Your AI works. Now build the work around it.

Most enterprise AI today is used the way earlier technologies were at first: as a personal tool or bolted onto an existing process. Getting past that stage means rebuilding the work in a set order: make it legible, make the data credible, move decision rights, then automate.

Ask a leadership team how its AI program is going and you will usually hear counts: licenses issued, weekly active users, pilots live, use cases in the pipeline. Most of those counts are rising. A recent U.S. Census Bureau study shows how shallow that use still is. Among firms using AI in business functions, 57% use it in three or fewer, and among those reporting task effects, 66% use it only to augment tasks people already perform.

A 2025 study from Denmark, revised in 2026, tests what that kind of use produces. Anders Humlum and Emilie Vestergaard linked surveys of 25,000 workers in 11 AI-exposed occupations to their pay and hours records. Adoption was broad, and many employers paid for enterprise tools or training. Yet two years after ChatGPT's launch, the study found no measurable effect on earnings or recorded hours. The same authors find that employers absorb AI by reorganizing tasks, which changes how work is done before it changes pay.

The counts leaders report measure adoption: whether people use the technology. What delivers the return is absorption: work rebuilt around the technology. Budget rules of thumb say how much effort to put into people and processes, and redesign frameworks say where the work should end up (see Details). Neither says in what order the work should change. That order decides whether the rebuild succeeds. The history of earlier technologies shows the order, which I call the Absorption Sequence.

Every general-purpose technology arrives in three layers

Computing and AI are general-purpose technologies, meaning they spread across many industries rather than one. Both entered business in three forms, and steam power, an earlier general-purpose technology, went through the last two.

The first is the personal tool. Spreadsheets helped justify corporate purchases of personal computers through the 1980s. Today it is chat, drafting and summarizing. Both spread quickly and proved useful, but neither changed how work moved from one step to the next.

The second is the technology bolted onto the existing process. When steam engines first reached British textile mills, many were used to pump water back over the mill's waterwheel, so the mill kept its riverside site and its old layout. Mainframes ran payroll and ledgers inside processes designed for clerks and paper. Today, AI pilots and agents are placed inside workflows nobody redesigned.

The third is work rebuilt around the technology. Once mills no longer needed a river, they could be sited near coal and workers and built around the engine. Ford rebuilt its accounts payable around a shared purchasing database, and Walmart rebuilt replenishment around data shared with suppliers such as Procter & Gamble. For AI, this layer means processes redesigned around what the technology makes possible.

Exhibit 1. Three layers of a general-purpose technology implementation

Productivity lagged the spread of computing, and firm-level evidence points to where the returns came from. U.S. productivity growth fell from 2.8% a year to 1.2% after 1973 and stayed low through the 1980s despite heavy spending on information technology. Studying large U.S. firms, Timothy Bresnahan, Erik Brynjolfsson and Lorin Hitt found that computers were associated with the largest productivity gains in firms that also reorganized the work around them. Those firms used more teams and handed more decisions to line workers.

Why the second layer is a trap

The second layer's limits are harder to see than the first's, because its savings are real. A predictive maintenance model catches failures that would have been missed. An agent drafts responses a team used to write by hand. The savings are measurable and modest, and they give leaders a reason to declare success and stop. The result is a portfolio of successful deployments, none of which changes the business.

Three signs show a deployment is stuck in the second layer. Its success is reported in model terms, such as accuracy or usage, rather than business terms. Its exceptions wait in a queue for someone with the authority to act on them. And the surrounding process would run the same way if the model were switched off. That last test, the switch-off test, is the most useful because you can answer it in a single meeting.

Take a claims team that deploys a model to flag suspicious claims. Accuracy improves, and claims handlers close files faster. The team reports success in model terms, yet nothing else changes: exceptions still queue for a manager's sign-off, and the process would run the same without the model. That team is still in the second layer. Getting it out means changing who decides.

The order that gets a process to the third layer

The move from the second layer to the third follows the Absorption Sequence. First the work becomes legible: described in data that exists outside people's heads. Then the data becomes credible: trusted enough that people accept its verdict, including when the verdict is unwelcome. Then authority moves: decision rights go to the people and systems closest to that information. Last comes automation, which pays off most when it acts on information people already rely on and decisions they have already agreed to hand over.

The earliest clear example comes from American gunmaking. In 1798 Eli Whitney won a federal contract for 10,000 muskets, due in two years. He later claimed his parts would fit any gun of the pattern. He delivered the last of the muskets in 1809, years behind schedule, and their parts were not interchangeable. The federal armories made the idea work over the next two decades, and their experience maps onto the Absorption Sequence.

Legibility came first. To machine a part to a standard, the armories had to get the standard out of the craftsman's head and into a physical object. By the 1820s, John Hall's rifle works at Harpers Ferry had done it with gauges, and in 1827 government inspectors confirmed that his rifle parts interchanged.

Credibility took far longer. Craftsmen resisted gauged work for years, because a gauge overruled a master armorer's judgment. A gauge could win acceptance only if it protected an honest worker as reliably as it caught a bad part.

With the gauges trusted, authority moved from the master craftsman to the gauge and the machine. The machines completed the sequence by making parts to the gauge.

Ford's 1980s redesign of accounts payable includes all four stages. Purchase orders went into a shared database, and the receiving dock's check against it became the record everyone relied on. A clerk's acceptance at the dock was then enough to trigger payment, and the computer did the matching and prepared the check. Where the new process was in place, accounts payable headcount fell by 75%, against the 20% Ford had first targeted.

Exhibit 2. The Absorption Sequence in practice

Credibility, the second stage, is where AI helps least. Models are more accurate and faster to test, but accuracy was never the slow part. The slow part is getting people to accept the system's verdict. That's where the armories took most of the time, and it is where many AI programs stall today.

Language models make credibility harder to earn. A gauge gives the same verdict every time it checks the same part. A language model can give different answers to the same question, and it can be wrong with full confidence. So a process built on one needs its own version of the gauge: a set of real cases to test the model against, a passing standard that the people relying on it agree on before launch and a rule for what happens when an answer falls short. Once those three are in place, people can act on the model's answers without checking them again.

One objection is that language models can read documents, emails, and call transcripts, so the first stage is no longer needed. But a model can only pull out what was written down and those sources rarely record what was decided, by whom and why. That record is what the first stage produces. A second objection is that AI-native firms have no old process to rebuild. They still have to go through the stages; they just start from a blank page. Skipping a stage doesn't save time either. A firm that hands decisions to a system before its people trust the data tends to slide back, or it quietly adds people to check the system's work.

What to do with the portfolio

Start by sorting every AI investment into the three layers. Most will land in the first two. Run the switch-off test on each one in the second. A deployment that fails it can keep running, because its savings are real. But scaling it adds savings without changing the process, so stop funding its expansion and redirect that money to the processes you will rebuild.

Then pick two or three processes to rebuild end to end. Choose processes where the work is already largely legible, so the program can start at credibility. Give the credibility and authority stages their own budget line, set before the pilots start and sized from your own past programs, not an industry benchmark. Name a patient buyer: an executive who owns the result, controls the budget for those slow middle stages and will still be in the role when the payoff arrives, which may be years away. Until then, judge progress by what changes in the process, not by how many people use the tools.

Measure absorption, not adoption

Most AI dashboards report usage and model performance, and those numbers rise smoothly through the long stretch before returns appear. To see whether a process is moving from the second layer to the third, track these instead.

Exhibit 3. What to measure at each stage of the Absorption Sequence

Because each stage depends on the one before it, score each process by the furthest stage it has completed in order. A process with excellent automation running on data nobody trusts has completed only the first stage, whatever the dashboard says.

Getting the work into data is only the start. The returns come once people trust that data, decisions move to the people who hold it and automation acts on both.

Ask one question of every AI investment at the next portfolio review: what changed in the surrounding process when it went live? If the answer is nothing, the pilot may work, but the business has not changed.