Rewired: why AI hasn’t reached the bottom line, and how to build companies where it does

[PLATE 01]

Gartner expects worldwide spending on AI to reach about $2.7 trillion in 2026, half as much again as in 2025, and more than half of it on infrastructure¹⁰. The same release places generative AI “firmly in the Trough of Disillusionment” this year. Alphabet, Amazon, Meta and Microsoft spent roughly $410 billion on capital projects in 2025 and have guided to between $720 and $745 billion for 2026¹¹. In 2024 David Cahn of Sequoia Capital doubled the run rate of data-centre chip sales for the buildings and energy around them, doubled it again for the margin their eventual customers need, and found a gap of about $600 billion a year against AI revenue¹². The sum was rough on purpose. Somebody, eventually, has to sell something.

Whether ordinary firms are selling more can now be read from large surveys, and they agree. The broadest is a survey of nearly 6,000 senior executives in the United States, the United Kingdom, Germany and Australia, run between November 2025 and January 2026 through business panels operated with the Federal Reserve Bank of Atlanta, the Bank of England, the Deutsche Bundesbank and Macquarie University¹. Sixty-nine per cent of firms said they actively use AI. Asked about the past three years, 89% reported no change in labour productivity, which the survey measures as sales per employee, and more than 90% reported no effect on employment. Converted to numbers, the realised productivity effect averaged about 0.29% over three years. The same executives expect 1.4% over the next three.

McKinsey’s 2026 survey of 1,719 respondents in 97 countries shows the same shape from another angle³. Eighty per cent said AI had improved their own productivity. Thirty-seven per cent said it had contributed anything at all to their organisation’s operating profit (EBIT), about the same as the 39% of the year before¹³, even though more organisations reported scaling AI. About 6% qualified as high performers, attributing at least 5% of EBIT to it. PwC’s survey of 4,454 chief executives in 95 countries found that 56% had seen neither higher revenue nor lower costs from AI in the previous twelve months, and 12% had seen both².

A third source is quoted often and should be read with care. MIT’s Project NANDA reported in 2025 that despite $30 to 40 billion of enterprise investment, 95% of organisations were “getting zero return”¹⁴. The figure rests on interviews at 52 organisations, 153 survey responses and a review of about 300 public initiatives, and the authors call their deployment figures “directionally accurate”. It is a useful signal and a weak measurement, and it points the same way as the larger surveys.

None of this says AI does not work. It says the benefit lands on individuals and does not reach the accounts. The cost of AI arrives as an invoice. The benefit, so far, arrives mostly as a feeling. A survey can record a feeling; an income statement cannot.

The cost of AI arrives as an invoice. The benefit, so far, arrives mostly as a feeling.

There are places where AI clearly pays. They are worth studying for what they share rather than for their headline numbers.

The firmest evidence comes from customer support. Erik Brynjolfsson, Danielle Li and Lindsey Raymond followed 5,172 agents at a Fortune 500 business-software company, across about 3 million chats, as a generative AI assistant was introduced⁴. Resolutions per hour rose 15% on average and about 30% for the least experienced agents, with little effect on the most skilled. (Earlier working-paper versions reported 14% and 34 to 35%, the figures still most often quoted.) Customers wrote more politely and asked for a manager less often, though survey-based satisfaction did not change. In a controlled experiment on professional writing, Shakked Noy and Whitney Zhang found that ChatGPT cut the time taken by 40% and raised graded quality by 18%¹⁵. Among 758 BCG consultants, those using GPT-4 on tasks inside the model’s competence finished 12.2% more tasks, 25.1% faster and at markedly higher quality¹⁶.

Klarna is the corporate case most often cited. According to its annual report for 2025, four in five of its customer-service chats that year went to the AI assistant, which did the work of more than 850 full-time agents and saved about $59 million, with no fall in satisfaction on internal surveys¹⁷. The same filing describes a “dual-track” approach that keeps human agents available, nine months after the chief executive said the push to cut service costs with AI had “gone too far”¹⁸. The more telling number sits elsewhere in the filing: customer service and operations costs rose 2% in 2025 while volumes rose 32%. That is output growing faster than cost, which is the kind of return this essay is about.

Look at what these cases share. The task is bounded: one conversation, one document. The work already leaves a record, so the model can be given context and its output checked. The firm already counts the result, so the gain is visible without new measurement. And the output drops straight into an existing flow: a resolved ticket is simply closed. Remove any of the four and the gain becomes hard to find.

There is a less comfortable feature too. Most of these gains are cost savings, and savings that every competitor can buy do not stay with the firm that makes them. If every support operation rents the same assistant, faster resolution becomes the industry standard and competition hands it to customers. That is good for customers and invisible in margins. Nicholas Carr made the argument about information technology in 2003: railways and electric power, once cheap and everywhere, “became commodity inputs”, and computing was going the same way. What makes a resource strategic, he wrote, “is not ubiquity but scarcity”¹⁹. The MIT NANDA report saw the other side of this. Its firmest savings came from cancelling outsourcing and agency contracts, and the tools “did not change team structures or budgets”¹⁴. The money moved; the firm did not.

[PLATE 02]

In 1987 Robert Solow remarked that the computer age was visible “everywhere but in the productivity statistics”²⁰. Three years later the economic historian Paul David offered an explanation by going back a century, to electricity⁶.

Electric motors were driving factory machinery by the 1880s, yet in 1899 they supplied less than 5% of the mechanical drive in American factories, and it took about two more decades to reach half. David found no visible effect on manufacturing productivity until the early 1920s, some four decades after the first central power stations opened. Then the effect was large: about half of the acceleration in manufacturing productivity growth between the 1910s and the 1920s can be accounted for statistically by the spread of electric motors. The delay, in David’s account, was organisational. Factories first laid the new system over the old, an “overlaying of one technical system upon a preexisting stratum” that he considered normal in such transitions. The overlay worked, after a fashion. It just did not pay.

Warren Devine’s history of factory power shows what the overlay looked like on the floor⁵. A steam-era factory ran one engine, which turned a long line shaft through the building, and every machine drew its power from the shaft by belt. The first electric factories swapped the engine for a large motor and kept the shafts, an arrangement called group drive. The saving was small, not least because fuel was only about 0.5% to 3% of the cost of a unit of output. The large gains came with unit drive, which gave each machine its own motor and joined the motors with wires. Without shafts, buildings could be single-storey and easy to change, with machines placed in the order of the work rather than the order of the power. Electric motors supplied about 5% of factory drive capacity in 1899, 25% in 1909, 53% in 1919 and 78% in 1929; output per hour of work in manufacturing grew 1.3% a year before 1919 and 3.1% a year after. Devine adds that the gains from unit drive “were not automatic”. Somebody had to redesign the factory.

[PLATE 03]

Economists call this a general purpose technology: one that spreads across sectors, keeps improving, and spawns further innovation in the industries that use it, with benefits that depend on complements built over years²¹. Brynjolfsson and Lorin Hitt found that the organisational investments that made computers pay could be “as much as an order of magnitude larger” than the spending on the computers²². Michael Hammer was blunter in 1990: “It is time to stop paving the cow paths”²³. Ford’s accounts-payable department employed more than 500 people when Mazda managed with five. Automating it would have saved about a fifth of the staff; redesigning the process so that invoices no longer had to be matched cut headcount by 75% where it was applied.

Read today, the parallel is uncomfortable. A licence for every employee is group drive: a better motor at the end of the same shafts. The shafts are the approval chain, the document passed from inbox to inbox, the weekly status meeting, the spreadsheet that reconciles two systems. Each person gets a little faster, and the work still moves at the speed of the shafts. Unit drive, in this analogy, gives each unit of work its own bounded capability, and that needs new wiring: a record the model can read, rules it must obey, and a way for it to act that can be checked and reversed. Section 6 is about that wiring.

A licence for every employee is group drive: a better motor at the end of the same shafts.

[PLATE 04]

If adopting firms are not capturing much of the value yet, someone else is. David Teece’s 1986 paper on profiting from innovation explains where to look²⁴. When an innovation is easy to imitate, the profits “may accrue to the owners of certain complementary assets” rather than to the innovator, and still less to the average user: manufacturing, distribution, service, whatever is scarce and hard to copy. RC Cola was first with cola in a can and first with diet cola; Coca-Cola and Pepsi, with their bottlers and their shelf space, followed almost at once and took the benefit. EMI invented the CT scanner, lost market leadership within six years of launching it in the United States and left the business soon after, while later entrants with stronger sales and service networks prospered.

Applied to AI, the stack has five layers: chips, cloud capacity, models, applications and the firms that adopt them. Returns today sit near the bottom, where supply is scarce. Nvidia reported revenue of $215.9 billion for its 2026 fiscal year, $193.7 billion of it from data-centre products, at a gross margin of 71%⁷. The cloud providers above it turn that hardware into rented capacity, and their capital spending is the demand that sustains those margins.

The model layer is less secure. The price of reaching a given level of capability keeps collapsing: Epoch AI found that the cost of matching a fixed benchmark score fell by between 9 and 900 times a year depending on the task, with a median of 50⁸, and the Stanford AI Index recorded a more than 280-fold fall in the price of GPT-3.5-level performance between November 2022 and October 2024²⁵. The frontier is still contested; the 2026 Index shows the gap between the leading closed and open-weight models widening again, to about 3.4% on one prominent ranking²⁶. A fair summary is that each level of capability becomes a commodity soon after it is reached, while the newest level stays, briefly, a product. Model companies can grow quickly, but they carry heavy commitments to the layers below and sell into a market where last year’s capability is this year’s discount.

For the firms at the top of the stack, Teece and Carr point the same way. A model that every competitor can rent is not an advantage. What an adopter can own is what the model needs and cannot supply for itself: the customer relationship and the right to serve it; the firm’s own record of what was done and what happened next; a track record that customers and regulators trust; and the workflows and distribution that put output in front of someone who pays. Those are the assets this essay’s design is meant to build.

Put a capable model into a firm designed for people and it meets four obstacles. None of them is about intelligence.

Authority. In most firms the right to act belongs to a role, granted slowly and recorded informally. A credit controller can extend a payment term; a buyer can raise a purchase order up to a limit. There is no mechanism for handing a model a small, measurable slice of that authority, so its output goes back into a human queue. Take a disputed supplier invoice. A model can draft the reply in seconds. Settling the dispute means reading the contract, checking the delivery record, applying the credit policy and posting a credit note, and the model has permission for none of it. The queue is as long as it was; the drafts are better written.

Context. What a firm knows is spread across a dozen systems, thousands of inboxes and the heads of long-serving staff. A model sees only what someone pastes in. MIT NANDA’s interviewees named “Too much manual context required each time” among the barriers to using these tools in core work¹⁴. Anthropic’s analysis of business use of its models reaches the same point from the supplier’s side: sophisticated deployments depend on curating the right context, and for some firms the data and organisational work needed to supply it may be the real bottleneck²⁷.

Verification. Output has to be checked by someone who could have done the work, and checking is not free. In METR’s randomised trial, experienced open-source developers working on their own repositories took 19% longer with AI tools, while believing they had been faster²⁸. BCG consultants using AI on a task outside the model’s competence were 19 percentage points less likely to reach the right answer than those working alone¹⁶. Consistency is a separate problem. Sierra’s τ-bench measures the chance that an agent succeeds on every one of k attempts at the same task; for the top agent tested in 2024 it fell below 25% at eight attempts in the retail domain²⁹. On 175 tasks set in a simulated software company, the top agent in 2025 completed about 30% on its own³⁰. Models have improved since, but the metric is the right one. A firm runs the same task thousands of times, and a process that works most of the time is a process with an incident rate.

Flow. In 1967 Gene Amdahl argued that speeding up the parallel part of a computation is “wasted unless it is accompanied by achievements in sequential processing rates”³¹. The organisational version is plain arithmetic. If drafting is a fifth of the elapsed time of a contract, infinitely fast drafting makes the contract at most 1.25 times faster. The rest is waiting: for approval, for a meeting, for the person who knows. The Danish evidence fits. Two years after ChatGPT’s release, Anders Humlum and Emilie Vestergaard could rule out effects on workers’ earnings or recorded hours larger than 2%, although most users said they saved time and spent it on other tasks³². Time saved inside a job is not output unless the job is redesigned to turn it into some.

Each obstacle is reasonable on its own. Together they explain why individual gains do not add up. The model saves a person forty minutes. The forty minutes dissolve into the same queue.

The model saves a person forty minutes. The forty minutes dissolve into the same queue.

[PLATE 05]

The alternative is not more AI. It is a firm built so that a fast, cheap, capable and occasionally wrong worker can do much of the routine work safely, with every piece of that work checked, bounded and replayable. We call this a unit-drive firm, after the factories that stopped turning one shaft and gave each machine its own motor and its own wiring. It is what “AI-native” ought to mean, and usually does not.

The idea is old engineering. In lectures given in 1952, John von Neumann showed how to build reliable machines from unreliable parts, treating error “as an essential part of the process under consideration”³³. Networks deliver reliable streams over lossy links; storage arrays survive failing disks. Reliability lives in the structure, not the component.

Models are unreliable parts in a precise sense. Researchers at Thinking Machines sent one prompt to one open-weight model 1,000 times at temperature zero, a setting that should always pick the same next word, and got 80 different completions; the cause was the way inference servers batch requests, not any creativity on the model’s part⁹. They also showed it can be fixed, at some cost in speed. The lesson is not that models are hopeless. It is that determinism is a property you build, and the place to build it is the firm.

Six design rules follow.

1. Write the firm down. Every flow a model touches runs on one typed record: customers, contracts, orders, stock, commitments and the decisions made about them, each with its source and history. The model does not receive a pasted document; it receives the result of a query against the record, with provenance attached. A firm that cannot describe its own state precisely cannot delegate any of it.

2. Make the rules executable. Policies, permissions and approval thresholds are code, versioned and tested like any other code. The rule that a goodwill credit above a set amount needs a manager is not a paragraph in a handbook; it is a check that runs on every request, whether a person or a model made it. A policy that cannot be executed is a suggestion.

3. Act only through verbs. The model never changes the world directly. It proposes calls to a short list of named actions, such as issue_credit_note(invoice, amount, reason), each with typed inputs, hard limits, an idempotency key and a log entry recording who asked, which model and rule versions applied, and what the record said at the time. Where the world allows, every verb has an inverse. Where it does not (a payment sent, an email delivered), the verb waits for a person.

4. Grant authority by evidence. A new employee gets a signing limit, not a blank cheque, and the limit rises with their record. Models get the same treatment, per verb and per amount. A verb starts in shadow mode, where the model proposes and a person decides, and the agreement rate is measured. When errors over a stated number of cases fall below a stated threshold, the limit rises; after an incident, it falls automatically. Trust becomes a number with a sample size.

5. Run verification as production. Checking is not a quality phase at the end; it is a line function with its own staff and budget. Every change of model, prompt or rule runs against evaluation sets drawn from the record before release. A sample of live decisions is audited every day. After an incident, the decision is replayed with the same inputs, model version and rules, and either it reproduces or the log says why not. If a decision cannot be replayed, it was a guess with a timestamp.

6. Keep people on purpose, exceptions and answerability. People decide what the firm is trying to do, handle the cases the rules did not foresee, and answer for outcomes to customers, auditors and regulators. Each exception a person resolves becomes a candidate rule or test, so the exception queue is also the firm’s learning loop. This follows Brynjolfsson’s case for augmenting people rather than imitating them: the design extends what people can oversee instead of removing the person who signs³⁴.

Three economic rules sit on top.

7. Rent the model; own the complements. Treat models as interchangeable suppliers behind a gateway, with pinned versions and the ability to switch. Put capital into what a model vendor cannot sell to a competitor: the customer relationship, the record of which actions led to which outcomes, the audit trail that lets a regulator or a large customer trust the system, and distribution. The record from rule 1 and the logs from rule 3 become exactly this asset.

8. Count output, not hours. Hours saved are an input measure, and they are easily absorbed. The measures that matter are output per person, revenue per unit of capacity, cycle time from request to result, and new things the firm can sell. A firm that reports its AI programme in hours saved has not yet changed anything customers pay for. A March 2026 survey of corporate executives found that the gains firms could already measure ran mainly through innovation and demand rather than cost-cutting³⁵.

9. Draw the boundary again. Ronald Coase explained the size of a firm as the point where organising one more transaction inside costs as much as buying it in the market³⁶. Models change both sides of that comparison, and the boundary is already moving. In McKinsey’s 2026 survey, 32% of respondents said their organisations had decided against buying at least one software product because coding agents let them build it³; MIT NANDA’s savings came from buying less outsourcing. A unit-drive firm makes these choices flow by flow and expects to be smaller per unit of output than its predecessors. Reports from young AI companies point the same way: Cursor reported more than $1 billion of annualised revenue with a little over 300 staff in November 2025³⁷, and Lovable $400 million of annual recurring revenue with 146 employees in March 2026³⁸. These are run-rate figures, not audited results, but they show how small a team can now be relative to its business.

A policy that cannot be executed is a suggestion.

If a decision cannot be replayed, it was a guess with a timestamp.

[PLATE 06]

Roles

A unit-drive firm does not need a department of AI. It needs a few roles, arranged around flows rather than functions; in a small company one person may hold three.

• Flow owner. Owns one end-to-end flow (order to cash, claim to settlement) and its output measures. An owner of the result, not a manager of steps.
• Rules engineer. Turns policy into tested code and keeps it current. Often someone from finance, legal or operations who learned to write tests.
• Verification lead. Owns the evaluation sets, the daily audit sample, the replay tooling and the authority ledger that records each verb’s limit and the evidence for it.
• Exception desk. Handles what the rules route out, and turns each resolution into a proposed rule or test.
• Model operations. Chooses, pins and routes models, watches cost per completed unit, and runs the switch when a cheaper or better model arrives.

Reference architecture

From the bottom up, as drawn in Figure c:

1. Record. An append-only, typed store of entities and events with provenance on every field. All model context is a query here.
2. Rules. A policy engine holding permissions, limits and approval thresholds as versioned, tested code.
3. Verbs. A narrow action interface. Each verb has a schema, limits, an idempotency key, an inverse or a human gate, and writes a full log entry back to the record.
4. Models. Interchangeable, version-pinned models behind a gateway. They read the record and propose verbs; they can call nothing else.
5. Verification. Evaluation sets, daily sampled audits, replay of any logged decision, and the authority ledger.
6. People. Intent and priorities above; an exception queue beside; sign-off on irreversible verbs.

A request runs like this: the model reads the record and proposes verbs, the rules check each one, permitted verbs execute and are logged, a sample goes to verification, anything refused or unusual goes to the exception desk, and each resolution feeds back into the rules and the evaluation sets.

Measures

A path for an existing firm

Incumbents cannot rebuild everything at once and should not try. Go one flow at a time. Choose a flow with high volume and a measurable outcome. Write it down: the record it needs, the rules it follows, the verbs it uses. Run the model in shadow mode until the agreement rate is known. Grant the first signing limits. Then redesign the flow itself, removing steps that existed only to move information between people. Only then measure output. Most programmes do these steps in reverse order, which is why most of them measure hours.

An essay that recommends verification should apply some to itself.

”This is the J-curve. Wait.” Partly true. Brynjolfsson, Rock and Syverson show that when a general purpose technology needs large unmeasured investment in process, skills and organisation, measured productivity first dips and then rises as those intangible assets pay³⁹. The revised Danish study reads the same way: work is being reorganised before anything shows up in earnings³². But the J-curve describes what happens once the complementary investment is made. It does not make it. Earlier work on computers found that firm-level benefits grew for years after the investment, and that firms spent several times more on reorganisation and training than on the technology⁴⁰. Waiting is not an investment.

”The macro gains will be small anyway.” Daron Acemoglu’s task-based estimate puts the ten-year gain in total factor productivity from AI at no more than 0.66%, and lower once harder tasks are accounted for⁴¹. Others estimate more; the March 2026 executive survey implies gains of around 0.8% in high-skill services for 2025³⁵. Our argument does not depend on who is right. A modest average is compatible with a wide spread between firms, and the spread is what matters to any one firm. If the pie is small, the slice depends on design.

”AI already works in support and software.” It does, and that evidence is the firmest part of the case. But look at why it works there: a bounded task, a record of every interaction, a metric the business already counts, and a short loop between output and outcome. Those are the design rules of section 6, arrived at by accident. Even in software the picture varies: METR’s 2025 trial found experienced developers slower with AI tools²⁸, and its February 2026 follow-up pointed to a speed-up while warning that the new data were an unreliable signal⁴².

”'AI-native' is the next empty label.” It may well be. Young companies publish run-rate revenue rather than audited results, the ones we hear about are survivors, and revenue per employee at a reseller of model capacity says little about margins. We would rather offer a test than a label. A firm is built around models if it can show, for each flow, its output per person, the authority each verb has earned and the evidence for it, and the share of last month’s decisions it can replay. A firm that cannot replay last Tuesday is a demonstration with a payroll.

”Incumbents own the complements.” Often they do: the customers, the data, the licences and the trust. Teece’s framework predicts that when imitation is easy and complements are scarce, the owner of the complements wins²⁴. MIT NANDA’s buyers said they would wait for existing suppliers to add AI rather than switch¹⁴, and Gartner notes that many enterprises are taking the simpler AI features their incumbent software vendors build in¹⁰. Against that, rebuilding a firm around models is what Henderson and Clark called an architectural innovation: the components stay familiar but the way they connect changes, which “destroys the usefulness of a firm’s architectural knowledge”⁴³. The likely outcome is not that entrants win everywhere. Incumbents who rebuild one flow at a time keep their complements; those who bolt on lose them slowly.

”Regulation will slow this down.” In the EU it will shape it, which is different. The AI Act requires high-risk systems to record events automatically over their lifetime (Article 12), to let people oversee them and override or reverse their output (Article 14), and requires deployers to keep the logs (Article 26)⁴⁴. Under the amending regulation adopted in July 2026, these obligations apply from December 2027 or August 2028, depending on the category of system. Most of what the Act asks for is what section 6 recommends for commercial reasons: logs, limits, a person who can stop the system, and the ability to explain what happened. A firm built this way treats compliance as a report it can generate, not a project it has to start.

Two limits of our own argument should be stated plainly. The design rules make models safer to use; they do not make an unwanted product wanted, and a well-logged process can produce unwanted output very reliably. And the design is not for every flow: work that is rare, ambiguous or built on relationships may be better left to people with good tools.

[PLATE 07]

JUMPHASH builds software and hardware that behave the same way every time. That has always been a statement about engineering discipline rather than about a particular technology, and it turns out to be the statement this moment needs. Models are probabilistic. Firms, ledgers, contracts and regulators are not. Something has to stand between them.

That is how we intend to build, for our own products and for the systems we build with others: the record first, rules as code, verbs with limits and receipts, authority earned by evidence, and replay as the default answer to “what happened?” We do not claim results from this yet. We will publish measurements when we have them, with the same inputs and limits we publish for everything else.

The history of electricity counsels patience, but not passivity. The factories that benefited were not the ones that waited longest for better motors. They were the ones that rewired.

Rent the motor. Own the wiring. Keep it boring.