The business case leadership signs

Why AI budgets get questioned

AI spending is rising fast. BCG’s 2026 index found that corporate AI spending roughly doubled in a year, from about 1.7% of revenue in late 2025 to 3.3%. At the same time, McKinsey’s 2026 survey found that only 37% of companies attribute any EBIT impact to AI.

That gap is why boards and finance teams now ask harder questions. The answer is a business case built like any other capital decision, on numbers finance recognises. BCG’s leaders do exactly that: they measure AI’s impact directly in their financial results.

One page, five blocks

The business case for an agent fits on one page, in five blocks.

1. The problem in money. The process, the P&L line it touches, and the baseline figure agreed with finance.
2. The value. The change measured in the proof, scaled to full volume, as a low, expected and high case.
3. The cost to build. The one-off cost to take the agent from proof to full use: engineering, integration, testing, training and change.
4. The cost to run. The monthly cost at full volume: model usage, infrastructure, monitoring, support and the people time to review the agent’s work.
5. Payback and risk. Months to payback in each case, the main risks, and the controls that cover them.

Everything else goes in an appendix.

Sizing the value

Value comes from four places. Most agents touch one or two.

• Cost out. Agency spend, overtime, outsourcing, rework, write-offs.
• Capacity freed. Hours moved from routine work to higher-value work. This only counts as value if the business says what those hours will be used for.
• Revenue. Faster quotes, more proposals, better conversion, better pricing.
• Risk reduced. Fewer errors, fines, disputes or non-compliant documents.

The hotel group in Chapter 1 shows the simplest kind of case. The group spent more than €1 million a year on agency staff. After the optimiser went live, agency spend fell by 70%. On a spend of that size, the saving is worth roughly €700,000 a year, from a project that took four months from kickoff to live.

Revenue cases need more care. The European RegTech platform from Chapter 4 cut the time per regulatory inquiry by 75%, from three to four hours to under an hour. The start-up’s revenue grew about three times in the year after launch. A business case should not claim the whole of that growth for the agents, since sales, market and product all played a part. It can claim the capacity freed and the faster response to customers, and note the revenue as supporting evidence.

Counting the cost

The cost to build is usually well understood. The cost to run is where business cases go wrong.

Model prices have fallen steeply. Stanford’s 2025 AI Index found that the cost of running a model at a fixed level of capability fell more than 280 times between late 2022 and late 2024. In most of Twistag’s engagements, model usage is now a small part of the monthly cost. People time to review and the engineering to keep agents current are usually larger.

Low, expected and high

A single number invites argument. Three cases invite discussion.

• Low case. The agent handles less volume than planned, or people review more than expected. What is the payback then?
• Expected case. The proof’s numbers, scaled to full volume.
• High case. The agent moves to approve-by-exception or supervise mode, as described in Chapter 2, and review time falls.

Next to each case, write the two or three assumptions that drive it. Leadership can then argue about the assumptions, which is the useful conversation, and the case can be updated as the numbers come in.

Payback and timing

A well-chosen first agent should pay back within a year. Value often starts slowly, as people learn to trust the agent and the review share falls. The case should show the monthly curve: build cost up front, run cost from launch, and value rising over the first months.

A case that only pays back in year three is a signal to look again. Either the use case is too small, the build is too large, or the value has not been sized properly.

Getting it signed

Three things help a case get approved:

• Finance in the room early. The baseline and the value logic agreed with finance before the case is written.
• A named owner. The executive whose budget or target moves signs the case alongside finance.
• Stage gates. Funding released in stages, with a decision at each one against the agreed numbers. This turns a large approval into a series of smaller ones.

Chapter 10 covers how to report the same numbers every quarter once the agent is live.

Common mistakes

• Counting hours as savings. Hours freed only become value when the business decides what happens to them.
• Ignoring the cost to run. A case with only a build cost looks great until the first monthly bill.
• Claiming the whole outcome. Revenue growth with many causes credited entirely to an agent.
• One number. No range, no assumptions, and no way to update the case as results arrive.
• No baseline from finance. The before number comes from the project team, and finance does not accept it.

What to do this month

1. For the first agent, write the one-page case in five blocks, using the proof’s numbers.
2. Agree the baseline and the value logic with finance before presenting it.
3. Build the low, expected and high cases, with assumptions written next to each.
4. Propose stage gates for funding, each with a date and a decision rule.

Questions for the leadership team

1. Which P&L line does each proposed agent move, and who in finance has agreed the baseline?
2. What will it cost each month to run the agent at full volume, including the people who review its work?
3. What has to be true for the low case to happen, and how likely is it?
4. How soon will the first value show up, and when does the investment pay back?
5. What would make the company stop or scale the agent, and who decides?