Article · Writing
Who Owns the 25%?
AI creates productivity. Realization is what management does with it, and most organizations never assign an owner.
The approval, and the review
In March, the VP of Customer Operations asks the CFO for $50,000, off cycle.
It is not a hype request. Two dozen people in her group spend much of their week summarizing cases, drafting responses and rekeying information between two systems. The AI tool she wants removes most of that. Conservatively, four hours per person per week. At the loaded hourly cost the finance team itself publishes, that is roughly $265,000 of annual capacity against a $50,000 license. Just over five times the cost.
The CFO does his job. He pressure tests the four hours, asks how the number was derived, asks what the vendor is claiming versus what the pilot actually showed, and knocks the estimate down to be safe. It still clears easily. He funds it from an underspend elsewhere and approves it.
Deployment finishes in July. The tool works. Adoption is good. In the Q3 review, the VP presents her P&L.
Labor cost is unchanged. Not slightly down. Unchanged. Headcount is flat, overtime is flat, the contractor line is flat. Volume is up 3 percent, which is roughly what the market did anyway. The department is over its operating plan, and the only new line in it annualizes at $71,000 once usage overages are counted.
Nobody in the room is lying. The CFO asks the team directly and they confirm it: the four hours are real, and most of them are glad the tool exists.
So he asks the only question left. Where did the time go?
Nobody has an answer that survives the meeting.
The business case was not wrong about productivity. It was silent on what came after. At no point in the approval chain was anyone asked which outcome the recovered time was supposed to produce, or who would be accountable for producing it.
That story is invented in its details. The pattern is not.
The invoice arrives and the offset does not
Some version of that review is happening in companies and public agencies all over the world this quarter, and it resolves the same way in nearly all of them, because the two halves of the ledger do not arrive with equal force. The AI invoice arrives every month, itemized, on a date, with a name attached to it. The offset usually does not arrive at all.
AI spending arrives on a real budget line owned by a real executive. The productivity it buys arrives as fragments of time across hundreds of individual days, owned by nobody.
Suppose AI makes a function 25 percent more productive. That is not a benefit. It is capacity, and capacity has to be deliberately converted into lower resource requirements, greater output, or better service and mission performance. What does not get converted stays unbanked.
That conversion is Realization: the deliberate conversion of measured capacity into lower cost, greater output, or improved mission performance.
Productivity is what the technology creates. Realization is what management does with it.
Efficiency is not a benefit until someone banks it.
The four doors
Every measured productivity gain eventually goes through one of four doors.
Door one, cost down. Same output, fewer resources. In an enterprise that is headcount reduction, attrition management, span of control changes, contractor reduction. In government it is vacancy non-backfill, overtime reduction, cuts to contractor and temporary labor, and not replacing retirements. The public version is slower, smaller in any single year, and politically different in kind. It is also the version most public leaders can actually use.
Door two, output up. Same resources, more work. A support organization deflects tier one tickets and the same team resolves more cases.
Door three, service or mission up. The capacity goes into speed, quality, reach or mission performance. A permitting office puts recovered administrative time into clearing resubmittals faster.
Door four, unbanked. The productivity is real and nothing in the organization changes.
Door four is not the failure case. It is the default. Door four is capacity without an owner.
One caution on the arithmetic. A 25 percent productivity gain does not automatically become a 25 percent cost reduction or a 25 percent output increase. Staffing ratios, statutory minimums, labor agreements and demand that does not yet exist all constrain the conversion. Which door you can actually open determines what you get.
Why door four wins by default
Door four requires no decision, no meeting and no owner.
Productivity from AI usually arrives in pieces too small to see. Five minutes on a document. Twenty minutes on a briefing. Four hours across a week, spread over two dozen people. Those gains fall below the resolution of every system an organization uses to manage resources.
The gains are frequently real, and that deserves conceding plainly. Staff are not exaggerating simply because the enterprise cannot find the result. The time was genuinely saved. It had no organizational destination.
A gain that stayed inside an individual's day left with them at 5pm.
It is why strong individual productivity and weak enterprise economics can sit side by side for years without contradicting each other.
You cannot walk through a door without a baseline
Doors one, two and three all require a credible before state: what the process cost, how long it took, how much of it there was. Door four requires nothing at all. That asymmetry, more than any technology decision, is why door four wins so often.
McKinsey's 2026 State of AI research, 1,719 respondents surveyed in May and June, shows the shape of it. Eighty percent of respondents who use AI in their work said it improved their individual productivity. Thirty seven percent said their organization could attribute any EBIT impact to AI. Six percent qualified as high performers, attributing at least 5 percent of EBIT to AI.
The distance between 80 and 37 is not value left on the table. It is two different things being counted: widespread individual productivity, and much narrower enterprise financial realization. What separated the two groups is the useful part. McKinsey's high performers were about twice as likely to have defined processes for measuring AI's impact, meaning baselines and controlled comparison rather than anecdote.
The public sector mirror is Code for America's 2026 Government AI Landscape Assessment. Nearly every state has run AI pilots. Seven were rated established on mechanisms to measure impact. None reached advanced. Experimentation has moved far faster than the ability to say what it produced.
Five measures, captured before deployment rather than after: volume, cycle time, labor hours by role, cost per unit, and quality or rework rate.
A faster bad process is not an efficiency gain, which is why the fifth one matters more than it looks.
Measure the process before you touch it. After you touch it, you are not measuring. You are negotiating.
Nobody outside the building can do this for you
There is a tempting shortcut here, which is to let the platform report the benefit. Most tools now ship with a dashboard that offers to do exactly that.
Look closely at what those dashboards count. Seats. Sessions. Prompts. Queries. Tokens. Documents processed. Suggestion acceptance rates. Sometimes an hours saved figure, derived by multiplying your usage by a benchmark the vendor developed somewhere else. Every one of those is a measure of consumption or engagement. None of them is a measure of what your organization received.
That is not an accusation of bad faith, and it does not need to be. A vendor is structurally incapable of computing your realization. They cannot see your labor cost structure, your vacancy authority, your overtime rules, your backlog, your service targets or your union agreement. They can see how much you used. And their revenue grows when that number grows, which means nothing in the reporting is designed to tell you that you bought more capability than you converted.
An hours saved figure on a vendor dashboard is the input to a business case, not the output of one.
A vendor can tell you how much you used. Only you can tell you what you got. Any measurement that matters has to be instrumented in your systems, against your baseline, by someone whose job is the outcome rather than the renewal.
The gain needs an owner before it exists
Measurement is necessary and it is not sufficient. An organization can measure a gain precisely, publish it in a dashboard, and still receive nothing from it, because measurement answers how much and says nothing about who.
If the team that produced the gain keeps all of it, the gain usually ends up in door four. So the instinct is to harvest it centrally. But if the organization harvests 100 percent of every measured gain, managers learn something quickly and permanently: reporting a productivity improvement produces a budget cut, a headcount reduction, or a higher workload expectation with nothing in return.
The rational response is to stop reporting gains.
A private sector manager understates AI enabled productivity because an accurate number could cost the team two positions it will need next year. A public agency department understates it because the budget office may claw back the capacity. Neither is a moral failure. Both are competent people responding correctly to the system they are in.
Take 100 percent of the gain and you will not hear about the next one.
Tell a manager that every reported efficiency becomes next year's budget reduction, and you have built an extraordinarily effective system for preventing managers from finding efficiencies.
So the question is not only what the gain is. It is who owns converting it, and what that owner is allowed to keep.
What a declared split looks like
There is no correct percentage, and 50/50 is not an answer, it is the avoidance of one. What matters is sequence.
Three requirements. The rule is declared in advance. It is written down. Some defined share of the value stays with the unit that produced it, in a form that unit genuinely values.
Cash is rarely what a team wants. Headcount protection. A backlog they get to clear. Training time. A tool they have asked for twice. Room to absorb work they have been refusing.
An enterprise might return part of the capacity to the original business case and let the team keep the rest for growth. A public agency might realize part through overtime and attrition and put the balance into service improvement.
The percentage matters far less than setting the rule before anyone knows how much there is to fight over.
Door three is not a consolation prize
The difference between the private and public cases is not profit versus mission. It is financial realization versus mission realization. Both are real economics. They post to different ledgers.
Take a transit agency where AI reduces the time service planners spend assembling ridership, reliability and performance data. Nothing about that requires eliminating planners. The same capacity can go into analyzing route performance more often, catching deteriorating service earlier, adjusting schedules faster, or absorbing new planning work without adding staff. Those are real outcomes, expressed in service performance rather than margin.
Or a benefits office where AI reduces the administrative work around determinations. The agency can reduce overtime, process more applications, shorten processing times, or give caseworkers more time on complex cases. Every one is a legitimate realization choice, and choosing among them is a management decision, not a technical one.
The difficulty with door three is not that the value is softer. It is that the value does not appear on a P&L by itself, which means it has to be defined and measured on purpose or it reverts to door four.
Public sector efficiency should not be a synonym for budget cuts. It should mean producing more public value from the resources government already has.
A mission gain that nobody measures is just as easy to lose as a financial one.
The Australian precedent
Since 1987 the Australian Commonwealth has applied an efficiency dividend to public service agencies: an annual reduction in the resources available to an organization, generally between 1 and 1.5 percent.
The sequence is what makes it instructive. The government banks the assumed efficiency first and asks the organization to produce it second.
That solves realization by brute force and creates a different problem. If the improvement does not actually exist in a given agency in a given year, the reduction still comes from somewhere, and it may come from capacity that was doing real work. Critics have long called across the board dividends a blunt budgeting instrument.
Door four wastes productivity. Forced door one invents it. Neither is management.
Why the bill is arriving now
For the first phase of enterprise AI, weak realization was affordable. Pilots were small. Innovation and one time funds absorbed the cost. Enterprise software contracts obscured marginal consumption. Vendors subsidized adoption to win accounts.
Stacia Garr's argument in Harvard Business Review this month is that the period is ending, as vendor subsidies decline and usage based pricing becomes the norm. AI stops being a license buried inside a technology budget and becomes a visible, variable consumption cost that competes directly with labor.
Variable cost is unforgiving of vague benefits. It arrives monthly, it grows with success, and it is trivially easy for a finance director to isolate.
The invoice is becoming variable faster than the benefit is becoming measurable.
Your AI is a variable cost. Your business case is fixed. One of those is wrong.
The instruction
Take your three largest AI deployments. For each one, answer four questions.
What was the baseline?
What measurable capacity has AI created?
Which door is that capacity supposed to go through?
Which executive owns making that conversion happen?
If you cannot answer all four, you do not have an AI ROI case. You have an AI productivity claim.
Decide where the gain goes before you know how big it is.
Sources: McKinsey, The State of AI, 2026, n=1,719, surveyed May to June 2026. Code for America, 2026 Government AI Landscape Assessment, May 2026. Australian Department of Finance, efficiency dividend. Stacia Garr, "How to Respond to the Coming AI Cost Shock," Harvard Business Review, August 2026.