Most AI automation projects do not fail because the tools are weak. They fail because the wrong process was chosen: something rare, something ambiguous, or something whose broken version was simply made faster.
Score candidates before you build
Score every candidate process from 1 to 5 on five axes:
- Frequency. How often does it run? Daily beats monthly.
- Rule clarity. Could you write the decision rules on one page? If not, an automation will guess.
- Cost of error. What happens when it gets one wrong — a small correction, or a lost customer?
- Data readiness. Does the information live somewhere structured, or in six inboxes?
- Time consumed. Real hours per week, measured, not estimated.
High frequency, clear rules, low error cost, ready data, high hours. Start there. Anything scoring low on rule clarity or high on error cost needs a human in the loop, not full automation.
The processes that almost always pay off first
- Lead qualification and routing. Inbound enquiries scored, enriched and routed within seconds instead of hours. Speed of first response is one of the few reliably decisive variables in sales.
- First-response handling. Instant, accurate answers to repeat questions across web chat, email and WhatsApp, with clean escalation.
- Data entry between systems. Form to CRM, CRM to invoicing, invoicing to reporting. Unglamorous, high volume, low error cost.
- Reporting assembly. Pulling numbers from several platforms into one consistent view, on a schedule.
- Follow-up sequences. Nothing sits unanswered because someone forgot.
What to leave to people
- Pricing negotiations and contract exceptions.
- Complaints, cancellations and anything emotionally loaded.
- Judgement calls with legal, safety or financial consequence.
- Anything where being wrong once costs more than being slow a hundred times.
The mature pattern is assistive: AI drafts, gathers and routes; a person approves. That is where most of the value sits and most of the risk does not.
Fix the process before automating it
Automating a broken process produces broken outcomes at scale. Before building anything, write the current process down step by step, mark the steps that exist only for historical reasons, and delete those. Frequently the deletion saves more time than the automation would have.
A sane sequence
- Map and measure. Two weeks. Document the process, count the hours, agree the success metric.
- Build one. Two to four weeks. One process, end to end, with logging and a human escalation path.
- Run it in parallel. Two weeks. Automation and manual process both live, outputs compared.
- Cut over and monitor. Keep the escalation path permanently.
- Add the next one. Only after the first is stable.
Teams that try to automate six processes at once usually finish none.
Non-negotiables in the build
- Logging. Every action recorded and reviewable. Unauditable automation cannot be trusted or debugged.
- Escalation. A clear, fast path to a human, triggered by uncertainty as well as by request.
- Ownership. A named person responsible for the automation's behaviour.
- Review rhythm. Monthly sampling of outputs. Systems drift as your business changes.
Measuring it honestly
Track hours returned, response time, error rate and cost per handled task — before and after, same definitions. "It feels faster" is not a result. If a process was taking eleven hours a week and now takes two, that is the number that justifies the next build.
Our AI Business Automation programme follows exactly this sequence, and Workflow Automation covers the integration layer between your existing systems.