The question this answers: if your AI pilot worked, which line in the P&L would show it, and who would answer for it?
Bottom line: the firms that show AI in their results do three things: they name a cost or revenue line first, they rebuild a whole process rather than add a tool, and they put a named person on the number. Most firms do none of the three yet.
This week’s three findings
1. The firms with a number name a cost line, not a tool
The number: 30% lower technology vendor spend at Citi, year on year.
What we saw: on its third-quarter call Citi also reported client response times 21% faster, with 14,000 engineers on an agentic coding tool. Santander has set more than €1bn of AI revenue and savings by 2028, with €35m realised in the first quarter of 2026. The IT-services firms now report AI as revenue: TCS $3.1bn a year, over 10% of the total; HCLTech about $876m a year, up 78%, while its total revenue grew 2.6%. All company-reported.
So what: a finance team can check vendor spend or a revenue line. It cannot check “productivity”. The firms with a number chose a line their CFO already tracks. look at what the four numbers have in common. None of them is “hours saved” or “employees using the tool”. Each is a line that existed in the accounts before AI arrived: vendor spend, revenue, a savings target with a date. That is why they can be reported at all. The firms that gave no AI figure may well have results; they just never tied the work to a line, so there is nothing to show. For a CEO the lesson is about sequence: the line comes first, the tool second. A pilot that starts with the tool will end with a usage chart.
The line comes first, the tool second. A pilot that starts with the tool ends with a usage chart.
Do this:
- Pick the one line in your management accounts that your AI project should move, and write down today’s value before the project goes further.
- Ask your finance lead to sign off that baseline, so the result is theirs as well as the project team’s.
2. Use is broad. Production is rare.
The number: 87% of 836 AI use cases at EU securities firms are internal only, and 45% of firms have none in production or development (survey of about 730 firms by the EU securities regulator).
What we saw: the European Banking Authority finds 92% of EU banks deploying AI. Yet Wells Fargo does not attribute its 8% fall in staff to AI, and Wipro gave no AI revenue figure. Across US firms using AI, only 15% trained their staff for it and 64% changed nothing about how the business runs (Census survey, reported second-hand). Citi, by contrast, talks about about 100 of its most manual processes re-engineered end to end.
So what: licences and pilots do not change how the business runs. A rebuilt process does, and that is where the number comes from. the gap between “deploying AI” and “changed how the business runs” is where most AI budgets sit today. Internal use cases (drafting, search, summaries) are cheap to start and easy to count, but they spread time savings thinly across many people, and thin savings rarely reach a cost line. Citi’s figure comes from the other end: about 100 whole processes rebuilt. The practical test for any pilot is whether a process map changes, with fewer steps, hand-offs or waiting time. If the map stays the same, expect the P&L to stay the same too.
Do this:
- Choose one process end to end (for example quote to cash, onboarding or claims), not one team’s use of a tool.
- Count its steps, hand-offs and waiting time now, and again after the change.
3. Accountability is moving to named people
The number: UK MPs want guidance on senior managers’ accountability for AI by the end of 2026; Singapore’s new rules make boards and senior management oversee AI risk from October 2027.
What we saw: the UK financial regulator’s own review found the senior-manager regime holds where AI supports human decisions but is challenged as AI acts more on its own. Santander’s targets are set out by a named chief data and AI officer.
So what: regulators are starting with banks, but the question will reach every board: who owns the AI number? accountability is the third ingredient, and it is the one the rules now force. Once a named senior manager answers for an AI system, that person will ask for the baseline and the process change before signing, which is exactly what the firms with a number already do. Founder-led firms face no such rule yet, which is an advantage only if they use it: they can name an owner this month without waiting for a regulator to make them.
Whose name is on the number? If the answer is "the project", nobody's is.
Do this:
- Name one budget owner per AI project this month, by role, not “the project team”.
- Put each AI project on the board pack once a quarter with one of three calls: funded, parked or stopped.
The surprise of the week
You would expect: funds that market themselves as investing with AI to beat the rest.
We found: 145 EU funds that say they use AI, holding about €13bn, returned 1.11% against 0.70% for other equity funds, a difference that is not statistically significant, and mostly lost money to outflows over three years.
What it means: “AI-powered” in a pitch is a claim, not a result. When a vendor or a team says it, ask what it was compared with.
Your next 30 days
If you run a company:
- List your AI projects and write next to each the P&L line it should move. Stop or park any that have none.
- For the project that remains, record today’s baseline and get finance to sign it.
- Name the budget owner and the process that will be rebuilt, not just assisted.
- Book the decision: a date when the sponsor will say funded, parked or stopped.
If you run an executive programme: candidates are asking for this. In GMAC’s 2026 survey of 4,253 prospective students, about two-thirds want AI taught through business simulations and practical application, and prompt engineering was among the least wanted. Most schools are still adding AI modules to existing courses (65% of deans).
- Have each participant arrive with a live problem from their own firm and its baseline number.
- Judge the module on the decisions participants take home, not on completion.
Where Civico fits
- For CEOs and COOs: the Civico AI Academy is an eight-week programme in which your leadership team takes one live AI project to a board-ready case: the work and decisions are redesigned around it, and your sponsors decide funded, parked or stopped at the end. Companies start with a half-day scoping workshop. Tell us the project.
- For programme directors: Vietnam Speed is a four-day venture residential in which EMBA and executive cohorts build, test and pitch a real venture or AI initiative, judged on the decisions they take home. 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 (filter by financial services).
Method and limits
The bank and IT-services figures are company statements from results calls and releases, not audited AI effects. Two figures are second-hand reports (the Census and GMAC figures) and say so. We hold back a JPMorgan row until we can confirm a figure in its call transcript. This is mostly large-bank evidence; founder-led firms rarely publish such numbers, which is why the questions above matter more for them.