The Two Percent Problem

This is the clearest industry-specific version of the number every AI ROI conversation eventually has to reckon with. Three out of four AI projects at the largest insurers in the world make something faster or cheaper. Roughly one in fifty makes something more money. If your board is measuring AI success by pilot count, this is the stat that reframes the question: faster and cheaper than what, and does it show up anywhere your CFO already tracks?

A short post titled "75% and 2%." State both numbers with no editorializing, then ask one question: which of those two your own AI budget is actually optimizing for. Works as a single-slide visual — two bars, one tall, one nearly invisible.

Notice what all three have in common: a real number today, and a much bigger number several years out. That gap is the honest version of an AI ROI roadmap — nobody credible is claiming the return arrived all at once. The dishonest version is a vendor pitch that skips straight to the 2030 number and calls it current performance.

A comparison table for board decks: company, current benefit, target benefit, years to target. Use it as a template a client fills in with their own numbers — most can't, and that gap is the conversation starter.

The savings are real and the reversal is just as real, and both facts matter more together than either does alone. Klarna didn't discover AI doesn't save money. It discovered that a cost number without a quality floor underneath it is a countdown to a correction. The question for any cost-savings business case isn't "how much did we save" — it's "what did we agree we'd protect no matter what, and did we actually protect it?"

Short video or post on "the Klarna correction" as a cautionary case study — not "AI cost-cutting fails," but "cost-cutting without a quality floor fails, and AI just made the mistake move faster." Pair with a one-question audit: what's the metric you'd notice degrading before your customers tell you?

This is the shape of cost-side AI that actually survives scrutiny: narrow scope, a specific document type, a specific decision it supports. Compare that to Klarna's story above — the tools that hold up under pressure are the ones aimed at one job, not "customer service" as a category.

A short piece on "the narrow win" — contrast broad AI cost-savings claims (headcount, "efficiency") with narrow ones (a specific contract type, a specific clause) and argue the narrow ones are the credible ones right now.

Individual productivity and enterprise earnings are not the same metric, and this is the clearest evidence yet that most organizations are only measuring the first one. The 6% who show up on the second metric didn't get there with a better model — they got there by redesigning the workflow around the tool before they scaled it, and by deciding what "success" meant before they started. That's a sequencing problem, not a technology problem.

Workshop exercise: before naming a single AI tool, have a leadership team answer three questions in writing — what workflow are we redesigning, what's the baseline we're measuring against, and who owns the number. Most teams can't answer any of the three. That gap is the actual readiness assessment.

This is the missing half of every AI productivity conversation. Training completion is not the same as judgment. If the people accountable for an AI-assisted decision were never taught how to explain or defend it, they will do the rational thing under pressure: quietly take more or less credit than the AI actually deserves. That's not a compliance footnote — it's a direct threat to whether the productivity number you're reporting is even true.

A piece for HR and risk leaders: "Ask your team this one question — can you explain, out loud, why the AI's recommendation was right or wrong in the last case you used it on?" Most can't. That's the training gap this research is describing, and it's cheaper to close than most people assume.

Four companies, four boards, one question underneath all of it: did the people approving the AI strategy actually understand — and disclose — what they were exposed to? That question used to be theoretical for most boards. It no longer is. The insurance and legal markets have both started pricing it, which means it's now a fiduciary question, not just a technology one.

A board-briefing one-pager: "Four questions your D&O carrier is already asking about your AI strategy." Use the four suits as the setup — training-data provenance, internal dissent handling, spend disclosure, and named accountability — and frame each as a question the board should be able to answer today, not after a suit is filed.

$1.5 billion is no longer an abstraction — it's a real settlement with a real per-unit price attached to a specific kind of exposure: not knowing, or not disclosing, what your AI was actually trained on. Every board approving an AI vendor relationship this quarter should be asking the provenance question this settlement makes concrete: what's actually in the training data, and who verified it?

A short explainer: "What a $1.5 billion settlement actually buys you in due diligence." Walk through the specific questions a board or GC should ask an AI vendor about training-data provenance, using this settlement as the reason the answer now has a real price tag.

This isn't really an AI story — it's an employee-mobility story that happens to involve two AI companies. But it's the cleanest recent illustration of a risk every regulated client should be thinking about: the exposure isn't only the AI tool your employees use, it's what leaves with an employee who's used AI tools trained on, or alongside, your proprietary information.

A short piece for GCs and CHROs: "The AI exit-interview question nobody's asking yet." Use the Apple/OpenAI case as the hook, then propose one concrete addition to offboarding checklists for any employee with AI-tool access to proprietary data.