The AI Metrics on Your Board Pack Are the Wrong Metrics

The slide will look something like this. AI Adoption: 87%. Copilot Seats Deployed: 4,200. Average Prompts Per User Per Week: 23. Employee Satisfaction with AI Tools: 4.1 out of 5. Hours Saved Per Week, Self-Reported: 6.3. Estimated Annual Productivity Gain: $14M.

The CFO will have signed off on the number. The CHRO will have a slide on adoption velocity. The CTO will have a slide on governance. Somewhere in the deck there will be a logo grid of the AI tools the company has rolled out in the last twelve months. The narrative arc of the section will be: we moved fast, we got broad adoption, employees love it, the productivity gain is material. Every board in the room will nod. The investors will note the strategic alignment with AI. The CEO will move to the next section.

The problem is that almost every number on that slide measures something the company does not durably own. The deeper problem is that the slide is telling the story of a bet the company has not consciously made.

Before getting to the bets, there is a more basic question that almost no board pack addresses. When an AI deployment captures value, where does the value actually accrue. The question sounds abstract. It becomes very concrete once you ask, for any specific deployment, what happens when the employee leaves.

There are four kinds of AI deployment currently visible inside large companies, and they have four different answers.

The first is the company brain. One AI, organisation-wide, with permissioned access to internal information. Anyone can ask it questions. Institutional memory finally made searchable: what has been tried, who knows what, what was decided and why. When the employee leaves, the company brain holds. The institutional knowledge that has been ingested, structured, and made queryable stays with the organisation. New employees onboard against it. The value compounds with usage and persists through turnover. The asset is the company's.

The second is the individual copilot. Every employee gets their own AI assistant for drafting, summarising, researching, thinking. Each person becomes meaningfully more productive at the work they were already doing. When the employee leaves, the copilot value walks out the door. The prompts the employee learned to write, the workflows they built around the tool, the judgment they developed about when to trust the output and when to redo it: all of that is portable. It goes with them to the next job, where they will be more productive on day one than the person replacing them. The asset is the employee's.

The third is the functional deployment. AI gets dropped into the function where it produces the most visible lift. Sales (call coaching, deal scoring, follow-up drafting). Marketing (content at scale, ad variants). Customer support (tier-one deflection). When the employee leaves, the value sits somewhere in between. Some of it is in the configured workflow, the integrations, the institutional knowledge of what works for this company's sales motion. Some of it is in the individual rep's learned ability to use the tool well. The split varies by how the deployment was built. If the company bought a SaaS tool and let each rep use it however they wanted, the value mostly walks. If the company invested in the underlying workflow design, more of it stays.

The fourth is the agent deployment. AI does the work, end-to-end, without a human in the loop for the bulk of it. The customer support ticket triaged, answered, and resolved before a human sees it. The reconciliation that closes its own loop. The fraud case adjudicated without human review. The agent is not assisting an employee. It is replacing a category of work. When the employee who built the agent leaves, the agent itself stays. The value accrues to the company, by construction, in the same way a piece of internal software does. The risk is different: not that the value walks, but that the company loses the capability to operate, debug, and improve the system it now depends on.

These four are not steps on a maturity model. They are four structurally different deployments with four different beneficiaries. The same dollar of AI investment can produce value that compounds inside the company, inside individual employees' careers, or inside the vendor that sold the deployment, depending entirely on where it gets deployed and who keeps the capability to operate it.

The current AI discourse keeps trying to sequence these as if they were a roadmap. They are not a roadmap. They are a choice about where the value lives.

Once the value-residence question is in view, the strategic question on top of it becomes clearer. There are three bets a company can be running with AI in 2026. Most boards are not asking which one they are on. The three bets look similar from the outside and even from inside the company. They have different operational shapes, different value-capture profiles, and different ways of failing. A board that cannot tell them apart cannot tell whether the AI strategy is working.

AI capacity gets converted into more work, more output, more revenue per employee. Headcount stays roughly flat or grows. Revenue grows faster. The story shows up as expanding operating leverage that is visible on the income statement.

This bet is the cleanest of the three because the value capture is observable. If revenue per employee is rising materially faster than the industry baseline, the bet is working. If it is not, the bet is not working, and no amount of seat-adoption metrics will obscure that fact. The growth bet is also the hardest to actually run, because it requires a credible path to absorb the additional capacity into more revenue. Early-stage software companies usually have that path. Mature enterprises in slow-growth categories mostly do not, which is why so few are running the growth bet honestly even when their board packs describe it that way.

The tell of a real growth bet is that the AI metrics on the board pack are leading indicators for top-line numbers that show up six to twelve months later. The tell of a fake one is that the productivity metrics keep going up while revenue per employee does not.

A company running the growth bet primarily through individual copilots is giving most of the value to its employees and then hoping the residue shows up as revenue. A company running it through agent deployments and product builds is keeping the value structurally. The same bet, deployed in two different places, produces two different outcomes.

AI capacity gets converted into reduced headcount at roughly flat revenue. Operating margin expands. The work that humans used to do is now done by agents, or by a smaller human team operating with agent support. This is the bet most large enterprises are actually running in 2026, whether or not their board packs describe it that way. It is the bet behind almost all of the visible AI-related layoff news.

The compression bet works when three conditions hold simultaneously. The work being removed has to be work an agent can actually do durably. The company has to retain enough internal capability to operate, debug, and improve the agent over time. And the institutional knowledge required to handle the cases the agent cannot manage has to be preserved somewhere rather than disposed of along with the headcount.

When the three conditions hold, the compression bet is the most powerful version of enterprise AI deployment, because the value capture is direct and visible. IBM replaced a few hundred human resources staffers with AI agents and consequently hired more programmers and salespeople. The HR back-office work was genuinely removed. The freed-up headcount was reallocated into functions where additional capacity converts into additional revenue. The compression on one side subsidised a growth bet on the other.

When one of the three conditions fails, the compression bet starts to unwind, often slowly. Klarna cut approximately 700 jobs and replaced them with AI-powered solutions between 2022 and 2024. CEO Sebastian Siemiatkowski publicly admitted that the AI-driven transition negatively affected service and product quality. Following increased customer complaints and operational issues, Klarna is now rehiring human staff. The work the agent was supposed to do turned out to require capabilities it did not yet have. The institutional knowledge required to handle the edge cases had been disposed of. The company is now rebuilding a capability it had publicly declared automated, against the backdrop of an IPO narrative built partly on the original claim.

Salesforce sits somewhere between these two cases, and the honest answer is that we do not yet know which version it is. Benioff revealed in late 2025 that he had reduced customer support headcount from around 9,000 to about 5,000, attributing the reduction to the efficiencies of Agentforce and stating that the company no longer needed to actively backfill support engineer roles. The headline compression is large. Whether the work is being done durably at acceptable quality, whether the freed-up headcount has been reallocated rather than simply eliminated, whether the support quality metrics hold over the next eighteen months: these are not questions the announcement could answer. They are questions that will be answerable by late 2027.

The IBM, Klarna, and Salesforce announcements look similar in the press release. They are three different points on the same bet, with three different probabilities of working, and the difference will only be visible later.

A company announces AI-driven layoffs that are, in significant part, cyclical headcount reductions repackaged with an AI narrative because the AI narrative plays better with investors than the cyclical one. The capacity claims are inflated. The durability of the substitution is overstated. The headcount comes back, quietly, under different labels, two to four quarters later.

This is harder to name from the outside because the announcements are indistinguishable from real compression bets. The same press release language, the same productivity claims, the same forward-looking statements about operating leverage. Forrester has noted that many executives are using "AI-driven restructuring" as a catch-all rationale for layoffs that may have been planned regardless of technological advancements.

The thing that distinguishes substitution theatre from real compression is not visible at the moment of the announcement. It is visible eighteen months later, in three signals. Whether service or quality metrics on the affected function have held or degraded. Whether headcount in the function, or in suspiciously adjacent functions like contractor spend, outsourced services, BPO contracts, or "transformation" cost categories, has crept back up. And whether the agent that was supposed to do the work is being operated by the company itself, or by a vendor whose monthly invoice now approximates the salary cost of the people who were laid off.

The reason theatre is worth naming as its own pattern, rather than as a degraded version of compression, is that the incentives that produce it are different. The compression bet that fails on impatience (the Klarna case) was a real attempt that ran ahead of the technology. Theatre is the case where there was never a real attempt; the AI narrative was always primarily about the announcement rather than the operational change. The two failure modes look the same but have to be diagnosed differently, and a company that confuses one for the other will draw the wrong lessons from its own rehires.

These three bets exist in every large company in 2026. Most companies are running more than one of them simultaneously, in different functions, with different levels of awareness. Almost no board pack distinguishes between them, because the reporting structure that exists was built for the growth bet, and is now being applied to compression-bet and theatre-mixed operations as if those were the same thing.

The $14M self-reported productivity number from the opening slide is a growth-bet metric. It expresses how much additional capacity AI has created. In a growth-bet company, that capacity gets converted into revenue and shows up on the top line, and the productivity number is a leading indicator of the income statement story. In a compression-bet company, the same metric is misleading. The capacity has not yet been converted into either more revenue or less headcount, so it sits in a kind of organisational limbo. Real in the sense that the work is genuinely getting done faster. Not real in the sense that it is not flowing through to either the top or the bottom line. In a theatre-mixed company, the metric is actively obscuring what is happening, because the productivity gain is being claimed in support of headcount reductions that were going to happen anyway.

The honest version of the board pack looks completely different for each of the three bets. The growth-bet pack shows revenue per employee against industry baseline, and the AI productivity metrics as the leading indicator of that ratio. The compression-bet pack shows work units removed from human workflow, headcount in the affected function over rolling four-quarter windows, service quality on the affected function held against baseline, and the contractor and outsourced and BPO lines that would indicate a quiet rebuild. The theatre pack, if any company were honest enough to produce one, would simply show cyclical headcount reductions reported as cyclical, with the AI narrative removed.

Almost no Fortune 500 board pack in 2026 looks like any of these three honest versions. Most still look like the version at the top of this piece. They are reporting growth-bet metrics on top of operations that are some unknown mixture of compression and theatre, and the board cannot tell the mixture. And underneath every one of those metrics, the more basic question still sits unanswered: when the work captured by AI produces value, who keeps it.

Look at the AI section of your most recent board pack. Sort every metric on it into one of three buckets.

The first bucket is employee-behaviour metrics. Adoption rate, seat counts, prompts per user per week, satisfaction scores, hours saved per week (self-reported), training completion rates. These metrics measure what employees are doing with AI tools. They do not measure what AI has changed about the company's operations. Almost all of the value these metrics describe is currently sitting in the employees' careers, not in the company's economic architecture.

The second bucket is operational-change metrics. Work units now done by systems the company owns (tickets resolved without human touch, transactions processed end-to-end, cases adjudicated without human review), headcount in specific functions over rolling four-quarter windows, service quality on affected functions held against baseline, contractor spend and outsourced services lines tracked alongside the headcount reductions. The value these metrics describe is the value the company is actually capturing.

The third bucket is financial-outcome metrics. Revenue per employee against industry baseline, operating margin expansion attributable specifically to AI deployment, gross margin movement in functions where AI is being deployed, the line items where AI savings should theoretically be appearing in the income statement. These are the lagging confirmation that the value has flowed through.

If the first bucket is materially larger than the second and third combined, the board is being shown the growth-bet version of a story that is operationally a compression bet, and is not yet equipped to tell whether the compression is real or theatre. If the second bucket is the largest, the board is at least asking the right questions of the compression bet, though it may not yet have the eighteen-month track record to answer them. If the third bucket is the largest, the board is running a genuinely mature reporting structure for AI, and is almost certainly in a small minority.

The test is repeatable. Run it on this quarter's pack. Run it again in six months. This publication will return to the test in six months, with a follow-up on what has changed across the public-company disclosures that are most observable from the outside: the rehire signals, the contractor spend creep, the support quality holds or degradations. The companies that ran compression bets patiently will be visible by then. So will the ones that ran them on impatience. So will the ones that were mostly running theatre.

What would change my mind on the test itself. If the operational-change metrics turn out to be substantially harder to define cleanly than they look, if "work units removed from human workflow" becomes a category that can be gamed almost as easily as adoption metrics, then the test moves toward the financial-outcome bucket as the only reliable signal, and boards have to wait longer to know what is real. This is plausible. If a new reporting standard emerges from the audit firms or the SEC requiring some form of AI operational disclosure, the test becomes redundant because the disclosure will do the work. This is not currently on the horizon.

Absent those, the test holds. The board pack is the artefact. The buckets are the diagnosis. Most companies in 2026 are not making the bet they think they are making, because they have never been asked to articulate which bet they are making.

The slide at the top of this piece is the bet most companies are accidentally running. It is not the bet most CEOs would consciously choose. And the productivity gain it is reporting is real, but mostly not theirs.