The short version
Enterprises have spent tens of billions on AI tools and still can’t show the productivity gains. The debate about why keeps coming back to methodology: weak baselines, contested studies, mixed results.
The simpler explanation is that most of that spending was never a productivity bet.
- What was bought: a large share of enterprise AI spend is career insurance. Executives buy it to avoid being the one who said no to the platform shift, pay for it with company money, and describe it in productivity language because that is the only vocabulary a budget approval accepts.
- Why it can’t be measured: you can’t find productivity gains in a purchase that was made to remove personal risk. That value arrives the day the contract is signed, when the board’s “what is our AI strategy?” has an answer.
- What to do: split AI spending by what each item was actually bought to do, and give each kind of bet its own test. Cost-savings bets need baselines. Insurance bets need a yearly review. Neither needs a single blended productivity number.
Five patterns you can check yourself
The career insurance explanation isn’t based on a survey. It rests on patterns that only make sense if productivity was not the goal. Each one can be checked inside your own organization.
| Pattern | What a productivity buyer would do | What usually happens |
|---|---|---|
| The return is always two years out | Set a target inside the current review cycle, so the bet can be judged by the people who made it. | Gains are projected 24 months out, safely beyond the next performance review. |
| No baseline | Measure the work before rollout. It takes a few weeks of analyst time. | Deploy first and attribute whatever improves afterwards. Ask for the pre-deployment baseline and most programs can’t produce one. |
| Uniform spend | Spend most where the threat or the opportunity is largest. | Per-seat spend on horizontal tools looks much the same in a packaged-goods company, a fintech and a government contractor. That is a tax, not a thesis. |
| Licenses before use cases | Find the highest-value workflows and start there. | License everyone, including functions with few documented use cases, then run internal campaigns to drive usage. |
| Weak results, renewed anyway | Cut tools that don’t deliver at renewal, as with any other software. | Check your own renewals. If weak results were followed by the same or a larger contract, the value was never the productivity. |
A real productivity buyer instruments the baseline aggressively, because the baseline is what proves the bet. When nobody bothers, the baseline was never the point.
The career risk calculus
The mechanism is an asymmetry in personal risk, and it explains both the size of the spending and why it persists.
Look at the bottom-left cell. An executive who approves the spend is safe whichever way AI turns out. An executive who declines is safe only if AI turns out not to matter, and even then the reward is a quiet “you were right.” If AI does matter, they become the leader who missed the most important platform shift in a generation.
The calculation a rational executive runs doesn’t need the productivity gains to be real. It only needs the downside of not adopting to look worse than the cost of adopting. At current license prices, in the current board and investor climate, it does for almost every senior leader. The spend is individually rational even when it is organizationally wasteful.
That is an incentive problem, not a character flaw. The people deciding carry asymmetric personal risk, and the currency of that risk is their professional reputation, not the company’s money.
Why every memo says “productivity”
The people approving these budgets aren’t confused about their motives. They are translating them into the only language the approval process accepts.
Capital allocation runs on discounted cash flow. Money goes in, savings or revenue come out, the net present value is positive, the investment is approved. Labor efficiency, developer throughput and cost per unit fit that model. Board reassurance, competitive narrative and an executive’s personal risk do not, so they can’t appear in a budget memo as they are.
So “we need investors to see us as an AI company” becomes “projected developer productivity gains of 25% over 24 months.” And “I can’t be the CTO who called AI a distraction” becomes a paragraph on competitive displacement risk.
Nobody is committing fraud. Organizations convert sensitive or embarrassing motives into legible categories all the time. Everyone involved understands the subtext, and nobody writes it down.
Vendors reinforce the translation. Microsoft, Anthropic, OpenAI and Google all supply ROI calculators, productivity benchmarks and TCO models, because that is the paperwork their buyers need to get a contract signed. Nobody sells a board-narrative calculator, even though that is what many buyers are really purchasing.
Where the savings are real
There is a real cost-savings case. It is narrow, and it is worth protecting from the broader critique.
It works when the task is high-volume and low-variance, and when the people doing it cost more per unit than the AI does. Examples include tier-1 support deflection with measurable ticket volume and handle time, contract triage where the classification is well defined, and extraction from structured documents where every output can be checked against a known answer. In these workflows the causal chain is short, quality can be audited, and the effect on headcount can be tracked.
The public evidence matches that shape. In a controlled experiment, developers using GitHub Copilot finished a bounded coding task 55.8% faster than a control group. Klarna reported that its AI assistant handled two-thirds of customer service chats in its first month, the equivalent work of 700 full-time agents. Both are specific tasks with clear before-and-after measures.
These gains are real. They are also a small share of total enterprise AI spending, they come from specific workflows rather than organization-wide rollouts, and they are routinely cited to justify spending far outside the conditions that produced them.
The debate is measuring the wrong purchase
The researchers who find weak productivity evidence and the executives who privately admit the numbers don’t add up are both right. They are measuring the right thing for a different purchase than the one that was made.
An insurance policy has no ROI to measure. Whether developers are 25% faster in two years is irrelevant to whether the purchase did its job, which was to give the board an answer.
So the productivity debate, held on its own terms, can’t be settled. The evidence will stay mixed because the spending was mixed: some genuine automation, some valuation narrative, some talent signaling, and a large share of career insurance that was never going to show up in a productivity metric.
The useful question is narrower: which part of our AI spending was meant to produce measurable gains, and are we measuring that part properly? For most organizations spending at scale, the honest answer is no.
Separate the portfolio
Naming the real purpose of each purchase doesn’t require anyone to confess bad faith. It requires sorting the spending into the different kinds of bet it already is, and judging each kind on its own terms.
| Kind of bet | Typical example | How to judge it | When to review |
|---|---|---|---|
| Cost savings | Automation of a high-volume, well-defined task | Measured change against a baseline captured before rollout | Quarterly, against the baseline |
| Revenue | A model inside a customer-facing workflow | Revenue or conversion attributable to the feature | Each product cycle |
| Risk mitigation | AI-assisted compliance monitoring | Did the feared event happen, and did the capability gap close? | Annually, like any insurance |
| Legitimacy | Copilot-style licenses for all knowledge workers | Did it clear the bar the board or market set? | Once, then stop reporting ROI on it |
| Option | An API sandbox for an R&D team | Capability built and what was learned, not output | At each funding decision |
Tracking all of this as “AI investment” against one productivity target measures none of it correctly. It applies a single financial test to a portfolio of bets that mostly aren’t financial bets.
If an organization won’t build measurement that matches what it actually bought, including the uncomfortable step of saying what that was, it should stop expecting productivity metrics to justify the spending. They weren’t designed to, and they won’t.
The organizations that are honest about what they bought will make better AI investment decisions than those still arguing over which productivity metric is least flawed.
If you want to sort your AI spending this way and build the baselines for the parts that are meant to pay back, the AI readiness assessment ends in a ranked roadmap with cost and payback for each item, and AI Portfolio Explorer keeps that portfolio scored and governed from then on. Or start a conversation.