Most AI business cases fail for the same reason: they ask leaders to fund a promise rather than evidence. Benefits are estimated from supplier claims, costs stop at the licence fee, and the risks are a single line in an appendix. Finance directors have seen enough of those to be sceptical, and they are right to be.
An investment-ready AI business case is different. It is built on evidence from discovery and pilots, structured in the format public sector decision-makers already use, honest about costs and benefits, explicit about AI-specific risks, and asks for funding in stages. This guide shows how to build one.
Key takeaways
Build the case on evidence: baselines from discovery and results from pilots.
Use the Five Case Model so the case fits existing approval processes.
Cost the whole lifecycle, not just licences: data, integration, change, assurance, running costs and exit.
Present benefits as cashable, non-cashable and enabling, each with a confidence level.
Ask for funding in stages, tied to evidence at each gate.
Why do AI business cases get rejected?
Five weaknesses come up again and again:
Benefits without baselines. Savings are claimed as percentages of a process nobody has measured.
Costs that stop at the licence. Data remediation, integration, change management, assurance and ongoing usage costs are missing.
No evidence it works here. The case relies on supplier case studies rather than a pilot in the organisation's own services.
Risk treated as an afterthought. Data protection, fairness, security and resident trust are mentioned but not assessed.
No owner for the benefits. Nobody in the service is accountable for making the savings real.
Each of these is fixable, and fixing them is mostly a matter of sequence: do the discovery and the pilot first, and the business case almost writes itself.
How should an AI business case be structured?
Use the Five Case Model from HM Treasury's business case guidance, which public bodies already know. Each case answers a question leaders will ask anyway:
Case | The question it answers | What it needs for AI |
Strategic | Why is this needed, and how does it fit our priorities? | The service problem, evidence of need from discovery, links to corporate and financial strategy |
Economic | Is this the best value option? | Options including "do nothing" and "do minimum", pilot evidence, benefits and costs over the lifecycle |
Commercial | Can we buy it sensibly? | Procurement route, contract length, data and exit terms, avoiding lock-in |
Financial | Is it affordable? | Whole-life costs, funding sources, the effect on budgets year by year, including usage-based costs |
Management | Can we deliver it and realise the benefits? | Governance, stage gates, risk management, benefits owners and the plan for live running |
For smaller investments, a shorter version covering the same five questions is usually enough.
How do you estimate the real cost of AI?
Cost the whole lifecycle. In our experience, the licence or usage fee is often a minority of the total. Include:
Cost category | What to include |
Licences and usage | Subscriptions, per-user fees and usage-based charges, with a realistic growth assumption |
Implementation | Configuration, integration with line-of-business systems, testing |
Data | Profiling, cleansing, ownership work and knowledge-base clean-up |
Security and assurance | Security testing, DPIA, equality impact assessment, transparency records |
Change and training | Process redesign, role-based training, communications, staff time |
Running costs | Support, monitoring, model or prompt updates, governance and reviews |
Exit | Data export, decommissioning and the cost of switching if needed |
Usage-based costs deserve special attention. They can grow quickly as adoption spreads, so model a range rather than a single figure, and set a monitoring and review trigger.
How do you present AI benefits credibly?
By separating benefit types, showing where each figure comes from and being honest about confidence.
Separate cashable, non-cashable and enabling benefits. Cashable savings, such as reduced manual effort or lower demand on high-cost channels, are what finance directors can bank. Non-cashable benefits, such as faster responses or better quality, matter but must not be presented as savings. Enabling benefits, such as better data or reusable components, support future cases.
Start from a baseline. Every benefit should trace back to a measured current cost, volume or time, ideally collected during discovery.
Use pilot evidence. Scale results from the pilot, not from supplier claims, and explain how you have extrapolated.
Give each figure a confidence level. High, medium or low, with the reason. A modest, high-confidence saving is worth more than a large, low-confidence one.
Adjust for optimism. HM Treasury guidance expects appraisals to account for optimism bias. Apply a sensible adjustment to costs, timescales and benefits, and say what it is.
Show the timing. Benefits ramp up as adoption grows. Profile them year by year rather than presenting a single annual figure from day one.
Name the benefits owner. A service leader, not the programme team, should own each benefit and report on it.
How should AI-specific risks be handled in the business case?
Treat them as part of the appraisal, not an appendix. The government's AI Risk Management Toolkit gives a ready-made structure: nine categories of AI risk, including legal compliance, fairness, transparency, contestability, security and technical robustness, scored for likelihood and impact against the organisation's risk appetite.
In the business case, summarise the highest risks, the treatments agreed, their cost and the residual risk. Include the cost of controls, such as human review, monitoring and the ability to pause or switch off the system, in the financial case. A case that shows risks have been assessed and priced is far more persuasive than one that claims there are none.
