The work around the machine.

Picture a purchase order pinned to the corkboard in a support center's operations office. It lists 120 seats of an AI assistant for the service desk, the connector to the ticketing system, hosting, and two days of onboarding from the vendor. Across the bottom, in rubber-stamp ink, is one word: activated. Taped beside it is a sheet torn from a notepad, in someone's handwriting. Repair the knowledge base. Train the supervisors. Decide when an agent escalates. Change how replies are reviewed. Both pieces of paper are invented, and both are accurate in their own way. The software is switched on. The list explains why switching it on did not make the organization ready to use it.

What the order buys is purchased technical capability: a tool that can do something, delivered and running. What the list describes is complementary organizational capability: the routines, knowledge and coordination that let the tool's ability turn into better work. My argument is that leaders who fund a new technology should treat those complements as an investment program in their own right, with named owners, evidence that each one exists, and a rule for stopping. A slow payoff can be a sign that an organization is learning and redesigning. Delay alone is not evidence that a payoff is coming.

This is not a claim that every tool needs a transformation program. Some products work out of the box, and the right response to a good, simple tool is to use it. Nor is it a universal checklist. Which complements matter depends on the workflow the tool enters, and a list copied from someone else's program will be wrong in the ways that count. The claim is narrower. The invoice names the software. It rarely names the work around it, and that work is often where the return is decided.

What a complement is

A complement is anything whose value rises when the tool arrives and without which the tool is worth less. For an assistant that drafts replies to customers, the obvious ones are concrete. The assistant can only suggest what its source material supports, so a knowledge base full of stale articles produces fluent, stale suggestions. Someone has to decide which documents the tool may draw on and who may see what. Someone has to check its suggestions against a standard, which means the standard has to be written down. Supervisors have to coach with the tool in the room rather than around it. And the boundary of each person's task moves: when should an agent accept a suggestion, rewrite it, or hand the conversation to someone with more authority?

None of these appears on the purchase order, because none of them can be bought from the vendor. Some can be bought from somebody (a contractor to clean the knowledge base, a trainer for the supervisors), but each has to be done to this organization's documents, people and rules. That is what makes them organizational rather than technical. It is also what makes them easy to underfund. They look like overhead, they are spread across several budgets, and when they are done well nobody notices them.

Why the payoff can arrive late

Erik Brynjolfsson, Daniel Rock and Chad Syverson gave this pattern a name and a model. In "The Productivity J-Curve," first circulated as a National Bureau of Economic Research working paper in October 2018 and published in the American Economic Journal: Macroeconomics in January 2021, they argue that general purpose technologies, a class in which they place AI, both enable and require large complementary investments: new processes, products, business models and human capital.¹ Much of that investment is intangible and poorly measured in the national accounts, even when it builds something valuable for the firm.

The consequence is a measurement pattern. Early on, while firms spend effort on complements that the statistics do not count as output, measured productivity growth looks too low. Later, when those investments pay off and their fruits are counted without their cost, it looks too high. Plot the measured figure against the true one over time and the gap traces a J. The authors applied the method to earlier technologies and estimated that, once intangibles related to computer hardware and software are counted, the level of US total factor productivity at the end of 2017 was 15.9 percent higher than official measures showed.¹

Two cautions keep the model in its place. It is an account of how investment and measurement behave across an economy, not a promise about a particular company. And it explains a possible timing pattern, not a guaranteed one. A J-curve describes what happens when the complementary investments work. It says nothing about the projects whose complements were never built, or were built around a tool that could not do the job. Those projects show a dip too. They simply never climb out of it.

A tool inside a working service

The strongest field evidence on a generative AI tool at work comes from a study by Brynjolfsson, Danielle Li and Lindsey Raymond, published online by the Quarterly Journal of Economics on February 4, 2025, and in its May 2025 issue. They studied 5,172 customer-support agents as an AI assistant was introduced to them in stages. With access to the assistant, agents resolved 15 percent more issues per hour on average. Less experienced and lower-skilled agents improved in both speed and quality, while the most experienced and highest-skilled saw small gains in speed and small declines in quality. The gains were largest for moderately rare problems, where agents had less experience but the system still had enough training data. Customers were more polite and less likely to ask for a manager.²

What matters here is the setting, because so much of the setting was complement. The assistant did not replace the service; it was placed inside one. As the working paper version describes it, the company was a Fortune 500 firm selling business process software, and the tool watched live chats and offered the agent two things: suggested replies, and links to the company's internal technical documentation.³ Its suggestions came from a model trained on a large set of the company's own past conversations, labeled by outcome: whether the issue was resolved, how long it took, and whether the agent handling it was considered a top performer. The output was shown only to the agent, who was free to ignore it.³

Read that description again with the purchase order in mind. The tool arrived in a service that already had a queue, a measure of resolution, a body of documentation, years of recorded conversations and a considered judgment about who did the work well. Some of what made the assistant useful was organizational knowledge the firm had accumulated before anyone licensed a model. One further result hints at learning in the people rather than the software. Using periods when the tool was unavailable, the working paper found that agents still performed better than their own baseline from before the assistant arrived.³ That is suggestive, not conclusive, but it points at the kind of complement no invoice can list: people who have absorbed better ways of working.

Be precise about what the study shows. It is substantial workplace evidence from one firm and one workflow, built on a staggered rollout rather than a randomized trial of the whole program. It measures a rate of work inside a service, not the firm's profit, and the authors write that the paper is not designed to shed light on the aggregate employment or wage effects of these tools.³ Nor does it estimate the return on any particular complement; it was not built to separate the value of the documentation from the value of the model. What follows, the register and the reviews, is my proposal for how to invest, not something the study tested.

Measure the mechanism, not just the curve

The J-curve has a hazard built into it. If a dip is expected, every dip can be explained. A program that is failing and a program that is learning look the same on a quarterly chart for a while, and the people who championed the purchase have every reason to say they are still on the downward part of the curve. The model cannot settle that argument, because a curve is only visible in hindsight. What can settle it is evidence about the mechanism: is the work that the payoff depends on actually getting done, and is it having the effect it was supposed to have?

Figure 2 sketches the problem. Two programs start with the same purchase and the same early dip in a service result. In one, the complements get built, and the result recovers and then rises. In the other, they are not built, or the core tool cannot do the job, and the result stalls. At the first review the two lines are close enough that the chart alone cannot tell you which program you are funding. Drag across the sheet, or tab to it and use the arrow keys.

The difference at the first review is not on the curve. It is on the list. One program can show that the knowledge base has been cleaned, that supervisors coach with the tool, that the escalation rule is written and followed, and that quality reviews now sample assisted replies. The other has the same list with few items done, or has done them and found that the suggestions are still wrong. Either finding is useful. The first justifies patience. The second justifies a redesign, a narrower scope, or a stop.

A register for the board

So give the complements the treatment the purchase already gets. A complement register records five things for each one. First, which capability has to change, stated for this workflow: not "change management" but "the 200 most used articles reviewed and the stale ones retired." Second, the evidence that will show the capability exists, something a skeptic could check. Third, who owns it, by name, with the authority and the budget to do it. Fourth, which measured service result it should move, and in which direction. Fifth, a review date, with an alternative explanation written down in advance: if the result does not move, what else might explain it?

The last item is the one most plans leave out, and it is the one that turns patience into a test. Writing down beforehand what would count as the mechanism failing is what separates a program that is learning from one that is merely waiting. It also protects the people doing the work. A team that can show its complements were built, and that the result still did not come, has learned something true about the tool, and deserves thanks rather than blame.

Two cautions about money. Calling this spending an investment does not make it an asset on a balance sheet, and nothing here suggests putting training costs on one; "intangible investment" is an economist's description of what spending does, not an accounting treatment. And a register is not a license to spend without limit. Each line carries a budget and a date, and the program as a whole needs a point, agreed before it starts, at which its sponsors will stop.

The wider picture

The economy-wide evidence is still thin, which is what the J-curve would predict and also what a disappointment would look like. In a note published on July 17, 2026, Federal Reserve Board economists Paul Soto, Mason Thieu and Jeffrey Allen sorted the public indicators of AI's impact into three groups: capabilities and costs, firm investment and adoption, and productivity and labor.⁴ They wrote that aggregate output and labor data currently show limited signs of broad-based transformation, while much of the evidence points to an economy reorganizing around the technology.⁴ That is the authors' analysis, not an official forecast, and a dispatch here in July read it as a lesson for board packets. Its use for this essay is the order of the groups. Capability arrives first, investment and adoption follow, and measured results come last, if they come. A company's own program has the same three layers, and the middle one, the work of adopting, is the one most fully under its control.

Objections, considered

A good product should work out of the box. Some do, and for those the advice here is to use them and move on. The complement argument applies where the workflow exposes specific dependencies: the tool's suggestions are only as good as the documents behind them, or its output changes who is accountable for a decision. Where the list beside the purchase order is empty, leave it empty.

This hides bad technology behind change management. It can, which is why the alternative path and the stopping rule are not optional. A tool whose core capability is broken is not cured by training people harder or rewriting the knowledge base again. The register is designed to expose that case early. If the complements are in place and the result still does not move, the evidence points at the tool.

The research is already old for AI. The assistant in the study was built on a model from an earlier generation, and the bulk of its rollout took place between November 2020 and February 2021.³ Its value is not the size of its estimate for current products. It is a documented field setting that shows the mechanism: a tool that drew on an organization's recorded knowledge and worked inside its existing routines.

A support-center result does not generalize. Correct, and the study does not claim that it does. The 15 percent belongs to one firm's chat service. What carries over is not the percentage but the method: measure a result that matters in your own workflow, before and after, and do not inherit someone else's number as your forecast.

The organization must build the capability that the invoice cannot deliver.

Before the next renewal

• Which complement is missing? Name it for this workflow, not from a checklist.
• What would show that it exists? Evidence a skeptic could check, not a status color.
• Who owns it? One named person with the budget and the authority to do it.
• Which result should move, and by when? One measured service outcome and a review date.
• What would make you stop? Write down the alternative explanation and the stopping rule before the first review, not after it.

Back to the purchase order on the corkboard. The stamp is still right: the software is activated. The handwritten list is still right too, but it has changed. Beside "repair the knowledge base" someone has written a name, the 200 articles to review, and a date eight weeks out. Beside each of the other items is an owner, a piece of evidence and a review. It is a shorter document than a transformation plan, and a more honest one. It says what the organization has to build, and how it will know whether it did.