What to ask before approving an AI budget
Most AI budgets get approved without a baseline, an owner or an exit price. Six questions separate a pilot with a decision attached from a purchase nobody will be able to evaluate a year later.
Contents
Before approving AI spend, ask six things. Which process it replaces, what that process costs today, whether this is a pilot or a purchase. Then who is accountable when the output is wrong, what leaving costs, and how the result gets measured. Most failed AI budgets fail on the second question.
Key takeaways
- Adoption is near-universal and financial impact is not. McKinsey's 2025 survey found 88% of organizations using AI in at least one function, and only 39% attributing any EBIT impact to it.
- Gartner predicted in July 2024 that at least 30% of generative AI projects would be abandoned after proof of concept by the end of 2025.
- MIT NANDA's 2025 study reported around 95% of organizations getting no measurable return, on a base of 52 interviews and 153 survey responses.
- Buying beat building in that same study: externally sourced deployments succeeded roughly twice as often as internal builds.
- Unit costs fall fast. Stanford's AI Index recorded a more than 280-fold drop in inference cost for a GPT-3.5-level system between November 2022 and October 2024, which is an argument against long lock-ins.
Which process does this replace
Every AI proposal describes a capability. Very few name the process that stops. That gap is where budgets go to die, because a capability without a displaced process becomes an addition to the cost base rather than a substitution.
The question has a strict form. Name the process, the team that runs it, the volume it handles per month, and the part of it the system will absorb. If the answer is “it will help the team be more productive”, the proposal is not ready for a decision. Help is not a unit of account.
There is a second failure inside the same question. A process that is absorbed but not retired still costs what it cost. Approving the spend means committing to the operational change that follows, and that commitment belongs to the same executive who signs the budget.
What is the baseline, and who measured it
A baseline is the current cost, cycle time and error rate of the process being replaced, measured before anything is bought. Without it, the pilot cannot be evaluated. With it, most pilots become much easier to approve or reject on the numbers.
Baselines are unpopular because they are slow and they occasionally embarrass the sponsor. They are also the only defense against the pattern the published evidence keeps describing. McKinsey's 2025 global survey covered 1,993 respondents across 105 nations. It found 88% of organizations reporting regular AI use in at least one function, and 39% attributing any EBIT impact to it.
McKinsey's 2025 State of AI survey found 88% of organizations using AI in at least one business function. Only 39% attributed any enterprise-level EBIT impact to it (McKinsey, 5 November 2025, n = 1,993).
Adoption and impact have decoupled. The organizations that can tell the difference are the ones that wrote down what the process cost before they changed it. If the sponsor cannot produce a baseline in two weeks, the honest reading is that nobody owns the process today, and that is the problem to fix first.
Is this a pilot or a purchase
A pilot has a fixed budget, a fixed end date, a pre-agreed success threshold and a named decision at the end. A purchase has a contract, a renewal and an owner. Trouble starts when a purchase is presented as a pilot, because a pilot that cannot fail is a procurement decision wearing a lab coat.
Gartner predicted in July 2024 that at least 30% of generative AI projects would be abandoned after proof of concept by the end of 2025. It pointed at poor data quality, inadequate risk controls, escalating costs and unclear business value. The same release put deployment approaches in a $5–20 million USD range, depending on scope.
In July 2024, Gartner predicted that at least 30% of generative AI projects would be abandoned after proof of concept by the end of 2025. The cited causes were poor data quality, inadequate risk controls and escalating costs (Gartner, 29 July 2024).
Abandonment is not the failure. Abandonment without a written decision is. A pilot that ends with “we learned a lot” and no recorded conclusion costs the same as one that ends with a documented no. It leaves none of the memory.
Who owns it when the output is wrong
Every AI system has a failure mode, and someone has to be accountable for it before the first output reaches a customer. The question is not who built it. It is who answers for the wrong invoice, the mis-sent message, the record changed by mistake.
That accountability has to sit with the process owner, not with the vendor and not with IT. Write it into the approval: named person, defined scope of authority, and a documented rollback path. The moment a system is allowed to write to a CRM without that structure, the failure mode becomes everyone's and therefore nobody's.
System-level effects matter too. DORA's 2024 research associated a 25% increase in AI adoption with better documentation and code quality. The same increase came with a 1.5% decrease in delivery throughput and a 7.2% decrease in delivery stability. Individual output and system reliability move independently.
Regulatory exposure belongs in the same paragraph. For anything touching high-risk categories in the EU, the dates are already fixed in the AI Act calendar. Compliance is cheaper to plan than to retrofit.
What does it cost to leave
Exit cost is the least discussed number in an AI approval and often the largest. It has four parts. Data you cannot take with you, integrations you would have to rebuild, retraining for the people who adapted to the tool, and process knowledge that stopped being documented.
Ask for the exit cost as a figure and a timeline, in the same document as the purchase price. A vendor who cannot answer has told you something useful. A sponsor who has not asked is not yet ready to sign.
Falling unit costs make this sharper. Stanford's AI Index reported that inference cost for a system performing at the level of GPT-3.5 fell more than 280-fold between November 2022 and October 2024. Anything priced on today's compute economics will look expensive within a year, which argues for short terms and portable data rather than multi-year commitments.
Stanford's 2025 AI Index recorded a more than 280-fold fall in inference cost for a GPT-3.5-level system between November 2022 and October 2024 (Stanford HAI, 7 April 2025). Long lock-ins are priced against a falling curve.
Build, buy or wrap
The build-versus-buy question is usually argued on control and cost. The published evidence points somewhere less flattering: the ability to finish. MIT NANDA's 2025 study of enterprise generative AI reported that externally sourced deployments reached production roughly 67% of the time, against roughly 33% for internal builds.
That study is worth reading with its method visible. It rests on 52 structured interviews, 153 survey responses from senior leaders, and analysis of more than 300 publicly disclosed initiatives. It reported about 95% of organizations seeing no measurable return on an estimated $30–40 billion USD of enterprise investment. The sample is small for the size of the claim. The direction is consistent with what the larger surveys show.
MIT NANDA's 2025 report found around 95% of surveyed organizations had seen no measurable return from generative AI. Externally sourced deployments succeeded about twice as often as internal builds (MIT NANDA, July 2025).
The practical conclusion is not “always buy”. It is that internal builds belong to the few cases where the process is genuinely proprietary. Everything else should be bought or wrapped, which is the same build, buy or wrap decision applied with an honest view of delivery capacity. Retrieval quality, evaluation and running cost decide whether the wrapped version holds up, and those are production concerns rather than demo concerns.
The six questions as a single page
An approval sheet that fits on one page beats a business case nobody rereads. It carries six lines and a signature.
Which process this replaces, with volume and owner. What that process costs today, with the date the baseline was measured. Whether this is a pilot or a purchase, with the end date and the success threshold. Who is accountable when the output is wrong. What it costs to leave, in money and weeks. How the result will be reported, to whom, and how often.
Two of those lines are usually blank on first pass. That is the useful part of the exercise. A blank line is a question the organization has not answered yet, and finding it before the money moves is worth more than the approval itself. The same discipline applies to how a department organizes its own delivery model. Decisions are cheap to record and expensive to reconstruct.
FAQ
What is the most common reason AI budgets fail?
No baseline. Without a measured cost, cycle time and error rate for the process being replaced, nothing can be evaluated afterwards. McKinsey's 2025 survey found 88% of organizations using AI but only 39% attributing any EBIT impact to it, which is the shape of a portfolio without baselines.
How do I tell a pilot from a purchase?
A pilot has a fixed budget, a fixed end date, a pre-agreed success threshold and a named decision-maker at the end. A purchase has a contract, a renewal and an owner. If a proposal called a pilot has no defined way to fail, it is a purchase and should be approved as one.
Should we build our own AI system or buy one?
Buy or wrap, unless the process is genuinely proprietary. MIT NANDA's 2025 study reported that externally sourced deployments reached production roughly twice as often as internal builds. Internal capability still matters for evaluation and integration, but building the model layer rarely explains the difference between success and abandonment.
What should exit cost include?
Four items: data you cannot export, integrations you would have to rebuild, retraining for people who adapted to the tool, and process knowledge that stopped being documented. Ask for it as a figure and a timeline in the same document as the price. Vendors who cannot answer have answered.
Does using AI improve delivery performance automatically?
No. DORA's 2024 research associated a 25% increase in AI adoption with better documentation and code quality. It also found a 1.5% decrease in delivery throughput and a 7.2% decrease in delivery stability. Individual productivity and system reliability need separate measurement.
Who should own an AI system inside the company?
The owner of the process it changes, not IT and not the vendor. Ownership means authority to stop the system, a documented rollback path and responsibility for the outputs. Where the system writes to production records, that ownership has to be named in the approval before deployment, not after.
Take the six questions in order, and refuse to move past the baseline until someone produces it. Approve pilots with end dates and thresholds, buy rather than build unless the process is proprietary, and keep terms short while unit costs keep falling. In six months it will be clearer whether the organizations reporting EBIT impact differ in technology or only in measurement discipline. Current evidence points at the second, the cheaper problem to fix.
