Most SME AI projects that fail were never properly evaluated before they started. The business case was either skipped entirely or built around vendor projections. When the results do not appear, there is no baseline to measure against, no decision gate to trigger a stop, and no way to learn what actually went wrong.

A good business case is not a financial exercise. It is a thinking tool. It forces you to define what success looks like, what it will cost to get there, and what circumstances would make you stop. Done well, it will either confirm the project is worth pursuing or identify the reasons why it is not, before you spend anything significant.

35% of UK SMEs now actively use AI, up from 25% in 2024
45% of those using gen AI report measurable cost savings
13 hrs/week per employee lost to low-value, pattern-based tasks
Sources: British Chambers of Commerce, September 2025 / OECD, 2025 / DocuSign, 2024

Start with a measurable before

The most important number in any AI business case is the baseline: what the process currently costs in time and money, before anything changes. Without it, you have nothing to compare your results to.

For each process you are considering, document the following before you look at any technology:

  • Weekly volume. How many times does this task occur in a typical week?
  • Average handling time. How long does one occurrence take, including checking and correcting the output?
  • Error and rework rate. What proportion of outputs need correction, and how long does that take?
  • Fully loaded cost per hour. Salary plus employer costs plus a reasonable overhead allocation.
  • Any downstream cost. If errors reach customers, clients, or suppliers, what is the typical consequence?

Multiply weekly volume by handling time by weeks per year, then apply the cost rate. That is your annual cost of doing this process manually. That number is the ceiling on what AI can save you.

The benefit categories that actually move the number

Not all benefits are worth pursuing equally. These four categories produce the clearest returns in SME contexts:

Time reclaimed

Hours freed from pattern-based work. The value is only real if those hours are redeployed into higher-value activity. If the saving just means the same person does slightly less work, the financial case is weaker than it first appears.

Revenue acceleration

Faster response to enquiries, faster quote turnaround, faster fulfilment. In businesses where speed of response affects conversion or retention, this can be the biggest number in the case.

Error and rework reduction

Where manual processes produce regular mistakes that cost time to fix or money to remediate, reducing error rate has a calculable value. Compliance-sensitive sectors have the most exposure here.

Capacity without headcount

Handling growth without proportional staff increase. If the business is capacity-constrained and would otherwise need to hire, AI may extend the runway before that hire becomes necessary.

Model all costs, not just the licence fee

The most common business case mistake is modelling only the software cost. The full cost picture for an AI project includes:

  • Software or API usage costs, including any usage-based pricing that varies with volume
  • Integration and setup effort, including internal staff time, not just external fees
  • Staff training time, at fully loaded hourly cost for everyone involved
  • Ongoing supervision and quality review, which does not disappear after go-live
  • Governance and compliance overhead, including data protection review if personal data is involved
  • Contingency for process redesign, because the process usually needs to be tidied up before automation works reliably

Add those up and you have a realistic total cost. Divide net annual benefit by annual run cost and you have a return ratio. Divide total setup cost by monthly net benefit and you have payback period.

A worked example: invoice processing in a 15-person business

Before

Finance administrator processes 40 supplier invoices per week. Each takes around 12 minutes: open email, locate PDF, cross-reference against purchase order, enter line items into accounting system, file. That is 480 minutes per week, or 8 hours. At a fully loaded cost of £20 per hour, that is £160 per week or roughly £8,300 per year. Error rate is around 5%, and each error requires about 25 minutes to resolve, adding another 30 minutes per week of rework.

AI solution considered

Document processing tool that extracts invoice fields and pushes them to the accounting system for human approval before posting. Administrator reviews and approves rather than manually entering. Expected to reduce handling time per invoice from 12 minutes to 3 minutes.

The numbers

Time saving: 9 minutes per invoice x 40 invoices x 50 weeks = 300 hours per year. At £20 per hour loaded cost, that is £6,000. Rework reduction adds approximately £500. Total annual benefit: around £6,500.

Costs: software £1,200 per year. Setup and integration: 12 hours internal time plus 8 hours external support, approximately £1,100 total. Training: 4 hours for one person, £80. Ongoing supervision: 20 minutes per week, £350 per year. Total first-year cost: approximately £2,730. Total run cost from year two: approximately £1,550.

Year one net benefit: £3,770. Payback on setup investment: around 4 months from go-live.

Good AI business cases include stop criteria, not only success criteria. Define in advance what result at what point would make you discontinue.

Use stage gates, not a single approval

Approve budget in phases rather than all at once. A four-stage model works well for SMEs:

  • Discovery: Baseline measurement and tool shortlisting. Low cost, time-bounded.
  • Pilot: Live test on a limited subset of the workflow. Release only if baseline metrics are met within agreed timeframe.
  • Controlled rollout: Expand to full workflow with monitoring. Release only if pilot showed target improvement with acceptable error rate.
  • Business as usual: Standard operational review and cost tracking.

Each gate has a clear pass criterion. If it is not met, you stop or redesign before spending more. This keeps downside limited and makes it much easier to get organisational support for the next project, because stakeholders can see that you are managing risk properly.

What makes a business case fail

  • Accepting vendor projection numbers without independently verifying the time saving assumptions
  • Calculating benefit at full hours saved without accounting for redeployment reality
  • Omitting staff time costs from setup and training because they feel like "internal" costs
  • Building the case on best-case accuracy numbers rather than actual performance on your data
  • No stop criterion: the project continues spending without a trigger to reassess

Sources

  1. Department for Business and Trade, Business population estimates for the UK and regions 2024
  2. Office for National Statistics, Business insights and impact on the UK economy (22 January 2026)
  3. Department for Science, Innovation and Technology, AI Opportunities Action Plan (January 2025)
  4. UK Government, AI Playbook for the UK Government
  5. Information Commissioner's Office, Artificial intelligence and data protection guidance