Enterprise AI: Most Firms Bolt It On. The Winners Rebuild

An episode of Dan's AI Intel

Almost every company now uses AI, but only about one in twenty profits from it, because most bolt it onto old workflows while the winners rebuild the work itself.

Published · By Dan Walter

Executive summary

Nearly nine in ten organisations in McKinsey's global survey now use AI, and only about one in twenty firms has been transformed by it. In McKinsey's 2026 survey, 89% of organisations use AI somewhere, yet only 37% can point to any effect on profit, and just 6% earn at least a twentieth of their operating profit from it. Boston Consulting Group's 2025 study of 1,250 companies finds the same shape from another angle: 5% get value from AI at scale, while 60% report minimal or no value at all. The technology works. The firms using it mostly have not changed the work it is used on.

That is the real story beneath the noise. Most enterprises are adding AI to workflows built for people, a chatbot here and a copilot there, and getting a little time back. A few rebuild the work itself: one governed data foundation, agents that run whole workflows inside guardrails, and people who handle the exceptions the agents flag. The gap between those two approaches, not the gap between models, is the sharpest dividing line in the major consultancy surveys. Electricity followed the same path a century ago. Factory owners who swapped their steam engine for an electric motor saw productivity barely move, and the payoff came about four decades after the first power stations, when factories were redesigned around the new power.

So enterprise AI is a change programme, not a technology purchase, and it is hard for the reasons change programmes are always hard: leaders who delegate it, data nobody owns, people asked to work differently without a redesigned job to do. What works is narrow and demanding: a few core workflows rebuilt end to end, owned from the top, built with people who actually build. What does not is the default: tools rolled out onto unchanged processes, scattered pilots and advice from people who have never shipped an agent. The economy-wide gains will arrive late and land unevenly, and the AI labs, whose revenue depends on companies getting this right, have started paying to do the rebuilding themselves.

Why the enterprise is where the AI revolution is decided

The AI revolution is usually told through models: who released what, which benchmark fell. Its economic test happens somewhere duller, inside companies deciding whether to change how they work. The labs' valuations assume that change happens. Anthropic, as its public filing showed, is mainly an enterprise business, which we covered in our Anthropic S-1 episode, number 78, released on 2 November 2026. Investors already price the other side of it: software shares fell sharply in early February 2026 as investors bet that agents threatened the seat-based software model, the subject of our Sierra episode, number 81, released on 3 November 2026.

If enterprises convert, the revolution shows up in profits, wages and productivity. If they stall, it stays a consumer phenomenon with a very expensive data-centre bill attached. The question that matters is not whether AI is capable, but why so few companies turn that capability into results.

Nearly every company uses AI; about one in twenty profits from it

The adoption numbers look like a revolution. McKinsey surveyed 1,719 people in 97 countries between May and June 2026: 89% of their organisations regularly use AI in at least one function, and 44% say it is scaling across the enterprise. Among companies with more than $1 billion in revenue, 40% are scaling AI agents, up from 27% a year earlier.

The results do not. Only 37% of McKinsey's respondents report any positive effect on operating profit, essentially unchanged from 2025, even though 80% say AI has made them personally more productive. McKinsey's "high performers", organisations that attribute at least 5% of operating profit to AI and call its value significant, are about 6% of the sample, also unchanged.

BCG's study of 1,250 companies draws the same picture from another angle. In a year, firms moved out of the bottom stages into "scaling", where some returns appear, but the top barely moved, from 4% to 5%: one company in twenty is "future-built", generating value at scale, while six in ten still get little or none. The middle is filling up faster than the summit. That matters because the summit is where the money is. The future-built firms report 1.7 times the revenue growth and 1.6 times the operating margin of the laggards, and 3.6 times their three-year shareholder return.

The most quoted figure, that 95% of enterprise AI pilots deliver no measurable profit impact, comes from MIT's NANDA initiative in July 2025. It is a small study, built on interviews and surveys of a few hundred people and reviews of about 300 public deployments, and its published sample sizes vary between versions, so treat its precision with care. Its diagnosis is more useful than its headline: the tools work for individuals but stall in companies because they do not learn from or adapt to the workflow around them.

Measured from outside, the picture is even more modest. The US Census Bureau found 19.8% of American businesses using AI in May 2026, a narrower measure of current use than McKinsey's, and 37% of firms with 250 or more staff. A study linking Danish survey data to payroll records, published by the National Bureau of Economic Research, found AI chatbots saved workers about 3% of their time, with no detectable effect on earnings or hours. Claims of fivefold productivity describe what one person with a good tool can do on one task. They do not describe what has happened to companies.

Every serious framework finds the same three tiers

Consultancies, universities and analysts have each drawn their own map of enterprise AI, and the maps agree. BCG has four stages, from stagnating to future-built. McKinsey sorts respondents into high performers and the rest. MIT NANDA calls the split the "GenAI Divide" between the few who cross it and the many who do not. MIT's Center for Information Systems Research has four maturity stages of its own, and finds companies in the first two perform below their industry's financial average while those in the last two perform well above it.

Laid side by side, they describe three tiers.

AI-native companies were built on AI from the start, so they have nothing to transform. Their economics are startling. The median private software company earns about $193,000 of recurring revenue per employee and the median listed one about $395,000. Lovable reported roughly $400 million in annual recurring revenue with 146 staff in early 2026, about $2.7 million per person, and outside estimates put Cursor in the same range. These are company-reported or estimated figures and headcounts move fast, but the order of magnitude is not in doubt.

Rebuilders are established companies that have redesigned how they work: BCG's 5%, McKinsey's 6%, and MIT's later stages. They are rare, they are pulling away, and they are the tier with lessons for everyone else. Their job is harder than an AI-native's because they carry legacy systems, contracts, regulators and thousands of people whose roles were designed around the old process. What marks them out is less technical brilliance than willingness to treat the operating model itself as something to redesign.

The bolt-on majority is everyone else, roughly 60% getting little and another third scaling slowly. They have AI everywhere and changed little. BCG describes them in a sentence: too many companies approached AI "as a way to do more of the same a little faster or better", when the need was reinvention.

These maps deserve some scepticism. Nearly every value figure in them is executives rating their own companies, and McKinsey's high-performer group is just 92 respondents, so the link between redesign and profit is a strong correlation, not a proven cause. And the bolt-on majority's biggest risk may not be its own slow rebuild at all, but an AI-native rival that never needed one.

The rebuilders change the work, not the toolset

The clearest finding in McKinsey's 2026 data is about workflows. Nearly three-quarters of high performers say they have fundamentally redesigned workflows because of AI, up from 55% a year earlier; among everyone else the figure is about one in four. High performers are also 3.3 times more likely to intend to use AI to transform their business within three years, and twice as likely to say senior leaders genuinely own the work.

Four traits recur across the rebuilders: agents with people at the exceptions, ownership at the top, narrow targets and shared data.

BNY shows what that looks like inside an old institution. The custody bank, founded in 1784, did not hand its staff a chatbot and wait. It built an AI platform called Eliza, and by 31 March 2026 it ran about 140 "digital employees" on work such as payment validation, each one doing a defined job alongside human staff. The agent runs the routine work; the people keep the exceptions and the judgement. Sierra, the customer-service company, builds the same idea into its price: it charges nothing when its agent hands a case to a person, so the human exception is part of the design rather than a failure.

Behind every case like this sits a decision made at the top. BCG's future-built firms run AI as a multiyear programme with cost and revenue targets set by top management, not as a collection of IT projects. Shopify's chief executive, Tobi Lütke, made it a rule in April 2025: teams asking for more headcount must first show why they cannot get the work done with AI, and AI use entered performance reviews. The rebuilders also choose their targets narrowly. BCG finds 70% of AI's potential value sits in core functions such as sales, manufacturing, supply chain and pricing, and its leaders concentrate on a few of those workflows rather than automating "here and there".

Underneath sits the data. BCG's leaders share "a data foundation" with "governed access to trusted enterprise data", because an agent can only act on what it can reach and trust. Some go further and change the organisation itself: in 2025 Moderna merged its human resources and technology departments under one executive, Tracey Franklin, so that the company plans work rather than separately planning people and systems.

So the answer to what works is a rebuild in layers: a few core workflows redesigned end to end, owned from the top, running on clean shared data, with people in charge of the cases an agent should not decide.

Everyone else stalls on people and data

Ask companies what holds them back and the models are not at the top of the list.

Of the fifteen challenges BCG asked about, the most-named was managing unstructured data, at 79%, followed by getting people to adapt to AI in their daily work and a shortage of AI talent. Security risks and wiring AI into existing systems came next. Complaints about the models themselves, hallucinations and accuracy, sat in the middle of the pack. BCG's long-standing rule of thumb puts 70% of a transformation's focus on people and processes, 20% on technology and only 10% on algorithms.

One result cuts the other way. "Lack of leadership support" was the least-named roadblock of all fifteen, at 42%, even though BCG's own analysis names weak top-management commitment as a leading cause of failure. Leaders believe they support AI. Supporting it is not the same as owning the redesign.

Four failure patterns recur.

Leaders delegate it. Among laggards, BCG writes, top management "may talk the talk" but hands AI to middle managers "who are unsure what to do or are fearful of the technology's future impact on them". The boardroom is not much better placed: in BCG's May 2026 survey of 625 chief executives and directors, 75% of directors rated their AI knowledge at least on par with their peers, while nearly 40% of chief executives said their boards lack an informed view of how AI is reshaping growth. Ownership is split too: 47% of chief executives said they lead AI implementation themselves, but only 39% of directors said their chief executive does. When the people at the top cannot judge what the technology can do, AI gets filed as an IT topic, and IT topics rarely redesign a sales process or a claims department.

The data is not ready. Gartner predicted in February 2025 that, through 2026, organisations would abandon 60% of AI projects not supported by AI-ready data.

Pilots multiply instead of compounding. BCG describes laggards spreading resources over scores of workflows, producing "a proliferation of disconnected initiatives". Gartner expects more than 40% of agentic AI projects to be cancelled by the end of 2027, many of them experiments driven by hype.

Automation is judged on cost alone. Klarna froze hiring, let its workforce shrink 22% to about 3,500 and leaned on an AI assistant for customer service. In May 2025 its chief executive, Sebastian Siemiatkowski, conceded that with cost "a too-predominant evaluation factor", the result was "lower quality", and the company began hiring human agents again.

One more pattern runs against instinct. MIT NANDA found that buying from specialised vendors and partnering succeeded about 67% of the time, while internal builds succeeded about a third of the time, half the vendors' rate, and that the best returns came from back-office automation even though more than half of budgets went to sales and marketing tools.

What does not work, then, is the default path: tools on unchanged processes, AI delegated downwards, dozens of pilots, cost as the only measure, and building everything in-house without people who have done it before.

Think electrification, and judge advisers by what they build

The most useful way to think about enterprise AI comes from economic history. In 1987 Robert Solow remarked that computers were everywhere except in the productivity statistics. In 1990 the economist Paul David took up the puzzle and found his answer in electricity.

Central generating stations opened in London and New York in 1882, yet by 1899 electric motors supplied less than 5% of the power driving American factory machinery. Where owners did adopt electricity, many simply ripped out the steam engine and installed a large motor in its place, keeping the old layout built around one central shaft, and productivity barely improved. The gains came when factories were rebuilt on a single floor around the flow of work, with a small motor on each machine and workers trained to run them. Electric motors passed half of factory drive capacity only around 1919. By Warren Devine's count, manufacturing labour productivity grew about 1.3% a year before 1919 and 3.1% a year after it, and the economic historian Paul David dates the surge to the 1920s, some four decades after the first power stations.

The parallel is close, though not exact: factories were held back by the slow turnover of physical plant, while software can be rebuilt faster. A chatbot added to an unchanged process is the motor in the steam-era factory. The rebuilt factory is the workflow redesigned around agents, on a data foundation, with people at the exceptions. Like electrification, it is a change programme: hard, slow, but entirely achievable by a company that treats it as one.

That frame also cuts through the noise, and there is a great deal of it. Deloitte agreed in October 2025 to refund part of a A$440,000 report for the Australian government after it was found to contain fabricated references and a misattributed court quote; a revised version disclosed that generative AI had been used to write it. Meanwhile the industry is quietly changing who does the work. Postings for "forward-deployed engineers", technical staff who sit inside a client and build until the system works, rose from fewer than 200 in 2023 to more than 7,300 in the first nine months of 2026, according to the labour-market analysts Lightcast. Deloitte and Accenture were among the biggest recruiters for the role, and Caylent's Val Henderson summed up why the role is scarce: "There is no compression algorithm for experience."

The deeper shift is that building has become cheap. A working agent, an internal tool or a prototype of a redesigned workflow can now be built in days by a small team using AI coding tools, which changes what good advice looks like. A strategy deck about AI transformation competes with a working system that a client can test the following week. MIT NANDA's finding that partnerships with specialised builders succeed about twice as often as companies building alone points the same way: the scarce skill is knowing how to make AI work inside a real process, and much of that lives in the scaffolding around the model, the subject of our AI harness episode, number 48, released on 25 August 2026.

To get past the noise, ask four questions of any claim: did a workflow actually change, or was a tool added; does the result show up in profit, not just in time saved; who at the top owns it; and has the person advising on it built and run an agent in production? Claims about productivity multiples and agent counts fail the first two almost every time.

What it means for the economy, workers, politics and the labs

For the economy, the history suggests patience and the data agrees. Rebuilding work is slow, so the productivity statistics will lag the hype, as they did for electricity and computers. The gains will also concentrate: BCG's leaders plan to spend more and expect double the revenue gains of laggards, so the gap is likely to widen before it narrows. Agents raise the stakes: BCG estimates they produced 17% of the value companies got from AI in 2025 and expects 29% by 2028, and agents only pay off where a workflow has been redesigned for them to run.

For workers, the change is arriving through the side door. MIT NANDA found companies "increasingly not backfilling positions as they become vacant", especially in customer support and administration, rather than announcing mass layoffs. That is quieter than the forecasts and harder to see in unemployment data.

For politics, the danger is the shape of the gains rather than their size. If a few rebuilt firms capture most of the value while everyone else adds tools and quietly stops hiring, the result looks like the deal voters already distrust, as our AI backlash episode, number 74, released on 27 October 2026, found: people use AI more every month but reject gains that seem to go to a few.

For the labs, the enterprise bottleneck is now their problem. OpenAI launched a deployment company in May 2026 backed by more than $4 billion. Anthropic is putting $100 million into training 10,000 forward-deployed engineers by the end of 2027, with first cohorts including engineers from Accenture, Deloitte, McKinsey and Morgan Stanley, and Amazon Web Services has committed $1 billion to its own programme. They are paying to do the rebuilding because their revenue depends on it.

The hinge is speed. If AI makes building cheap enough that a mid-sized company can rebuild a workflow in weeks rather than years, the electricity lag compresses and the bolt-on majority catches up. If rebuilding stays scarce, slow and dependent on a small pool of experienced builders, the gap between the 5% and the rest becomes a lasting divide. The share of companies redesigning workflows, now about one in four outside the leaders, is the number to watch.

Bottom line

Enterprise AI is not failing; most enterprises are using it the way factory owners first used electricity. The few that profit rebuild the work around it, starting with the data and ending with people handling the exceptions, and they treat it as the change programme it is. Everyone else is paying for the motor and keeping the line shaft.

Sources

Transcript

Sam: A company froze hiring, let its workforce shrink by about a fifth, and leaned on an AI assistant for customer service. Then its CEO said cost had become a too-predominant factor, quality fell, and they hired humans back.

Alex: That's Klarna, and I've signed off on a project with exactly that logic. Cheaper first, think later. Enterprise AI: Most Firms Bolt It On. The Winners Rebuild.

Sam: Before we get into it, I need to tell you what landed in my inbox on Monday. People Inc. is building custom feeds just for AI agents. Plain text, with who wrote it and when it was updated stamped on top.

Alex: Why would a publisher bother?

Sam: Their chief innovation officer, Jon Roberts, told Digiday that nearly all of a web page is junk meant for a browser. Agents don't want the junk. And Time and Forbes are doing versions of the same thing.

Alex: Huh. So the thing that was invisible to my engineers for twenty years, the page itself, is now the thing they're rebuilding.

Sam: Right, and it's my desk. If an agent is the one reading my brand's content, my brand voice has to survive being read by a machine. I don't have a plan for that yet. I have a Slack channel with a lot of question marks.

Alex: You always have a Slack channel with a lot of question marks.

Sam: That's the job, Alex.

Alex: Welcome to Dan's AI Intel. We take the one question about the AI revolution that actually matters, and try to get past the hype and the... the fear, I guess, to what's going on.

Sam: Past both, yes. And what it means for the economy, for politics, for the labs, not just the tech. I'm Sam. I run brand and lifecycle marketing at a big online pet retailer, so I'm the one who signs off the AI tool budget and knows what each line costs.

Alex: And I'm Alex. I'm the CTO at an activewear brand, and I still write code at weekends, which my team has asked me to stop mentioning.

Sam: Nobody asked you to stop. Your engineers asked you to stop merging it, and they've told everyone.

Alex: Fair. So here's what we're opening up. Nearly every company says it's using AI. The money, though, is landing somewhere very specific. What separates the firms where it lands from the firms where it doesn't?

Sam: And is it the technology, or is it something dull and human? We'll go through what the big surveys find, what the winners actually did, what stalled everyone else, and then a history lesson from a hundred years ago that I did not expect to care about.

Alex: And it matters now, because the labs' valuations assume companies change how they work, and software shares fell sharply in February on a bet that agents threaten seat-based software. If companies don't convert, that's a very expensive data-centre bill.

Sam: There's a twist about who's now getting paid to do the rebuilding, too. Hit follow wherever you're listening, it's free, and it means the next one shows up on its own. Before the numbers, let's put our cards down. Where do you start on this, Alex?

Alex: My honest starting position is that a lot of this is just early. I run engineering. I've watched my people go from skeptics to using coding agents every day in about a year. Give companies more time and better tools and the profits follow. And software isn't a factory. You can rebuild a workflow in days, so the lag doesn't have to be long. The tools are the bottleneck.

Sam: See, I think that's the exact story every vendor wants us to tell. I've had these tools on my team for two years. My copywriters are faster. Nobody can show me where the time went. It didn't become profit. It became more drafts.

Alex: That could be a measurement problem.

Sam: Or it could be that we gave everyone a faster pencil and left the process exactly where it sat. I think the part that matters isn't the tool. It's who changed the work, and whether anyone owns that.

Alex: So I'm saying wait for the better motor. You're saying the motor was never the problem.

Sam: Hold that thought. You're going to hear that word again.

Alex: I'm suspicious already.

Sam: You should be. Let's see who's right.

Alex: Start with the biggest survey. McKinsey ran a huge global survey this spring. Eighty-nine percent of their organisations use AI in at least one function.

Sam: That's the headline I'd put on a slide. Which is exactly why I don't trust it by itself.

Alex: Fair. So next number. Only thirty-seven percent say it moved operating profit at all. And the group McKinsey calls high performers, firms that put at least a twentieth of their operating profit down to AI, is about six percent.

Sam: Six percent. Out of nearly everybody.

Alex: And most of those same people say AI made them personally more productive.

Sam: Right, that's the pencil again. Everybody writes faster. Nobody's profit moved.

Alex: Boston Consulting Group studied a big panel of companies and got a similar shape from a different angle. Five percent are what they call future-built, getting value at scale. Sixty percent get little or nothing.

Sam: And the middle is filling up faster than the top. They moved out of the bottom stage, but the summit barely moved.

Alex: That's what I'd put to my CEO, because the summit is where the money is. Those top firms report faster revenue growth, better margins, and a far better shareholder return over three years.

Sam: That I'll buy, because it's a gap you can see in a result, not just in a feeling.

Alex: So you take the skeptic's side then. What's the catch?

Sam: The catch is, who's rating whom. Most of those value figures are executives scoring their own companies. I don't trust anything self-rated, and the vendor-sponsored ones least. I've been sold both kinds.

Alex: That's a stance you've held.

Sam: And I'll defend it. So let me bring in what you can measure from the outside. The US Census Bureau counted just under one in five American businesses using AI in May, and bigger firms are higher but nowhere near McKinsey's number.

Alex: Which is far lower than McKinsey's.

Sam: Because it's a narrower question. Different asks, different answers. And then there's a Danish study, linking survey answers to payroll records, published by the National Bureau of Economic Research. Chatbots saved workers about three percent of their time. No detectable effect on earnings or hours.

Alex: Three percent. That is not the fivefold productivity people quote.

Sam: Because the fivefold story is one person with a good tool on one task. It's not what happened to a company. When a vendor says five times in my budget meeting, I ask five times what.

Alex: Hm. I want to push on that. Three percent was an early chatbot. My engineers are not using early chatbots.

Sam: Agreed. And that's the whole fight, isn't it. Is it the early tool, or the unchanged work around it?

Alex: Right. Don't let me wriggle out of it.

Sam: I won't. I've got a spreadsheet for you wriggling.

Alex: Use is easy to count, though. It doesn't prove anything.

Sam: No, and it gets weird fast. My daughter Maya came home asking whether she should make her college essays sloppier so a detector won't flag her as an AI. And in a survey of job seekers, about sixty percent changed or thought about changing their applications so they wouldn't read as machine-written.

Alex: So everyone uses it, and everyone's hiding it.

Sam: And I'm supposed to put "AI adoption" on a slide for my CMO as a win. It's a bad scoreboard.

Alex: Then there's the number everybody quotes. Ninety-five percent of enterprise AI pilots deliver no measurable profit. That's from MIT's NANDA initiative, July last year.

Sam: And it's the number I quote at budget meetings when I want a line removed.

Alex: Which is a bad habit. It's a small study. Interviews and surveys with a modest number of people, plus a look at public deployments, and the sample sizes shift between versions.

Sam: So the precision is soft.

Alex: The precision, yes. The diagnosis is better. The tools work fine for individuals, but they stall in companies because they don't learn from, or adapt to, the workflow around them.

Sam: Which is the pencil, again.

Alex: Which is the pencil. And that diagnosis is one my own record has some scars on, because I've bought tools that were great in a demo and invisible in the order system.

Sam: Name one.

Alex: No.

Sam: The legal department would like a word.

Alex: They would.

Sam: So everyone has a map. BCG has four stages, from stagnating to future-built. McKinsey splits high performers from the rest. MIT NANDA calls it the GenAI Divide. And a separate MIT group, the Center for Information Systems Research, has its own four stages.

Alex: Which, laid side by side, say the same thing. Three tiers, and I've sat in the wrong one in a board meeting.

Sam: Let me guess first where my own place sits. Tier three. We have AI in every corner and I can't point to a process that's different.

Alex: Tier one, AI-native companies. Built on AI from day one, so there's nothing to transform. Their economics are strange. A typical software company earns a few hundred thousand dollars of recurring revenue per employee. Lovable reported roughly two point seven million a head.

Sam: Per person. I'd settle for my team shipping one thing a quarter without a committee.

Alex: Those are company-reported or estimated, headcounts move fast, but the order of magnitude isn't in doubt.

Sam: And tier two is where I'd like to be and probably can't afford to be.

Alex: The rebuilders. Established companies that redesigned how they work. BCG's five, McKinsey's six. They carry legacy systems and contracts and regulators and thousands of people whose jobs were designed around the old way. That's why they're rare.

Sam: And tier three is us.

Alex: Tier three is most of us. The bolt-on majority. AI everywhere, work barely changed. BCG's line is that too many companies treated it as a way to do more of the same a little faster.

Sam: A little faster. That's every presentation I've given about AI.

Alex: Including the ones at my board.

Sam: Especially those. There's a caution on the maps, though. Nearly all of this is executives rating themselves, and McKinsey's high-performer group is a small slice of the sample. So the link between redesigning work and making profit is a strong correlation. It isn't proof.

Alex: Which I'm happy to hold. But here's the part that keeps me up. The bolt-on majority's biggest risk may not be its own slow rebuild at all. It may be an AI-native rival who never had to rebuild anything.

Sam: That lands, and it scares me. A competitor with no thirty-step approval chain to defend.

Alex: Same for me. A rival store that never had my legacy order system.

Sam: And I'd face a pet-food brand that doesn't carry years of how we send emails.

Alex: Which brings us to the ones who rebuilt anyway. Here's the cleanest finding in McKinsey's data. Nearly three-quarters of high performers say they've fundamentally redesigned workflows because of AI. Everyone else, about one in four.

Sam: Everyone I manage has the tool, and not one job description has changed a word.

Alex: And those high performers are three times more likely to say they intend to transform the whole business within three years. And twice as likely to say senior leaders genuinely own the work.

Sam: Owning it is the phrase I keep circling. Own versus sponsor.

Alex: Engineering owns things. Marketing sponsors them. I'd say that's the difference, not the word.

Sam: Ha. No. Sponsoring is a slide. Owning is a person whose bonus changes if the process doesn't, and in your building that's a pager, not a bonus. At my place, I'd say we sponsor.

Alex: Mine too, if I'm straight about it. I get paged, but nobody's pay changes if the process doesn't.

Sam: Good. Now give me the real example before I get too smug.

Alex: BNY. The old custody bank. It built its own platform, called Eliza, and by the end of March it was running about a hundred and forty digital employees, each with a defined job, on things like payment validation.

Sam: Digital employees. Hm. That's marketing language for agents, isn't it.

Alex: It is, a bit. But the design underneath is real. The agent does the routine work, and the people keep the exceptions and the judgment. I trust a case like this because it has real numbers behind it. The ones that only talk, I tune out.

Sam: Which fits Sierra, the customer-service company. They charge nothing when their agent hands a case to a human.

Alex: Right, so the handoff isn't a failure, it's priced in. Mine bills me for the bot either way.

Sam: Which is what my customer care team would actually like to hear. Nobody there wants to be the thing the bot gives up on.

Alex: That's the part I love. Elegant. Okay, I can feel a tangent about 1999 coming on.

Sam: No.

Alex: A very short one.

Sam: Shopify instead. Please.

Alex: Fine. Shopify. In April last year its CEO, Tobi Lütke, made a rule. Any team asking for more headcount first has to show why the work can't be done with AI. And AI use goes into performance reviews.

Sam: I'd get a revolt if I tried that rule on my copywriters. Harsh, until it works.

Alex: It's an ownership decision made at the very top. BCG's future-built firms do the same. They run it as a multiyear programme, with cost and revenue targets set by top management, not a pile of IT projects.

Sam: And they pick narrow. BCG says the bulk of the value sits in core functions like sales, supply chain and pricing. The leaders concentrate on a few workflows. They don't automate here and there.

Alex: Under all of it, data. Governed access to trusted enterprise data. An agent can only act on what it can reach and trust. That's my data director's whole year, and it's the line on my budget the CEO likes least.

Sam: And Moderna went further. Merged human resources and technology under one executive, Tracey Franklin, so the company plans the work instead of planning people and systems separately.

Alex: I would have hated that, as an engineer, in 2012.

Sam: And now?

Alex: Now I think it's smart. Planning people and systems apart is exactly how you end up with a pencil.

Sam: You've given yourself a new nickname.

Alex: Yes. Pencil guy. So the recipe. A few core workflows redesigned end to end. Owned from the top. Running on clean shared data. With people in charge of the cases an agent shouldn't decide. The directors want the demo and none of the rest of that list.

Sam: And here's where it hits my desk. People Inc. rebuilding its content for agents, the thing I opened with, is that recipe on a small scale. They didn't add a chatbot to the website. They rebuilt what the product is, for a reader that isn't human.

Alex: And they cut about six percent of staff the week before that came out.

Sam: Which is the half I'd have to explain to my copywriters. A rebuild like that moves people around, and I'd be the one in the room.

Alex: That's the part I skip when I pitch this to my executive team. It isn't a nicer process.

Sam: A nicer process is what I want. A rebuild is what I'd have to explain to my team.

Alex: Yeah.

Sam: So why does everyone else stall? And this one's the part I like, because it's not about the models. BCG asked about fifteen challenges. The most-named was managing unstructured data, seventy-nine percent.

Alex: Second, getting people to adapt to using AI daily. Third, a talent shortage, which is my hiring board in one line.

Sam: And model complaints sat in the middle. Hallucinations, accuracy.

Alex: Which flips what I assumed. I came in thinking the model was the main issue.

Sam: BCG's old rule of thumb puts the bulk of the effort on people and process, and a sliver on algorithms.

Alex: A sliver on the thing my engineers argue about all day.

Sam: One result cuts the other way. "Lack of leadership support" was the least-named roadblock of all fifteen. Forty-two percent.

Alex: But BCG's own analysis names weak top management commitment as a leading cause of failure.

Sam: So leaders believe they're supporting it. Supporting isn't owning the redesign. I've sat in that meeting and nodded.

Alex: Which is that same word. Among the laggards, BCG says top management may talk the talk but hands AI to middle managers who are unsure what to do, or afraid of what it'll do to them.

Sam: I have been that middle manager. Handed a thing with a deadline and no authority to change the process around it.

Alex: There's a boardroom version, too. BCG surveyed CEOs and directors in May. Most directors rated their own AI knowledge at least on par with peers, while a lot of CEOs said their boards lack an informed view.

Sam: And the CEOs say they lead AI implementation personally, but fewer directors say their CEO does.

Alex: They can't agree on who's driving. I'd hate to walk into that boardroom with my budget ask.

Sam: So it's filed under IT, and IT topics rarely redesign a sales process or a claims department. That's where my last two AI projects went to die, Alex. Your side of the building.

Alex: That sentence should be on a mug.

Sam: I'll expense it.

Alex: And it gets worse than leadership. Let's do the whole list. Data first. Gartner predicted, back in February last year, that through this year organisations would abandon most AI projects not supported by AI-ready data.

Sam: Pilots next, and I've got a graveyard of those. BCG describes laggards spreading effort across scores of workflows, "a proliferation of disconnected initiatives". I've paid for four tools that each did one thing for two weeks. And Gartner expects a large share of agentic AI projects to be cancelled by the end of next year.

Alex: Many of those experiments driven by hype, as Gartner puts it.

Sam: Then cost. Klarna froze hiring, let its workforce shrink by about a fifth, and leaned on an AI assistant for customer service.

Alex: Then in May last year its CEO, Sebastian Siemiatkowski, said that cost had become "a too-predominant evaluation factor". The result was lower quality. The company started hiring human agents again.

Sam: Okay, so I'll say what I'd say as a customer. A wrong answer from a bot is worse than no answer, because I blame the retailer. Not the bot.

Alex: There's a number behind that?

Sam: An industry exec said it about shopping agents this week, that a wrong answer is worse than none and the shopper blames the retailer. And the same note says most commerce organisations haven't fully defined what success even looks like for their AI.

Alex: So they're shipping without a dashboard. I wouldn't let my team do that for a week.

Sam: Flying with a very enthusiastic slide.

Alex: And one that runs against my instincts. MIT NANDA found that buying from specialised vendors and partnering succeeded about two times in three. Internal builds, about one in three. Half the rate.

Sam: And you build things.

Alex: I build things. I would like it noted that I took that personally.

Sam: Noted.

Alex: They also found the best returns came from back-office automation, even though most of the budgets went to sales and marketing tools.

Sam: Which is my side of the building. I'd like to argue, but I've seen my tool budget.

Alex: And my renewals list. A long wall of SaaS contracts, and I couldn't tell you which ones an agent could replace. Time for the history lesson, and I'm going to enjoy this far too much.

Sam: I've been warned.

Alex: Back in the eighties, the economist Robert Solow said you could see computers everywhere except in the productivity statistics. A few years later Paul David took up that puzzle, and his answer was electricity.

Sam: Electricity. The thing that's in the walls.

Alex: Central power stations opened in London and New York in the 1880s. Nearly two decades later, electric motors still supplied a sliver of the power driving American factory machinery.

Sam: Obviously better, though. So why so slow?

Alex: Because many owners simply ripped out the steam engine and put in a big electric motor, keeping the old layout built around one central shaft. Productivity barely moved.

Sam: Hold on. That's the pencil.

Alex: That's the pencil. Then eventually somebody rebuilt the factory on a single floor, with a small motor on each machine, and workers trained to run them. Electric motors passed half of factory drive capacity only around the end of the First World War. Warren Devine counts manufacturing labour productivity growth more than doubling after that. The surge lands in the 1920s, roughly four decades after the first power stations.

Sam: That long. And I thought my approval chain was slow. So come back to the bet. You said wait for the better motor.

Alex: And the factory story says the motor was fine. A better motor in the same shaft layout is still a shaft layout. I came in thinking time and better tools would fix it, and I'd have said that with full confidence an hour ago.

Sam: A full hour.

Alex: I'm owning it. Fair, I deserved that. The waiting wasn't for technology. It was for somebody to redraw the floor. That part moved me.

Sam: But you're keeping something.

Alex: I'm keeping the factory difference. They were held back by how slowly physical plant turns over. Software doesn't have that. So the lag can compress in a way it couldn't in 1920.

Sam: And I'll give you something real back. I'd been treating "we're slow" as a fact about us. It's partly an excuse. If building is cheap, nobody in my building gets to say the redesign is too expensive to try.

Alex: You were right about the order, though. Floor first, then motor. I'm not putting that in writing.

Sam: It's being recorded.

Alex: Naturally.

Sam: Then there's the noise. There is so much of it. Deloitte agreed last October to refund part of the fee on an Australian government report after it was found to contain fabricated references and a misattributed court quote. A revised version disclosed that generative AI had been used to write it.

Alex: That makes me wince, because I've bought that kind of advice. Meanwhile it's quietly changing who does the work. Postings for forward-deployed engineers, technical people who sit inside a client and build until the system works, went from a couple of hundred in 2023 to over seven thousand this year, according to Lightcast.

Sam: Deloitte and Accenture are among the biggest recruiters for the role, which is a little funny.

Alex: Caylent's Val Henderson said why it's scarce. "There is no compression algorithm for experience."

Sam: I'd put that on the mug next to the other one.

Alex: There's a deeper shift here. Building has become cheap. A working agent or a prototype of a redesigned workflow can be built in days by a small team using AI coding tools. Which changes what good advice looks like.

Sam: And a deck costs me what a pilot costs me, and tells me less than something I can test next week.

Alex: And I'll be honest. A deck I paid for last year is in a folder, and a prototype my team built in a week is in production.

Sam: Shocking. The thing you paid for with no builders in the room did less.

Alex: Mm.

Sam: So give me something I can use on Monday. Four questions to ask of any claim.

Alex: Mine I'd ask a vendor first. Did a workflow actually change, or did you add a tool? I've sat through enough demos to know the answer is usually the tool.

Sam: Mine is about money. Does it show up in profit, not just in time saved? Because time saved is what I get invoiced for.

Alex: Third, who at the top owns it. That one's uncomfortable for both of us.

Sam: And four goes to the consultants. Has the person advising you built and run an agent in production?

Alex: Every pitch deck I get that brags about productivity multiples and agent counts fails the first two.

Sam: So "five hundred agents deployed" gets a no.

Alex: Unless a workflow changed and the profit shows.

Sam: Good. I'll print it for my next vendor meeting.

Alex: I'll take the fourth to my next adviser call.

Sam: Please do. I'd like to hear the silence. So what does this do to the budget I'm about to defend? If the payoff is slow, my line for AI tools doesn't come back in a quarter. The productivity statistics will lag the hype, the way they did for electricity and for computers, and that's the sentence I'd have to put in front of my CMO.

Alex: And the gains won't spread evenly. BCG's leaders plan to spend more and expect double the revenue gains of the laggards, so the gap probably widens before it narrows.

Sam: That's the part I'd hate to explain to my CMO.

Alex: Agents raise the stakes. BCG estimates they produced about a sixth of the value companies got from AI last year and expects closer to a third within three years. But they only pay off where somebody redesigned the workflow, and I've got agents in my shop with no workflow to run in.

Sam: Which brings it to people. MIT NANDA found companies "increasingly not backfilling positions as they become vacant", especially in customer support and administration. No layoff announcement, just a role that quietly doesn't come back.

Alex: And that's what changes my hiring plan. My daughter is a junior engineer at a payments start-up in Austin, fast with coding agents and scared the rung she's on is the one being automated. I used to plan junior hires by the dozen. Now I ask what a junior is for.

Sam: Because junior work is the most codified.

Alex: Which I said on episode seventy-three and still believe. Nobody fires her. They just don't hire the next one.

Sam: I've got the same cliff with my junior copywriters, and I'd rather find out now. And past my own desk, it's Maya's generation I think about. If a few rebuilt firms capture the value and everyone else adds tools and quietly stops hiring, she graduates into exactly the deal voters already distrust.

Alex: It's what we found on episode seventy-four. People use AI more each month, and they reject gains that seem to go to a few. I'd hate to be wrong about speed and have that be the result.

Sam: Hold that, because the labs know it. The enterprise bottleneck is now their problem.

Alex: Right. OpenAI launched a deployment company in May backed by more than four billion dollars. Anthropic is putting money into training ten thousand forward-deployed engineers by the end of next year, and Amazon Web Services has its own programme.

Sam: They're paying to do the rebuilding because their revenue depends on it. I'd like to know whether their engineers end up in my building or on my invoice.

Alex: Which, to me, is the cleanest single sign that the bottleneck isn't the model, and it's the opposite of what the model vendors tell my board.

Sam: Or the labs are buying their way into every company's operating model, and I want to know who pays for that.

Alex: Who pays for that is yours. You can have it.

Sam: Thank you. I already did.

Alex: Here's where I'd use the thing you gave me. The hinge is speed. If AI makes building cheap enough that a mid-sized company can rebuild a workflow in weeks, the electricity lag compresses, and the bolt-on majority catches up.

Sam: And if rebuilding stays scarce, slow, and dependent on a small pool of people who've done it before, the gap between the five percent and everyone else turns into a lasting divide.

Alex: I land on the first. Building got cheap. A small team can prototype a redesigned workflow in days now, and I've stopped being shocked by that.

Sam: And I land on the second, for my own building. The building was never the slow part. The slow part is the meeting where somebody has to say "this job is now different". Cheap prototypes don't give me that meeting.

Alex: But you're underrating the prototype. A week of building makes the argument for you. People stop debating a slide when they can click the thing.

Sam: Clicking it isn't owning it. Somebody has to give up a team, or a budget line, or a Tuesday ritual. Nobody gets promoted for shrinking their own job.

Alex: Some of my best engineers did exactly that and got promoted.

Sam: In your building. In mine, zero.

Alex: Fine, that's a real disagreement. You think it comes down to who owns it, and I think it comes down to how fast a thing can be shown.

Sam: Right. And the number to watch is the share of companies redesigning workflows. About one in four outside the leaders right now.

Alex: If that moves, I'm right. If it doesn't, you are.

Sam: I'll bring snacks either way.

Alex: So, the short version. Almost everyone uses AI and about one in twenty profits. The ones that do rebuild the work around it, on one governed data foundation, with agents running workflows and people handling the exceptions.

Sam: Everyone else stalls on people and data, not models. Leaders who sponsor and don't own, pilots that multiply, cost as the only measure.

Alex: And the electricity story says the waiting isn't for a better motor. It's for somebody to redraw the floor.

Sam: So ask four things. Did a workflow change, does it show in profit, who at the top owns it, has the adviser built.

Alex: Elena sent me a Pew survey this week, the one on Americans being open to AI in healthcare but wanting a human to make the final call. She's an emergency doctor, and she said it's the whole design argument in one line.

Sam: People keep the exceptions.

Alex: People keep the exceptions. It's why that BNY setup reads as sensible to me and the Klarna one reads as a warning.

Sam: And I still disagree with you about how fast it happens.

Alex: Good. Keeps the show interesting. If you came away understanding a little more of where this is heading, then the show did its job. What companies like ours do in the next few years decides a lot of how this turns out.

Sam: A quick note on how this is made. The show is AI-generated. Dan builds a custom stack of AI tools to research, analyse and verify these questions, mostly so he can learn the subject himself, and then shares it with anyone who finds it useful. Worth a second look at anything that matters to you, and not a substitute for a professional's advice on your own decisions.

Alex: And if this was useful, follow Dan's AI Intel in Spotify, Apple Podcasts, or whatever app you're using. It's free, one tap, and a follow is the single biggest thing that helps a small independent show grow.Sam: One last thing from us. If there's something here you'd push back on, or a question you want us to go deeper on, write to podcast@connectiveshift.com. Every message gets read, and what you send shapes what we research next. Thanks so much for listening.

Sam: See you next time.