Business process automation: where to start, what to automate, how to measure it

Good candidates repeat dozens of times a month, follow stable rules, produce verifiable output and currently consume low-value manual hours. Typical cases: re-keying the same data into two systems, manually assembling recurring reports, routing inbound requests, routine document checks. Unstable, exception-heavy processes should be simplified first — automating disorder only makes it faster.

For each candidate you need three numbers: monthly frequency, average minutes per run, and detected errors. The first two give the hours at stake; the third gives the hidden cost of rework. Without that baseline the project cannot be judged afterwards, which is the most common reason working automations are perceived as disappointing.

Integrations between systems via API remove re-keying and offer the best value-to-cost ratio. Rules and workflows encode states, approvals, deadlines and alerts. Document automation extracts data from PDFs, invoices and contracts. Generative AI classifies and summarises free text. In most SMEs the first two levers deliver the majority of the value.

Use AI when the input varies in form but not in meaning: request emails, heterogeneous supplier documents, free text to classify. Do not use it where a deterministic rule exists — calculations, thresholds, routing on structured fields — because explicit logic is cheaper, verifiable and stable. Any AI component in production needs human review on low-confidence cases and continuous accuracy monitoring.

Weeks 1-2 process map and baseline; weeks 3-5 build with full execution logging; weeks 6-7 run in parallel with the manual process to surface real exceptions; week 8 go live and measure the delta. Return is hours freed × loaded hourly cost, plus avoided error cost, minus amortised build and running cost. The most durable benefit is data reliability feeding your KPI dashboards.