Key takeaways
“A budget is not a forecast of what will happen. It is a statement of what the organisation has decided to find out.”
Somewhere in your organisation, an AI budget is about to be approved. It will have a confident headline saving, a vendor quote and a timeline that fits neatly on one slide. What it probably will not have is the cost of data preparation, the cost of running the system after launch, or a way to tell the board later whether it worked. A CFO who signs that document is not being careless; they are being asked to sign half a business case. This guide is for finance and technology leaders who want to put the other half in front of the CFO before the signature, not after the overrun.


Why AI Budgets Get Approved on Incomplete Numbers
Finance teams are good at testing numbers they are shown. The trouble with AI proposals is what they leave out. A typical request arrives with a vendor price, an estimate of hours saved, and a multiplier that turns those hours into money. Each figure looks reasonable on its own. Together they describe the cheapest, fastest and most favourable version of the project, because the people writing the proposal are the people who want it approved.
There is also a structural reason. AI spend rarely sits in one place. The licence is in IT, the data work is in an analytics team, the pilot is in a business unit, and the cloud bill is somewhere else entirely. No single owner sees the whole cost, so no single approval covers it. By the time a CFO is asked to sign, they are often signing one slice of a larger commitment without being told so.
What a CFO is usually shown, and what is missing
- Shown — the vendor quote. a licence or build fee, often with an assumption about usage volume.
- Shown — the benefit estimate. hours saved multiplied by a loaded salary rate, usually before anyone has tested whether those hours convert to cash.
- Missing — the baseline. a measured picture of how the process performs today, which is the only thing a future saving can be compared with.
- Missing — the run cost. what the system costs per month once it is live, and how that grows with adoption.
- Missing — the owner. a named person accountable for the benefit actually arriving.
None of this means AI is a poor investment. It means the approval process is being applied to a kind of spend it was not designed for, and the gaps are predictable enough to close in advance.
Why the Usual Business Case Template Fails for AI
Most capital templates assume a project with a defined scope, a delivery date and a steady-state benefit. AI initiatives behave differently in four ways that matter to a finance reader.
- The cost is not one-off. models need monitoring, evaluation, updating and, in many designs, usage-based fees. A system that is cheap to launch can be expensive to keep.
- The benefit is probabilistic. an AI system is right most of the time, not all of the time. The saving depends on how often it is wrong and what each error costs to catch and fix.
- The value depends on data you may not yet have. the quality of inputs decides the quality of outputs, and cleaning inputs is often the largest piece of work.
- The scope moves. successful pilots attract new use cases, and each one brings its own integration and governance needs.
Templates also reward optimism. A single-point estimate of payback invites a single-point answer, and a project that pays back quickly and one that pays back very slowly can look identical on the slide. A CFO needs a range, the assumptions behind each end of it, and the signals that would show which end the project is heading towards.
Common fixes do not fully help. Adding a contingency percentage to the budget hides uncertainty rather than describing it. Asking the vendor to guarantee savings usually results in guarantees that are conditional on factors the buyer controls. Running a pilot first is sensible, but a pilot with no agreed success criteria simply postpones the same argument.
The Costs a CFO Will Not See in the Vendor Quote
The most useful thing a finance leader can do before approval is ask what is not in the number. In our experience, the omissions fall into a small number of categories, and each can be estimated before work begins.


- Data preparation. extracting, cleaning, labelling and connecting the data an AI system needs. In many programmes this takes longer than building the model or agent itself.
- Integration. connecting the system to the CRM, ERP, ticketing or finance tools where work actually happens. A tool that cannot write back to those systems creates a new manual step instead of removing one.
- Run costs. inference or usage fees, cloud infrastructure, monitoring and support. These recur every month and tend to rise as more teams adopt the tool.
- Evaluation and quality control. someone must check that outputs stay accurate as data and models change. That is skilled, ongoing work.
- Governance and security. access control, audit trails, data handling reviews and policy work, especially where the use case touches customer or regulated data.
- Change management. training, process redesign and the productivity dip while people adapt. Savings only arrive if behaviour changes.
Governance deserves particular attention because it is the cost least likely to appear and the most expensive to add later. If leadership cannot see which AI uses exist, who is submitting what and what the controls decided, the organisation is carrying risk that no budget line acknowledges. Visibility is a design requirement, not an afterthought. As an adjacent example, an executive dashboard built to make prompt governance visible to leadership shows what that looks like when it is part of the scope from the start.
How to Stress-Test an AI Business Case Before Approval
A CFO does not need to understand model architecture to test an AI case. They need a short list of questions that expose whether the numbers have been built on evidence or on enthusiasm. The ones below are quick to ask and can change the decision.
Questions that expose weak cases
- What is the measured baseline?. Ask for the current cost, volume, cycle time and error rate of the process, from real data, not recollection. If none exists, building it becomes the first deliverable.
- What happens to the saved time?. If a team saves a few hours a week, does headcount change, does the backlog shrink, or does the time disappear into other work? Each answer has a different financial value.
- What does a wrong answer cost?. Estimate the frequency of errors and the cost of catching them. A system that is right most of the time can still be uneconomic if the exceptions are expensive.
- What is the monthly run cost at several times today's volume?. This tests whether the model scales economically or only works at pilot size.
- What would make us stop?. Agree in advance the metrics and dates at which the programme is paused, redirected or ended. A project with no exit criteria has no real approval.
- Who owns the benefit?. Name the business leader whose targets include the saving, not just the technology lead who delivers the system.
Then run the case at the pessimistic end. Reduce the benefit, increase the run cost, delay adoption by a couple of quarters, and see whether the investment still makes sense. If it does, you have a robust case. If it only works under the central assumptions, you have learned something valuable before the money is spent rather than after.
A Business Case Framework a CFO Can Defend
The aim is not to make approval harder. It is to make the decision defensible to the board, the audit committee and the CFO's own future self. A sound AI business case has a consistent shape, and it works best when funding is released in stages tied to evidence.
- Start with a measured baseline. document the current process cost and performance before any build begins.
- Value the benefit in ranges. state a low, expected and high case, and separate cash savings from capacity released and risk reduced.
- Cost the whole lifecycle. include data, integration, run, evaluation, governance and change, not just build or licence.
- Stage the funding. release budget in tranches, for example discovery, a bounded pilot, production rollout, each unlocked by agreed results.
- Build visibility in from the start. define the dashboard or report that will tell leadership how the system is performing and being used, and treat it as part of the scope.
- Review on a fixed rhythm. compare actual run cost and benefit with the case every quarter and publish the variance.
This is the discipline Prospiretech's AI strategy and advisory work is built around: tying each use case to a measurable outcome before engineering starts. Where the case then proceeds to build, the same thinking carries into intelligent workflow automation and autonomous AI agents that are designed to integrate with the systems your finance and operations teams already use. For production systems, MLOps and model engineering supplies the monitoring that makes ongoing cost and quality visible.
An Enterprise-Grade Dashboard That Makes Prompt Governance Visible to Leadership
A governance cost is hard to justify when nobody can see what it is governing. An enterprise SaaS company in the USA, which builds AI governance and prompt management tools for large customers in regulated industries, had a working rules-based governance engine but no interface for the compliance leads, product managers and executives who needed to act on it. Prospiretech designed and built the customer-facing platform: an executive dashboard showing prompt submission status, validation results and governance metrics, a structured submission flow for non-technical users, and a results view that translates engine output into clear pass, fail or escalate decisions. It connects to the client's engine through REST APIs, runs each tenant in full isolation and was built to be Okta SSO/SAML-ready. For a CFO, it shows what visibility looks like in practice.
- Leadership can see the state of prompt governance across the organisation for the first time, without touching the engine
- Deployed across multiple enterprise tenants without changes to the underlying engine
- Delivered Okta-ready, with no additional development needed for identity provider integration
- Non-technical users can create, submit and track prompts with no training required
Frequently Asked Questions
Why do AI business cases tend to understate cost?
Cost is spread across teams and budgets, so nobody sees the total. Proposals usually quote the licence or build fee and omit data preparation, integration, monitoring, security review and change management. Run costs also grow as adoption spreads. The remedy is to cost the full lifecycle, name an owner for each line and review actuals against the case every quarter rather than only at approval. Revisiting the case at each stage gate, with real data replacing assumptions, is the most reliable way to keep cost and benefit honest as scope grows.
How should a CFO value time saved by AI?
Treat hours saved as capacity, not cash, until a decision converts them. Ask whether the time reduces headcount, avoids a hire, clears a backlog or improves service levels, then value that outcome. Multiplying hours by a salary rate overstates benefit when the time is absorbed elsewhere. Present cash savings, released capacity and risk reduction as separate lines so the board can see what each assumption is worth. Whatever the method, document the assumptions beside the number so that a later reader can see what the benefit depended on and challenge it with evidence.
Is a pilot enough to justify full funding?
Only if the pilot was designed to answer the financial questions. A pilot on clean, prepared data with an enthusiastic team shows what is possible, not what is typical. Define success criteria first, test on representative data, record real effort and error rates, and measure run cost. Then use those results to set the range for the production case instead of extrapolating from the best week. Seeing actual spend against the case at least quarterly also prevents the common drift in which small approvals add up to a commitment nobody agreed.
What governance should sit alongside an AI budget?
At minimum, an inventory of AI uses, clear ownership, access controls, an audit trail of inputs and outputs, and a regular report to leadership on usage, quality and cost. The depth should match risk: a tool touching customer or regulated data needs more scrutiny than one drafting internal notes. Including governance in the original scope is cheaper than adding it after an incident or audit question. Agree in writing, before work starts, which results trigger the next tranche of funding and which would end the programme, so the decision is not reopened emotionally later.
How can finance and technology teams work together on the case?
Build it jointly from the start. Technology supplies the architecture, data and run-cost estimates; finance supplies the baseline method, valuation rules and approval gates; the business owner commits to the benefit. A shared template with ranges, assumptions and stop criteria keeps everyone honest, and a quarterly review keeps the numbers live. Outside advisers can help set the structure, but ownership should stay inside. Whichever structure you choose, schedule the first review before the contract is signed, so the habit of checking numbers against reality is established from day one.
Give your CFO the whole number
A budget that includes data, run cost, governance and exit criteria is easier to approve and far easier to defend. Prospiretech will help you build an AI business case your finance team can stand behind, starting with a complimentary consultation.
