The ROI argument is really an attribution argument
Most enterprise AI ROI disputes are not about whether a model works. They are about whether anyone can trace a dollar of saved cost or new revenue back to a specific AI decision. A CEO sees a demo that drafts contracts in seconds and assumes the savings are obvious. The CIO knows the savings are invisible until someone instruments the workflow, counts the decisions the AI actually made, and shows which ones a human still had to redo. Until that point you have a story, not a return. The teams that win the budget fight are the ones that decided, before launch, exactly which event they would count as value and how they would capture it.
Pick use cases by measurability, not by excitement
The use case that demos best is rarely the one that proves out first. A good first candidate has three properties: the task happens often enough to produce a sample size in weeks, the outcome is observable without a six-month study, and a baseline already exists to compare against. Renewal-clause review, support-ticket triage, and invoice coding all qualify because the before-state is already measured. A vague 'productivity' play across the whole org does not, because there is no clean baseline and no single owner who can be held to the number. Rank your candidates on how cheaply you can attribute their results, then start at the top.
Instrument the decision, not the token count
Vendor dashboards tend to report usage: prompts sent, tokens consumed, seats active. None of that is value. Value lives one layer up, in the decision the AI influenced and what happened next. For each governed use case, capture the input, the AI output, the human action that followed (accepted, edited, rejected), and the downstream result. That record is also your audit trail, which means the same instrumentation that proves ROI to finance is what lets you defend the system to a regulator. Build the measurement and the control together; retrofitting either one later costs far more.
Separate gross value from the cost of running it
An AI workflow can create real value and still lose money, because the spend is unbounded by default. An agent that retries, calls tools, and chains steps can quietly run up an invoice that swamps the saving. Track cost per governed use case next to the value it produces, not as a single platform line item. When you can show finance a per-use-case P&L, the conversation shifts from 'is AI worth it' to 'which workflows earn their keep,' which is the conversation you want. The ones that do not clear the bar get cut without drama.
Move into the hard accountability phase on purpose
There is a predictable arc. First a wave of experiments funded on faith. Then a reckoning when finance asks for the number and most teams cannot produce it. Budgets in the second phase survive only for use cases with a measurable, governed outcome. You can wait for that reckoning to happen to you, or you can run your portfolio as if it already arrived: every funded use case carries an owner, a baseline, a value event, and a cost ceiling. The work of getting AI into production that delivers business value and the work of proving the return are the same work done well.
What good looks like after one quarter
A defensible ROI position at the ninety-day mark is concrete. You can name the two or three use cases in production, show the baseline each beat, point to the captured decision records behind the claim, and report cost per use case against value created. You can also show the ones you killed and why. That portfolio view, value proven where it exists and spend cut where it does not, is what earns the next round of funding. It is also the foundation of a governed AI operating model rather than a pile of disconnected pilots.
Frequently asked questions
Why is enterprise AI ROI so hard to measure?
Because the value sits inside individual decisions and most teams never instrument those decisions. Without a captured record of what the AI influenced and what a human did next, there is no way to attribute a saving or a gain, so the return stays anecdotal.
What is a realistic timeframe to show AI ROI?
Choose use cases with a high decision frequency and an existing baseline and you can show a credible signal within a quarter. Broad, infrequent, or baseline-free use cases can take a year or longer and often never produce a clean number.
Should AI ROI include the cost of governance?
Yes. Count the cost of running and controlling each use case against the value it creates. Governance is what lets the workflow reach production and stay there, so its cost belongs in the same per-use-case P&L as the model spend.