizon to be knowable, and that is exactly what makes it worth following closely.

Alex: And one honest note on how this is made. It's AI-generated. Dan builds a custom stack of AI tools to research, analyse, verify and illustrate the questions worth understanding, mostly to learn them himself, and publishes it for anyone who'd like to follow along. AI-assisted, fact-checked, worth a second look.

Sam: Before we go, one useful thing you can do. Follow the show. Whatever app you're in, there's a follow or plus button. One tap, it's free, and each new episode arrives the moment it lands. For a small independent show like this, a follow is the biggest lever there is for reaching other people trying to make sense of all this. So if that was worth your time, go ahead and hit it.

Alex: And one more. If there's something in here you'd push back on, or a thread you want us to pull harder on next time, tell us: podcast@connectiveshift.com. We read every message, and it decides what we dig into next.

Sam: And Alex, read the footnote.

Alex: I will read the footnote.