Deloitte asked 3,235 executives across 24 countries, between August and September 2025, how many of their AI pilots had reached production. Only 25% had moved at least 40% of their pilots into production. McKinsey's August 2026 edition of its annual survey found that 37% of companies attribute some EBIT impact to AI, and that just 6% attribute a meaningful one, the same share as the year before.
The money keeps flowing regardless. Bain surveyed more than a hundred CFOs in April 2026: 83% plan to raise AI spending by more than 15% within two years, while only 31% are satisfied with the results so far. Rising budgets on top of flat returns mean one thing for any executive committee: this is the year AI stops being approved on faith.
The pilot has already proved what it had to prove
A pilot answers a technical question: does this work with our data, our processes and our people? Once the answer is yes, the pilot's job is done and what follows belongs to a different genre. A business case answers a financial question: does this deserve the money more than everything else that same money could do this year?
Most pilots that end up in a drawer had passed the first question; they stalled because nobody turned their result into a fundable decision with a number, an owner and a date. IBM's May 2025 study of 2,000 CEOs puts it bluntly: 64% admit to investing in technologies before they understand the value they bring, for fear of falling behind. A board that approves on that basis has given up its function.
The 2026 backdrop makes the conversion more urgent. January's trend lists agreed on two things: AI now ships inside the ERP, the CRM and the customer-service platform, and process automation is evolving into what the industry calls augmented automation, with bots, machine learning and natural language in the same flow. When AI comes bundled with software you already pay for, "having AI" stops being an achievement. As the innovation blog of Spanish insurer Santalucía put it in January: the question is no longer how much gets adopted, but what actually gets transformed.
What a board can approve, and what it cannot
What a board approves are concrete decisions that compete with each other for the same capital: hiring two salespeople, opening a market, renewing a production line, paying down debt. "Artificial intelligence" appears nowhere on that list, any more than "electricity" does. The business case exists to make the comparison possible, which is why it has to speak the language of comparison: the P&L and the balance sheet.
Framing matters more than it looks. Gartner surveyed more than two hundred finance executives in July 2026: 45% aim their AI investment at productivity and only 20% at better decision quality. Yet just 17% of the productivity camp report significant value, against 31% of those who invested in deciding better. The case that sells most easily in the committee room is the one that reports the least return afterwards.
What separates the companies that do capture value? McKinsey finds the same thing edition after edition: close to three quarters of high performers have fundamentally redesigned their workflows rather than layering the tool on top of the existing process. A serious business case books that redesign as a cost and as a condition, because without it the value line is paper.
"A pilot proves the technology works. A business case proves it deserves the money over everything else that money could do. A board can only sign the second."
The five value lines and their admission rules
Every line in an AI business case has a rule that decides whether a euro gets in. Without that rule the document inflates the first line, forgets the next two and ends up defending a payback nobody believes.
1. Savings. Only what disappears from a ledger entry counts: a contract cancelled, a vacancy left unfilled, a supplier invoice that drops, inventory that shrinks. "Hours saved" multiplied by an hourly rate turn into savings only once someone decides what those hours will do. A CFO spots that move in seconds, and rightly so.
2. Productivity. This is the most common line and also the most fragile. It gets in only if the freed hours have a written destination (more volume with the same headcount, a shorter cycle, new work that previously did not fit) and if at least one outcome KPI is committed to move: orders served per person, cycle days, error rate. The full rule is in the people plan that decides your ROI.
3. Revenue. This is the most wanted line and the most abused, as Deloitte measured in January 2026: 74% of companies hope to grow revenue through AI and only 20% are doing so. It enters the case only with an explicit mechanism (conversion, average ticket, retention, time to quote) and a way of measuring it that survives seasonality: a control group, a region or a segment where the system stays switched off.
4. Risk avoided. Valued as probability times impact, from a source that can be audited: your own incidents over the last three years, audit findings, insurance premiums, fines in your sector. Payment fraud, billing errors, knowledge walking out of the door with a key person, and regulatory compliance all belong here; the EU AI Act calendar, reshaped by the Omnibus that entered into force in July 2026, now places high-risk obligations in December 2027, and NIS2 is already in force. The admission rule is that the risk is quantified with your own data and that nothing already booked as savings is counted twice.
5. Payback. Months until the full project cost is recovered through the four lines above, with explicit assumptions and one sensitivity: what happens if adoption is half of what was planned. As a committee rule of thumb, a process automation that fails to return the money within 12 to 18 months needs extra justification; a data or integration foundation that serves several cases can accept 24 to 36 months, provided the second case is already identified.
The full three-year cost
The second half of the business case is the one that almost never gets written in full. KPMG's quarterly pulse of June 2026, covering 2,145 executives in twenty markets, found that only 7% achieve measurable returns, that 42% have only partial visibility of what they spend on AI, and that those who do control the cost are five times more likely to achieve ROI (15% versus 3%). Half of the return lies in knowing what it costs.
A full three-year cost has at least seven items: the licence or subscription; model consumption (tokens, calls, platform), which is variable and grows precisely when the project works; integration with the ERP, the CRM and in-house systems; data preparation; human supervision of what the system produces; training and change management; and security and compliance, including the cost of leaving the vendor if it comes to that. Any return formula that leaves out the consumption item is describing a different project.
One effect deserves to be written down in the committee: usage pricing turns success into a cost. An agent that resolves 60% of queries burns more tokens than one that resolves 20%. If the value line grows linearly and so does the cost line, the project's margin is decided by the slope of each, and that arithmetic has to be done before signing, with the vendor's price list and a volume hypothesis.
Three decisions that deserve the money, and two that do not
The three cases below are archetypes, with rounded figures to illustrate the method. What matters is the shape of the reasoning and how each line passes, or fails, its admission rule.
Accounts payable at a €60 million manufacturer. Four people process 30,000 invoices a year on an ERP that forces them to key everything in. Automating capture, matching to purchase orders and the posting proposal handles 70% without intervention. Admitted savings: 1.5 positions left unfilled after two retirements already scheduled (about €60,000 a year) plus early-payment discounts currently lost to delays (another €25,000). Full first-year cost: €45,000 of ERP integration, €20,000 of licence and consumption, €15,000 of supervision and change. Payback: 10 months. It deserves the money because the saving sits in a ledger entry that disappears and can be checked at the next close.
First-line support at a services company with 40,000 tickets a year. An agent that resolves 55% of repetitive queries and hands the rest to a person with the context already gathered. Admitted savings: the outsourced contact-centre contract, renegotiated downwards (an invoice that changes). Admitted risk avoided: SLA penalties the company has paid in two of the last three years, with a known amount. Revenue: left out of the base case even though response times improve, because the mechanism to retention has never been measured. Payback: 11 months. The stop criterion is the correct-resolution rate falling below 90% for two weeks.
An assistant over the company's own ERP and CRM data for the sales director. This is the case that looks most like a data investment: the customer master has to be cleaned, price lists unified, and two systems that have ignored each other for ten years connected. The quote cycle drops from five days to one, and pricing errors, which today cost margin and returns, halve. Admitted revenue: only the improvement in quote win rate, measured against the region where the system stays off during the first half. Payback: 22 months, and it still deserves the money, because the same clean data foundation carries the next two cases (demand forecasting and pricing), which cannot exist without it.
And two that fail in the form they usually reach the committee. The "corporate AI platform so we are ready", with no first case carrying a value line: pure cost with the hope that someone will use it. And the copilot for the whole workforce justified with a satisfaction survey and a theoretical hours saving: with no destination for those hours and no outcome KPI, it passes nobody's rule. Both are missing the same thing: a business case.
The one page that goes to the board
An AI business case fits on one page, and when it spills over, one of the lines is usually inflated. The page carries eight elements, in this order.
1. The decision. One sentence with a verb: "automate supplier-invoice capture and matching in the ERP". Without the word "explore".
2. The number. Three-year net value and payback in months, with the five lines broken out and the admission rule each one has passed.
3. The assumptions. Three to five, each with the data behind it: annual volume, resolution rate observed in the pilot, consumption price, expected adoption.
4. The full cost. The seven items over three years, with variable consumption tied to the volume of success.
5. The dependency. Who owns the data, what switching model vendor costs, and what happens to the process if the service is down for a week.
6. The risks and their controls. System errors, sensitive data, compliance, and who reviews what before anything goes out to a customer or to the tax authority.
7. The stop criterion. Which number, on which date, forces the project to close, and who has the authority to do it without going back to the board.
8. The owner. A named business person who answers for the number at the 90-day review. Neither the vendor nor the IT department.
A board that receives this page can do its job: compare, condition, approve or reject. A board that receives forty slides with the word "transformation" can only trust, and trusting is no part of a board's job.
Dependency and scalability: the two missing questions
The 2026 conversation turns on three words: return, dependency and scalability. The last two almost never reach the board, and they decide whether the first case is the start of something or an isolated expense.
Dependency. Gartner predicted in June 2025 that more than 40% of agentic AI projects will be cancelled by the end of 2027, through escalating costs, unclear business value or inadequate risk controls, and warned about "agent washing": familiar products (assistants, RPA, chatbots) relabelled as agents. The board's question is simple: if this vendor disappears or triples its price, in how many weeks and at what cost do we keep operating? If the answer depends on data that only exists inside the vendor's platform, the case has a hidden line item.
Scalability. BCG estimated in September 2025 that only 5% of companies are built to scale AI, 35% are beginning to generate value and 60% report minimal gains. The distance between that 5% and that 60% is almost always decided by the second case: whether it costs 20% of the first or 100%. A business case that aspires to be the first of a series has to say how much of its cost (clean data, integrations, controls, team skills) remains available for the next one.
When those two questions are asked before signing, with someone at the table who is selling nothing, the case changes shape and, often, size. A business case of this kind takes two weeks of serious work and avoids the pattern PwC recorded among 4,454 CEOs in January 2026: 56% still see no significant financial benefit from AI and only 12% see it on both the cost and the revenue side. That 12% buys the same models as everyone else and approves them with better cases.
If there is a pilot waiting for budget and a vendor proposal on the table, that is precisely the moment when an independent reading returns more than it costs: a Second Opinion before signing, or a business case closed before anything gets built, which is how our Enterprise AI Implementation starts. And if the job is to put the whole portfolio of initiatives in order, that is what our advisory programs are for.
