Stop Asking "How Much AI Are We Using?"

Most companies are still measuring AI adoption the wrong way, and this session named the problem in one clean before-and-after. It's a short session, six slides, but it's one of the tightest frameworks from the whole conference.

Change the Question, Change the Outcome

The old question companies have been asking: "How much AI are we using?" That question measures adoption, activity, and motion. It rewards consumption. It produces leaderboards — impressive-looking dashboards that don't actually tell you if anything got better.

The 2026 question, the one this session argued should replace it entirely: "What does each outcome cost, and what is it worth?" That question measures value per outcome. It rewards results. It produces decisions instead of vanity metrics.

The presenter's framing of the shift: "We have gone from the adoption cycle to the value-maximizing cycle." That's a genuinely different way of running an AI program, not just a rhetorical flourish.

Count the Whole Iceberg: The Token Bill Is Only What You Can See

This is the most useful reframe in the whole session. Most ROI conversations start and end with the token bill, because that's the number that's metered, visible, and sitting right there on the invoice. But that's only the part above the waterline.

Below the waterline, invisible on any invoice but very real in cost: platform and integration engineering, retrieval and vector storage, tool and browser sessions, observability and evaluation infrastructure, human review of what the agent got wrong, rework and retries after failure, and governance, risk review, and change management.

Why this matters for the math: rate-card ROI compares token spend to value. Real ROI compares the whole iceberg to value. A "cheap" agent with a heavy review burden and a 40% failure rate is one of the most expensive systems a company owns, the cost is just booked under three other budget lines instead of the AI line. The session's guidance: budget successful outcomes, not advertised prices. Model the whole workflow, not just the rate card.

Discipline One: Architecture — Stop Paying to Rethink Solved Problems

Three specific moves, in order of impact:

1. Route by task. Frontier models for frontier problems. Small, fast models for classification, extraction, and routine steps. Routing is described as the single biggest lever available.
2. Cache what repeats. Prompt and context caching cut re-read costs dramatically. A general agent re-derives known answers every time. A tuned one starts from what it already knows.
3. Narrow the agent. Purpose-built agents with pre-paid thinking beat general agents that plan from scratch on every single run. Constrain loops, cap retries.

The illustrative number from the session: the same monthly workload, run through one general agent versus a routed, cached, and constrained setup, dropped from $48,000 to $9,000, with equal or better task success. That's not a marginal optimization. That's an 81% cost reduction for the same outcome quality.

Discipline Three: Governance Belongs Inside the ROI Equation, Not Next to It

The session's formula for governed ROI:

Governed ROI = (Realized Value − Risk-Adjusted Downside) ÷ (Token Cost + Governance Cost + Platform & Integration Cost)

Three reasons governance changes the math directly, rather than sitting off to the side as a compliance line item:

• It multiplies value. Trust drives adoption, and adoption is what actually drives value. Ungoverned tools get quietly abandoned by the people who don't trust them, no matter how capable they are.
• It prices the downside. Governance cost should reflect probability times the cost of a harmful or non-compliant action. In healthcare specifically, one bad outcome could erase a full year of savings.
• It scales by risk tier. Governance cost should be a step function of risk, not a flat tax on volume. Heavy review where risk is genuinely high, lightweight review everywhere else.

The One Metric Worth Tracking Next Quarter

The session closed with a single recommendation, deliberately simple: if you measure only one thing next quarter, measure what a successful, governed outcome is worth, divided by what it truly costs to produce.

Each of the three disciplines feeds that one number differently. Architecture lowers the cost of each outcome. AI FinOps makes the cost visible and attributable in the first place. Governance raises what each outcome is actually worth by protecting the trust that drives adoption.

What This Means for Any Company Running AI at Scale

If your current AI reporting is a usage dashboard, logins, queries run, tokens consumed, you're measuring the wrong cycle. None of that tells you whether the outcomes were worth what they cost, review burden included. Before your next budget conversation about AI, try running the iceberg exercise: pick one AI workflow and actually tally what's below the waterline. Most teams are surprised by what they find.

Not sure what your AI workflows actually cost once review time and rework are counted? Winsome helps companies find the real number behind the token bill. Talk to Winsome about your AI ROI.