How to build an AI ROI model your Series C board will trust

This week one number did the talking. Anthropic filed confidentially to go public, and on June 5 CNBC’s tech column called the listing the first big test of AI boom valuations. The same day, the AI to ROI newsletter put the company’s run rate at a 47 billion dollar annualized pace, up from about 9 billion at the end of 2025, with its CFO citing historic demand.

Here is what happens next inside a company. A board member reads that headline, forwards it, and asks a deceptively simple question: our AI spend keeps climbing, what did it return. I have watched founders freeze on that question, not because the answer is bad, but because nobody built the model that answers it. The freezing is the tell. This is how to build that model before the Q3 board pack is due.

How to build the AI ROI model, step by step

A model a board trusts is not a slide claiming AI saved 30 percent of someone’s time. It is a chain: a baseline measured before launch, a revenue or cost number that actually moved, the cost to run and govern the thing subtracted out, and a named human who signs the before and the after. Six moves get there.

1. Set the baseline before the agent ships

   The speed of any honest ROI read tracks how measurable the metric was before deployment. When a number already exists, an agent's effect shows up in weeks. When it does not, the company guesses for quarters. Write down the current value of the one metric this AI is meant to move, dated, before a single prompt runs.

2. Measure whether revenue sticks, not whether people log in

   Adoption is the easiest number to grow and the easiest to fake. The 2026 Benchmarkit data is the warning: across 342 SaaS and AI-native companies, gross revenue retention just fell to 84 percent, the largest single-year drop in the study's history, even as AI features shipped everywhere. Pick the metric that proves the revenue held, not that the dashboard got clicks.

3. Subtract the run-and-oversee tax

   Gross value is not the number. Token and compute bills are now a finance-level surprise, and the monitoring and governance layer bolted on to keep agents safe is a real cost too. Net AI value is gross value minus what it costs to run and watch. A model that reports only the savings and hides the bill will not survive one sharp board question.

4. Separate the vendor's revenue from the return

   Anthropic running at 47 billion is the vendor's P&L, not proof of anyone's return. Confusing the two is the most common category error I see in board decks. The model measures the change inside the business, attributable to one workflow, not the growth of the company that sold the tool.

5. Attribute to a workflow with a named owner

   Company-level AI ROI is unprovable and everyone in the room knows it. Workflow-level ROI is provable. Tie each AI bet to a single workflow, a before number, an after number, and one human who owns the result. Five clean workflow stories beat one heroic company-wide claim every time.

6. Put it on one page the board can attack

   The output is a single page per bet: baseline, metric moved, gross value, cost to run and oversee, net value, and the kill criteria if net value goes negative. If a director cannot challenge a line on that page, it is marketing, not measurement.

Why most AI ROI models fold at the board

The mistake that looks smart is leading with the biggest gross number available. A team finds the workflow where AI saved the most hours, annualizes it, and walks into the board meeting with a triumphant figure. Then a director asks what it cost to run, what the baseline was, and whether that revenue renewed. The number folds.

The deeper error is measuring the wrong layer entirely. In this week’s roundup, MarketingProfs gathered the proof gap in one place: MIT found 95 percent of AI pilots deliver zero measurable P&L impact, S&P Global found 42 percent of companies abandoned most of their AI projects, and IBM put initiatives delivering expected ROI at about 25 percent. None of those failures are model-quality failures. They are measurement failures. The pilots may have worked. Nobody set the baseline, so nobody could prove it.

That 84 percent is the line I would stare at longest. AI features went into nearly every product this year, and the revenue retained got worse, not better. Shipping AI and keeping revenue are not the same achievement, and a board that has read the same data will know the difference.

A board does not fund the promise of AI anymore. It funds the proof, net of the bill.

What an AI ROI model should measure

What good looks like is boring on purpose. A clean board view of a single AI bet is a before number, an after number, the cost to run and oversee it, and the net. Here is the same bet reported two ways.

For benchmarks to anchor the page: the 2026 Benchmarkit set put median Rule of 40 at 25 percent and median CAC payback at 16 months, so a board already lives in a world of disciplined, netted numbers. An AI line that arrives without a baseline or a cost row looks out of place next to those. The fix is not a better model. It is a baseline written down before launch and a cost row nobody is allowed to delete.

The first AI ROI move before the Q3 board

If there is one move to make before the Q3 board pack, it is the baseline. Pick the three AI bets that matter most, find the metric each is supposed to move, and write down today’s value with today’s date. That single act, done before anyone touches the model, is what separates a board update that holds from one that folds.

The Anthropic headline made AI return a public-market question this week, which means it became a board question inside every company by the weekend. That is not a reason for dread. It is permission to do the unglamorous thing well: measure the baseline, net out the bill, name the owner, and put it on one page. A board that can challenge every line is a board that can defend the spend. That is a much calmer room to walk into than the one holding a slide full of gross numbers and a prayer.