Business cases and measuring value
AI investments compete with everything else. This module shows how to build a business case that survives scrutiny — and how to prove value.
Learning objectives
- Structure an AI business case
- Identify the full costs of an AI initiative
- Estimate benefits with ranges and assumptions
- Define business, adoption and quality metrics
- Design pilots with clear success criteria
3.1 What a business case must show
A credible AI business case answers four questions: what will change, what it's worth, what it will cost (in full), and how we'll know — with the decision points for stopping or scaling.
3.2 Count the full cost
| Cost | Often forgotten parts |
|---|---|
| Technology | Licences or API usage (per user or per token), hosting, vector databases, monitoring tools |
| Data | Cleaning, labelling, integration, access controls |
| People | Build team, subject-matter experts' time for testing and review, ongoing operations |
| Change | Training, process redesign, communication, temporary productivity dip |
| Risk & governance | Assessments, legal review, security testing, audits |
Usage-based pricing scales with adoption: model costs carefully at the expected volume, not the pilot volume.
3.3 Estimate benefits honestly
Time saved is the most common benefit — and the most often overstated. Ask: what will people do with the time? Unless it's redeployed to valuable work, reduced overtime or hiring avoided, it isn't a financial saving. Use ranges (low / expected / high) rather than single numbers, and state your assumptions.
200 support agents each spend 2 hours a day drafting replies. A pilot shows AI drafting cuts that by 30% with no drop in quality scores. That's 120 hours a day. If half of it absorbs growth that would otherwise need new hires, at a loaded cost of £30/hour over 230 working days, that's about £414,000 a year — before subtracting licence, integration and training costs.
3.4 Measure what matters
Define metrics at three levels before the pilot starts, with a baseline for each:
- Business outcomes — cost per case, revenue, customer satisfaction, error rates
- Process and adoption — active users, tasks completed with AI, time per task, override rates
- Model quality and safety — accuracy on test sets, groundedness, harmful-output rate, latency, cost per request
Adoption without outcomes is a vanity metric; outcomes without quality metrics hide risk.
3.5 Design pilots to learn
- A clear hypothesis and success threshold agreed up front ("≥ 25% faster with no fall in CSAT")
- A comparison group or before/after baseline
- Real users, real data, real volume — not a curated demo
- A fixed time box (often 6–12 weeks) and a scale / iterate / stop decision at the end
For your top use case from Module 2, write a one-page business case: the change, benefit range with assumptions, full cost (all five categories), three metrics with baselines, and the pilot's success threshold and time box.
See a worked example
Example — support drafting: Change: AI drafts replies, agents edit and send. Benefits: £250k–£500k/year (assumes 20–35% drafting time saved, half redeployed). Costs: licences £90k/yr, integration £60k one-off, training £20k, QA reviews 0.5 FTE. Metrics: handling time (baseline 7.5 min), CSAT (baseline 4.3/5), edit rate. Pilot: 30 agents, 8 weeks; scale if ≥ 25% faster and CSAT ≥ 4.2.
Key takeaways
- Show what changes, its worth, its full cost and how you'll know.
- Count data, people, change and governance costs — not just licences.
- Time saved is only money saved if it's redeployed.
- Measure outcomes, adoption and model quality together.
- Pilots need a hypothesis, baseline, time box and decision.