Productivity is not value

Why record AI spend and reported productivity gains still do not add up to visible business value, and what the missing measurement is really about.

July 14, 2026 8 min read

Two numbers are sitting on desks across the enterprise this quarter and they do not agree. The first is the AI bill, which is the largest it has ever been and still climbing. The second is the answer to a simple board question: what did we get for it. In Harness's 2026 State of Engineering Excellence survey of around 700 engineering practitioners and managers, 89 percent said their current metrics accurately reflect AI's impact, and in almost the same breath 94 percent admitted the factors that actually matter, including tech debt, validation time, and developer burnout, are missing from those metrics. Proving ROI to leadership was named one of the single biggest challenges. They believe they are measuring AI, and they know they are not measuring the part that counts.

The consequences are already visible and they are not subtle. Uber burned through its entire 2026 AI budget in four months and started capping per-employee spend. Starbucks retired an AI inventory system after nine months because it miscounted the shelves it was built to count. These are not stories about weak models. They are stories about organizations that could not connect what the AI was doing to what the business needed, and found out late.

The value attribution gap

Here is the gap, named so we can stop talking around it. You can measure AI output. You cannot attribute that output to business outcomes. Lines of code, pull requests, tickets closed, tokens spent: all trivially countable, all going up. Whether any of it advanced the thing your company is actually trying to do: uninstrumented. So the dashboard fills with green while the honest answer to "is this working" stays "we cannot tell," and the two coexist without anyone lying. That is the value attribution gap, and no amount of additional output closes it, because output was never the missing measurement.

It is worth being precise about why the two halves diverge, because the instinct is to treat it as a reporting problem that a better dashboard will fix. It is not. Output becomes value only when it is the right output, meaning the work advanced a decision the business actually made: a strategy, a constraint, a commitment to a customer. An agent generates the most probable work given what it can see, and what it can see is the codebase and the ticket, not the reasons behind them. So it produces volume against inferred intent. Some of that volume happens to serve a real decision and becomes value. Much of it serves a plausible guess and becomes activity. From the outside the two are indistinguishable, which is exactly why you cannot measure the difference: nothing recorded which decision each unit of work was meant to serve.

Both failures this quarter were the same failure

Look at Starbucks and the measurement gap together and they stop being two problems. The inventory agent was measured on a proxy, whether it produced counts, and it produced counts enthusiastically. What went unmeasured was whether those counts served the operational decision they existed for, keeping the right things in stock. It optimized the visible output and quietly failed the outcome, and because the outcome was not instrumented, the failure showed up in barista complaints months later rather than in a metric on day one.

That is the same shape as the boardroom question. When work is not tied to the decision it was supposed to serve, you can watch the output climb and still have no read on whether the business moved. The enterprise version is just slower and more expensive, because it is distributed across every team using an agent, and the lag between "output looks great" and "this did not land" is a quarter instead of a shift.

Why a dashboard cannot close it

The reflex, once leadership starts asking, is to buy measurement: an AI ROI tool, a productivity analytics layer, another panel. Vendors are already shipping exactly this. It will not close the gap, and the reason is structural. Attribution is a reconciliation. To connect a unit of work to a business outcome, both sides have to be recorded and something has to link them: the work, and the decision it served. The output side exists in abundance; every tool emits it. The other side, the decision the work was meant to serve, mostly does not exist as a record at all. It lives in a head, a thread, a call that happened once. You cannot reconcile against a record that was never kept. A dashboard built on one side of that reconciliation produces confident charts of the half you have, which is precisely the 89-versus-94 result: leaders measuring output accurately and outcomes not at all.

This is the deeper reason the cost of context-less development shows up as unmeasurable rather than merely large. The waste and the blindness have one root. The agent that cannot see the decision builds the wrong thing, and the organization that never recorded the decision cannot see that it did. You are paying for both halves of the same missing object.

What the missing measurement actually is

If attribution is a reconciliation and one side of it was never recorded, the fix is not a better view of the side you have. It is to make the missing side real: capture the decisions, constraints, and commitments the work is supposed to serve, as durable records with enough structure that a piece of work can point at the one it advanced. Do that and two things become true at once that executives usually treat as separate asks. The agent conditions on the actual decision before it generates, so more of its output is the right output rather than a plausible guess. And every unit of work carries the decision it served, so for the first time the output and the outcome sit in the same ledger and the reconciliation resolves. Alignment and measurement turn out to be the same investment seen from two ends: you cannot measure whether work served the strategy until the strategy is written down in the form the work can reference, and once it is, the agent can reference it too.

We built Brief to be that record, the governed layer of decisions an agent reads before it builds and a business reads to see what the building was for. But the conclusion does not depend on us. The enterprises that will answer the board's question next quarter are not the ones with the most AI output or the fanciest ROI dashboard. They are the ones that wrote down what they decided, so that for once the work and the outcome are talking about the same thing.

What this does and does not claim

It does not claim measurement fixes strategy. If the underlying decision was wrong, attributing work to it faithfully will only tell you sooner, which is the point, not a flaw. It does not claim every outcome is cleanly attributable; some value is diffuse and some is luck, and honest measurement includes wide error bars. And it does not claim AI is not producing real gains: the reported productivity is likely real at the task level. The narrow, load-bearing claim is that task-level productivity is not the same quantity as business value, that the enterprise is currently measuring the first and calling it the second, and that the gap between them is a decision you never recorded, not a chart you never bought.

Frequently asked questions

Why can't we prove the ROI of our AI investment? Because you are measuring output, which every tool emits, and not outcomes, which almost nothing records. Attribution requires connecting a unit of work to the business decision it served, and that decision usually exists only in someone's head or a one-time conversation. With one side of that link missing, no dashboard can complete it, so spend keeps climbing while its effect on the business stays unmeasured.

Is AI actually making our teams more productive? Most likely yes, at the level of individual tasks, which is what surveys capture when engineers report feeling faster. The catch is that task productivity and business value are different quantities. More code produced faster is only valuable if it was the code the strategy needed, and nothing in a task-productivity metric tells you whether it was.

Why does more AI output not translate into business results? Because output becomes value only when it advances a real decision, and an agent working from the codebase alone cannot see the decisions that live outside it. It produces plausible work against inferred intent, some of which serves the strategy and much of which is merely activity. The two look identical in the metrics, so volume rises without a matching rise in outcomes.

What is the value attribution gap? It is the inability to connect AI output to business outcomes, even while output is easy to measure. It arises because attribution is a reconciliation between work and the decisions that work was meant to serve, and the decision side is rarely recorded. The result is organizations that can chart their AI activity in detail and still cannot say whether it moved the business.

How do you connect AI development to business outcomes? By making the decisions themselves first-class records: capturing the strategy, constraints, and customer commitments the work is supposed to serve, in a durable and structured form. Once those records exist, agents can condition on them before generating and each piece of work can reference the one it advanced, which closes the loop that measurement has been missing.