The question this answers: when every results release mentions AI, how do you tell an AI result from an AI mention, and when should a pilot stop?
Bottom line: in the third-quarter results we read, AI was named everywhere and priced in two releases. The firms with the most usable numbers say where the money lands and measure it net of cost. Few firms have written down when a pilot should stop.
This week’s three findings
1. Insurers name AI in their results. Two of the five we read put money on it.
The number: about C$250 million a year of recurring pre-tax earnings is credited to AI at Intact Financial, with a target of about C$500 million within 24 months. Company-reported on its 4 November call, with no base period or method stated, and read from a summary of the call.
What we saw: Ping An credits smart fraud detection with RMB10.04 billion of claims savings in nine months, up 9.7% on the year, and says AI handles 82% of its customer-service volume (company-reported, from the full text of its nine-month release). The other three releases name AI and price none of it. Allstate’s release says its AI agent now handles millions of service calls, gives no saving, and its property-liability underlying combined ratio rose to 80.5 from 78.7. AIG’s general insurance expense ratio fell to 30.6% from 30.9%, and the release attributes none of it to AI. Hippo’s expense ratio fell 10.3 points to 41.7% while its loss ratio rose to 56.7% from 47.6%, with no figure crediting AI. The Allstate, AIG and Hippo figures are from the companies’ own releases; Intact’s comes from a summary of its call.
So what: insurers are a fair test because they report the ratios AI is supposed to move. Intact is the only one of the five that says where its figure lands: about 60% through the loss ratio and the rest through expenses. That is what makes a number usable, even though Intact gives no method. The others show why naming AI is not enough. Allstate describes AI at scale while its underlying combined ratio went up. Hippo’s expenses fell, but its losses rose faster, so the expense gain did not protect the result. A five-firm sample proves nothing about insurance as a whole. It does show the gap between AI as a topic on the call and AI as a line in the accounts.
Do this:
- Before an AI project starts, name the ratio or line it should move and write today’s value next to it.
- When a peer quotes AI at scale (calls handled, models in production), ask what happened to the cost line in the same quarter.
2. The targets worth copying are measured net of cost
The number: AXA targets €500 to 700 million a year of recurring pre-tax value from AI by 2029, which its investor-day deck footnotes as “net of investments”. A target from its 15 September investor day, not a result.
What we saw: Chubb plans up to 20% fewer staff out of about 43,000 over three to four years, with AI and automation covering 85% of major underwriting and claims processes and about 1.5 points off the combined ratio (an investor-presentation target, second-hand). Gartner’s survey of 2,002 CIOs, as published, finds only 13% report significant or transformational value from AI, 58% face more pressure to deliver AI cost improvements, and 51% expect AI to raise total cost of ownership.
So what: most AI business cases count the gain and leave out the running cost of the tools, which is exactly the cost half of CIOs now expect to grow. AXA’s footnote points the right way: the value is counted net of what it invests to get it. Chubb’s target is the other useful shape, because it names two things that can be checked later, a headcount and a ratio. Neither is a result yet, and both firms will report against their own definitions. For a founder-led firm the lesson is cheap to apply: a target that is net of cost and tied to a named line can be checked in a year; a gross “productivity” figure cannot.
Do this:
- Rewrite your main AI business case net of what the tools cost to build and run, per month.
- Tie each target to one line and one date, the way Chubb ties its target to a headcount and a ratio.
3. Few firms have a rule for stopping a pilot
The number: 31% of CIOs have guidelines for when an AI pilot should not move to production, and 17% for stopping low-value AI already live (Gartner, 2,002 respondents, read from a published copy of the release).
What we saw: in the same survey, the CIOs Gartner calls the vanguard (better business outcomes and high perceived AI value) are 2.3 and 2.6 times as likely to have these rules, and Gartner puts about one in five AI use cases down as successful. DBS shows how hard the comparison is even for a careful bank: its chief executive said fee income grew 19% in corporate accounts where its banker AI agent was used, against 11% elsewhere, with no control group or selection method described (company-reported, primary transcript).
Volume is not value.
So what: a pilot without a stop rule tends to continue by default, which is how a firm ends up with many pilots and few results. The firms Gartner rates highest are the ones that decided in advance what failure looks like. The DBS figure is the other half of the problem. A 19% against 11% gap is promising, but if the agent went first to the accounts the bank expected to grow, the gap measures that choice as well as the AI. A stop rule is only as good as the comparison behind it, so the two have to be designed together, before the pilot starts.
Do this:
- For every live pilot, write one sentence: “We stop if, by this date, this line has not moved by this much.”
- Decide how accounts, teams or sites are chosen for the pilot before it starts, and keep a comparison group that does not get it.
The surprise of the week
You would expect: a widely quoted Vietnamese bank result, “SME credit from 10 days to 2 days”, to be an AI result across small-business lending.
We found: the bank’s own statement, as reported by Tạp chí Kinh tế Tài chính, covers “some” corporate loan files on its business platform going from about 10 days to 2 days, on MB’s internal data. The article credits data and AI in its lead, and digitisation and automation in the body, and MB says the figures do not yet cover all processes or customer groups.
What it means: the claim is not false; it is narrower. Before a figure goes into a board paper or a classroom, ask three things: which files, credited to what, and compared with when. The same goes for the other Vietnamese bank figures this week, such as TPBank’s roughly 50% faster software processing and two to three times the productivity, which come with no baseline, period or method.
Your next 30 days
If you run a company:
- List your live AI pilots and write next to each the line it should move and today’s value.
- Restate the largest business case net of the monthly cost of the tools.
- Add a one-line stop rule to every pilot: the result, the date and who decides.
- For the next pilot, choose the comparison group before you choose the pilot group.
If you run an executive programme: your participants will sit in rooms where AI is named and rarely priced. Gartner finds only 31% of CIOs have a rule for when a pilot should not go to production.
- Have teams write the stop rule and the comparison group for their own initiative, and defend it to a panel.
- Use a real results release as a case: find where AI is named and whether any figure says where it lands.
Where Civico fits
- For CEOs and COOs: the AI Transformation Studio is a 12 to 16 week programme in which your leaders build their own AI initiatives on their own work, and a Pitch Day ends it, where your leadership funds the strongest initiatives for a first stage or stops them. Companies start with a half-day scoping workshop. Tell us the initiative.
- For programme directors: Vietnam Speed is a four-day AI venture residential in which EMBA and executive cohorts build, test and pitch a real AI venture or initiative, with founders, operators and investors in the room. Talk to us about an edition.
- For anyone checking the numbers: every figure above links to its source, and the rows are in the Use cases view.
Method and limits
This week 18 rows and 4 funding rounds were added and 16 existing rows were rebuilt on primary sources (13 of them changed a date, figure, base or scope). Company figures are company statements, not audited AI effects, and AXA and Chubb figures are targets. Second-hand in this note: the Intact and Chubb figures (summaries or trade-press reports of the call and investor presentation), the Gartner figures (a published copy of the release; gartner.com did not load) and the Vietnamese bank figures (translated by us). The DBS figure is from the bank’s own transcript. Five insurers is a small, large-firm sample.
Corrected 22 November 2026: checked against the companies’ own releases and AXA’s investor-day deck. AXA’s target is footnoted “net of investments” (we had written “net of implementation costs and the extra running cost of the tools”, from a press report); Allstate’s release gives its underlying combined ratio (80.5 against 78.7), not the 17.6% expense-ratio figure we had from a summary; Hippo’s expense ratio fell 10.3 points. The Ping An figures held, now from the company’s own release. The findings do not change.