This paper is written for the people who have to show that AI spending produced a return. The CFO who signs the budget, the Chief AI Officer who promised the lift, and the operating leader who runs the team where the tool landed. It answers one question. Your people are measurably faster with AI, so where did that speed go, and why has almost none of it reached the income statement.
The Realization Gap, the Reallocation Gap, Slack Reabsorption, Captured and Diffused Gains, the Realization Rate, and the Realization Architecture are original constructs of this work, published under the license on the back cover. They build on measured evidence that individual productivity gains from AI are large and real [1][3], and on the economics of why such gains appear late at the aggregate level [2]. Every quantitative claim is cited to a named source or labeled directional, without exception.
Every finance leader I meet has the same quiet problem with AI. The pilots work. The teams are faster. The demonstrations are real. And when the year closes, the margin looks exactly as it would have without any of it. The gain was true at the desk and invisible at the bottom line.
This is not a story about AI failing to deliver. It delivers. A support agent handles more cases. A developer ships a feature sooner. A writer finishes a draft in half the time. The gain is measured, repeated, and large. The failure is downstream of the gain. The saved time refills with the same work, the saved capacity scatters across too many people to remove, and the better output never moves a line that finance can book. The value was created and then it quietly diffused.
This paper is my answer to the question I now ask in every value review. Not whether the tool works, it usually does, but whether anything in the organization is built to catch what the tool frees. I give that catch a name, a measure, and an architecture. The constructs are original work. The numbers are cited or labeled directional, without exception. I hold this to the standard of the rooms it will be read in.
Your AI investment is probably working and probably not showing up in the numbers. The gain is real at the task and lost on the way to the income statement. The distance between the productivity you can measure and the financial result you can book is the subject of this paper. Five findings follow.
The gains are real and large. Support agents using generative AI resolved work 14 percent faster on average [1], and developers completed a coding task 55.8 percent faster [3]. The question is never whether AI creates value. It is whether the value is captured.
The gain diffuses before it is booked. Freed time refills with the same work, saved capacity spreads too thin across a team to remove, and better output never moves a revenue or cost line. The paper names six such leaks. The largest is Slack Reabsorption, where saved minutes simply become a calmer pace.
Capture is an architecture, not a hope. The Realization Architecture meters the gain against a baseline, consolidates fractional savings into whole units, decides to redeploy or reduce, and books the result to a line finance recognizes. Each step is owned.
The aggregate gain arrives late by design. The Productivity J-Curve shows that general-purpose technologies look unproductive early, while the complementary work is built, then pay out later [2]. A low result this year is a signal to build the complement, not to stop.
Measure the Realization Rate, then raise it. The Realization Rate is the share of a measured gain that reaches the P&L. Report it per use case, chase the gains that convert, and let the ones that only feel good stay felt. One quarter is enough to start.
Start with what is not in doubt. Support agents using a generative AI assistant resolved work 14 percent faster on average, and the least experienced staff gained the most [1]. Developers with an AI assistant finished a coding task 55.8 percent faster [3]. In a controlled experiment, professional writing time fell sharply as well [4]. The gain is measured, repeated, and large.
Now ask the finance question. Did cost per resolution fall, did the release date move in, did the team get smaller or sell more. In most organizations the honest answer is no. The gain happened at the desk and stopped there. That distance, between a gain you can measure and a result you can book, is the Realization Gap.
A productivity gain is not a return until a line the controller recognizes actually moves.
A gain this size should show up in the numbers. Exhibit 2 shows why it usually does not.
A measured gain does not reach the P&L by itself. On the way it passes through six leaks, each a specific place where real value turns into something finance cannot book. Naming them is the first act of capture, because a leak with a name has an owner.
Saved minutes refill with the same work at a calmer pace. The gain becomes comfort, not capacity, and leaves no trace in the numbers.
Time saved is never pointed at higher-value work. Freed hours exist but convert to no new output, so the saving evaporates.
A tool saves a fifth of ten people. Two roles of value exists, spread too thin across the team to remove or reassign.
Faster throughput is absorbed by demand that was always waiting. More gets done, but unit cost and headcount hold flat.
The gain is spent doing the same volume better. Real, and unbooked, unless a revenue, retention, or risk line actually moves.
The gain is real but never metered at the financial line, so finance cannot book what no one thought to count.
The six leaks share one cause. Nothing in the organization is responsible for catching what the tool frees. The Realization Architecture fixes that with four owned steps that move a gain from the desk to the income statement, and one metric that tells you how much survived the trip.
| Step | What it does | The leak it closes |
|---|---|---|
| Meter | Measures the gain at the task against a pre-AI baseline, so it exists as a number. | Measurement Blindness. A gain no one counts cannot be booked. |
| Consolidate | Pools fractional savings across people into whole, removable or redeployable units. | Fractional Headcount. A fifth of ten becomes two whole roles. |
| Redeploy or Reduce | Points freed capacity at a named higher-value output, or takes it out as cost. One of the two. | Slack Reabsorption and the Reallocation Gap. Capacity gets a destination. |
| Book | Ties the change to a revenue, cost, or risk line a controller recognizes, with an owner. | Queue Backfill and Quality Substitution. The line moves, or the choice is made on purpose. |
A board can see real desk-level gains and no movement in the yearly numbers, and that is not failure. General-purpose technologies follow a known shape, and understanding it changes what a weak first year means.
Early on, measured productivity dips, because the organization is paying for complementary work, the process redesign and new ways of operating, that is not yet counted as output. Later, once that complement is built, the gain is harvested and measured productivity rises above trend [2]. This shape is the Productivity J-Curve.
It means a weak result in year one is often a signal to finish building the complement, not to stop. And capture is the complement. The Realization Architecture of Section 03 is the operating work that turns the dip into the harvest.
Some gains convert to booked value far more readily than others. A gain that lands as direct cost reduction books at a high rate. A gain that only improves quality books at a low rate, unless a revenue or risk line is wired to it. Knowing the pattern tells you where to spend the capture effort.
Capture does not need a transformation program. It needs three phases on one use case, each with a deliverable and a gate. Finish the phase or do not advance. By the end of a quarter you can report a Realization Rate to finance for your first AI use case.
Not every gain should be chased to the income statement. Four variables tell you which ones convert and which are better left as a quieter, faster desk.
Is the saving in a few roles, where it can be removed, or spread thin across many, where it cannot?
Can freed capacity point at a revenue or cost line, or does it only make existing output better?
Is there latent demand that will backfill the throughput, holding unit cost and headcount flat?
Can the gain be metered against a baseline at the financial line, or can it only be felt?
The Realization Rate is the share of a measured gain that reaches the P&L. What completes all four steps is a Captured Gain. What stops short is a Diffused Gain, and it books nothing.
| Probe | Question | Yes | No |
|---|---|---|---|
| 1 METER | Do you have a pre-AI baseline for your highest-value AI use case? | ||
| 2 COUNT | Is the gain recorded as a number, not just reported as faster? | ||
| 3 POOL | Have you pooled fractional savings across people into whole units? | ||
| 4 DEST | Does every unit of freed capacity have a named destination? | ||
| 5 CHOICE | Have you chosen redeploy or reduce for each gain, on purpose? | ||
| 6 BOOK | Is the gain tied to a revenue, cost, or risk line finance recognizes? | ||
| 7 OWNER | Does each booked gain have an owner who confirms it in the numbers? | ||
| 8 RATE | Do you report a Realization Rate per use case? | ||
| 9 TRIAGE | Do you skip gains that cannot convert, rather than chase every one? | ||
| 10 JCURVE | Does the board read a weak year one as a signal to build, not to stop? |
Your score places you on the posture model of Section 06. The stage you land on is the one to finish before you climb.
This paper is a conceptual contribution. The Realization Gap, the Reallocation Gap, Slack Reabsorption, Captured and Diffused Gains, the Realization Rate, and the Realization Architecture are original constructs of this work. They rest on measured evidence that individual productivity gains from AI are real and large, in customer support [1], software development [3], and professional writing [4], and on the economics of why such gains appear late at the aggregate level, the Productivity J-Curve [2].
Every exhibit is marked. Solid fill is cited data. Hatched fill is directional illustration, shown to make a mechanism legible, not to assert a measured value. Red marks a gain that leaks away. The hard figures in the paper, the measured task-level gains, are cited to named studies [1][3]. The six-leak shares, the Realization Rate by gain type, and the capture curve are directional by construction.
Arjun Jaggi, The Realization Gap, Enterprise AI Research, White Paper No. 08, arjunjaggi.com, November 2026.
Arjun Jaggi writes and advises at the intersection of enterprise AI strategy, the economics of adoption, and the operating work that turns capability into result. His frameworks turn the hardest questions of running AI at scale into terms a board can decide on and a team can execute. Engage at arjunjaggi.com.