How to Build the Business Case for AI Automation (Without Overselling It)

Step 1: Measure the Current State Precisely

The foundation of any automation business case is an accurate measurement of the current state. Not an estimate — a measurement. How many times does this process run per month? How long does it take (median, p90, p99)? What is the loaded cost per unit? What is the error rate? What downstream costs does each error generate?

Teams that estimate current-state metrics often get them wrong by 2-5x. The process runs more frequently than anyone remembers. The actual time per unit is longer when you measure it, because people account for interruptions, context-switching, and rework. Measuring before you model pays for itself in credibility when the numbers are challenged.

Step 2: Model Benefits Conservatively

The common mistake: project 80-90% automation of the process and model the benefit as (0.85 × current cost). The reality: automation captures certain parts of the process while creating new overhead in others (monitoring, exception handling, data quality), and adoption is rarely 100%.

— Model two scenarios: conservative (50-60% of tasks automated, 90% adoption) and realistic (70-80% automated, 95% adoption)

— Subtract new costs: monitoring, system maintenance, exception handling, retraining over time

— Include transition costs: parallel running period, staff retraining, integration work

— Do not model first-year savings as full-year — ramp-up typically takes 3-6 months to reach steady state

— Use the conservative scenario as the primary case; use the realistic scenario as the upside case

Step 3: Build a True Total Cost of Ownership

Business cases that underestimate TCO create budget problems that surface at the worst moment — mid-implementation, when costs are already committed.

— Build cost: engineering time, design, testing, data preparation, integration work

— Deployment cost: infrastructure, vendor onboarding, security review, compliance sign-off

— Year-1 operation: monitoring, bug fixes, model evaluation, user support, prompt tuning

— Ongoing: model retraining or prompt updates as the process evolves, infrastructure scaling, annual vendor cost escalation

— Hidden: management overhead, exception handling by humans for low-confidence outputs, audit and compliance reporting

The Metrics That Make the Case Defensible

The most defensible business cases connect AI automation to metrics the business already tracks and cares about — not AI-specific metrics that require context to understand.

— Cost per unit processed (reduce from £X to £Y)

— Processing time (reduce from N hours to M minutes)

— Error rate and error remediation cost (reduce by Z%)

— Staff hours reclaimed and their redeployment value

— Customer experience impact (if the process is customer-facing)