The 12% and the 48%

Everyone is watching the cost side of the ledger. Almost no one is watching the value side — and the firms reporting real returns are the ones who do. The gap between 12% and 48% is the whole thesis in one number: ROI is not something AI produces, it is something a management practice produces. The harness data says the same thing from the other direction: the firms with returns built the controls first.

Post: "The 12% and the 48%." Lead with 61% reviewing cost at approval, 59% monitoring it in operation, 12% checking value. Close with the one habit that moves a firm from the first group to the second: put a value metric next to every cost line before approval, not after the pilot.

Investors are now betting that "prove it pays" is a budget line, not a nice-to-have. The irony worth saying out loud: the announcement's own ROI figures can't be checked, which is exactly the problem the product claims to solve. Any percentage without a baseline is a claim, not a result.

Short post: "Before you buy an AI ROI tool, ask what baseline it measures against." Use the 47/24/86 figures as the example of numbers that sound precise and can't be audited, then give readers a three-question checklist: baseline, sample, window.

Firms deployed AI where it is easy to deploy, not where it is easy to measure. Writing and summarization tools run at 87% of these lenders; underwriting support, fraud detection and credit analytics barely register. In a regulated business the control that decays fastest is the one that matters after launch — written policy is at 87%, ongoing monitoring at 58%, board reporting at 45%.

Carousel for regulated-industry executives: "Where AI is deployed vs. where it pays." Two columns — easy to deploy (87% productivity tools) against hard to deploy but measurable (underwriting, fraud, servicing). Finish with the 87% / 58% / 45% policy-to-monitoring-to-board drop-off as the regulated-industry kicker.

"We can attribute it" and "we track it informally" cannot both be true. This is the over-attribution problem — leaders crediting AI for results without the baseline to prove it. The tell is the C-suite-versus-manager gap: the people closest to the work report less advanced adoption than the people reporting upward.

Post: "Perception vs. proof." Open with 72% versus 50% as a same-page contradiction, then ask one question: could your AI result survive your CFO asking for the baseline?

The use cases that got picked were the ones with a measurable baseline — uptime, batch records — and the blocker everyone named was data, not AI. That is a better ROI predictor than any model benchmark: choose the workflow where the "before" number already exists.

Post or workshop exercise: "Name the before-number." Ask each participant to write the baseline metric for their AI use case in one line. Anyone who cannot has found the real first project.

Form ADV is where an adviser says in writing what it does with AI, and the smaller the firm, the quieter the brochure. The exposure is rarely using AI — it is the gap between what a firm does in practice and what its disclosure says. Closing that gap costs an afternoon and is the cheapest compliance win in the AI conversation.

Post for advisers: "Does your Form ADV match your tech stack?" Offer a three-line self-check: which tools touch client data, where AI touches client communications, and what the brochure currently says about both.

Pilots rarely stall for lack of results. They stall because nobody agreed in advance what result would justify the next dollar. Writing the decision date and the decision-maker into the pilot before it starts is a cheaper fix than any measurement tool.

Post: "Write the ending before the pilot starts." Show a one-page pilot skeleton — metric, baseline, 12-week window, decision date, who decides — and invite readers to fill it in for the AI project they are funding now.