AI ROI Brief

Financial services self-reports the most aggressive AI returns of any sector, while the leading cross-industry survey puts organizations with significant earnings impact at six percent. Both findings can hold only if financial services is genuinely well ahead — or if the people who answer an AI vendor's survey are the ones who already bought in. Treat the optimistic number as a ceiling, not a baseline.

Put the two figures side by side in a single graphic — 64% and 6% — and name the difference in survey population out loud. The lesson for executives is methodological: before you benchmark against an industry number, ask who was asked and who paid for the asking.

For a five-billion-dollar firm, '0.5 to 1 percent of AUM in incremental inflows' is twenty-five to fifty million dollars a year in new assets. That is a growth number, not a productivity anecdote — and it reframes AI from a cost-center project into a distribution project, which is a conversation the people who run the firm will actually sit still for.

Run the arithmetic live against three firm sizes, then turn immediately to the catch: leaders capture three times what laggards do, and the difference is operating model, not software licence. The math is the hook; the gap is the lesson.

This carrier did not buy one AI product. They built eighty models against a single domain, and the return shows up as twenty-three days off liability assessment and a sixty-five percent drop in complaints. That is the shape of real AI ROI: one narrow domain, many bounded jobs, one measured output. The organizations that rolled a general-purpose assistant out to everybody got the opposite result.

Workshop exercise: have each executive name the single bounded, high-volume, document-heavy process in their business that already has a measured cycle time — then estimate what twenty-three days off it would be worth. The carrier example works as the anchor because the numbers are specific and dated.

Here is the cost story that did not make it into most business cases: agentic AI consumes five to thirty times the tokens of a chatbot for the same task, and a fifth of enterprises say those bills are already limiting what they will run. Spend that used to go to predictable per-seat software licences is now going to variable per-task consumption. An AI budget without a unit-economics line is not a budget.

Explain the shift from fixed seat licences to variable inference cost in plain language, then hand executives three specific questions to put to their CFO before approving the next agentic pilot. This is the least-covered finding in an otherwise heavily-covered report.

Eighty percent of your people will tell you AI made them faster. Six percent of companies can find that speed on the income statement. The time saved is real — it is simply sitting on your employees' desks rather than flowing into more deals closed, more claims resolved, more work shipped. Productivity you do not reinvest is not productivity. It is slack.

Lead with the 80-versus-6 split as a single line. Explain the mechanism in three sentences — time saved stays where it was saved — and close on the one behavior that separates the two groups: roughly three-quarters of high performers fundamentally redesigned workflows, against roughly one-quarter of everyone else.

The wealth management version of the ROI question is not 'how much do we save,' it is 'how many more real relationships can one advisor carry.' But look closely at what got measured at that top-ten manager: hours saved and preparation time. Those are inputs. Nobody published the revenue those twenty thousand hours produced. Every AI business case that stops at hours saved is unfinished.

Build the piece around one question — 'twenty thousand hours saved, saved into what?' — then give the audience the three output metrics a firm should be tracking instead of hours. The distinction between input and output metrics is the practical skill most executives leave an AI conversation without.

Executives have now overestimated AI-driven headcount reduction two years running — thirty-two percent predicted cuts, fourteen percent delivered them. The technology, retraining and institutional inertia required to actually remove a role at scale is far harder than the boardroom slide suggests. Anyone selling AI on a headcount-reduction business case is selling the least reliable number in the category.

One chart, two numbers: predicted versus actual workforce reductions. Close on the point that the honest ROI case for AI is capacity and quality, not headcount — deliberately counter-programming the displacement content saturating most feeds.

Leakage is the most underrated AI ROI line in insurance — a thirty to fifty percent reduction, and unlike headcount it is a number finance already tracks, already dislikes, and already has a baseline for. If you want an AI business case that survives contact with the CFO, find the metric they were measuring before anyone said the word AI.

Workshop exercise: have each executive list three metrics their CFO reports monthly and already considers a problem. That list — not the vendor's use-case catalogue — is where the first AI deployment should go. Leakage serves as the worked insurance example.

Here is a compliance question almost no boardroom has asked yet: is your AI deployment actually covered? Most standard liability policies now carve AI out. That means the cost side of your AI business case is missing a line — and unlike token spend, this one only appears after something has already gone wrong. Lloyd's is writing against the gap, which tells you the market has priced a risk your board has not.

Short piece titled for the executive who has never thought about it: 'The AI line item your insurance broker hasn't mentioned.' Works equally well as a newsletter section or as a risk-quantification module in an executive readiness session, since it converts an abstract worry into a priced, checkable question.

Inference became roughly two hundred and eighty times cheaper in under two years. Whatever advantage you believed you were buying with model access, you were not. The only thing that compounds is proprietary data that changes what your product does: use generates data, data improves output, better output attracts more use. If your AI initiative does not close that loop, you have bought a very expensive commodity.

Explain the loop in three steps, then end on a diagnostic rather than a lecture: name the data your business generates that nobody else has. For most executive audiences that question produces a longer silence than any statistic will.

The most reliable AI ROI case we have seen this year is one we ran on ourselves: build the deliverable from the client's own information, hand it over before anyone signs anything, and let the artifact do the selling. It converts because it is not a claim about what AI could do — it is a demonstration on their business, dated and specific.

Walk the pattern end to end without naming any client: the handful of discovery questions that reveal what someone is already straining to keep up with, the personalized brief that comes back, and the conversation that follows. The value of the piece is that it is reproducible by the reader, not just admirable.