Which workflow in your business would you improve if you could only change one thing this quarter — and how would you know it worked?

That question matters more than which AI model you use or how much you budget. Most first AI projects fail for a boring reason: the team picks a workflow they cannot measure, changes it, and cannot say whether anything improved. The fix is unglamorous. Choose one workflow with volume, repetition and a measurable output, baseline it before you change anything, and agree in advance what success looks like.

What makes a good first candidate

A workflow is a good first AI project when three things are true:

  • Volume. It happens often enough that a modest improvement per occurrence compounds. Ten times a week is workable; ten times a year is not.
  • Repetition. AI assists well where judgement is applied to patterned work — summarising, classifying, extracting, drafting — rather than genuinely novel situations.
  • Measurable output. You can count the thing: emails processed, leads entered, tickets routed, hours spent. If you cannot count it, you cannot baseline it, and if you cannot baseline it you cannot prove value.

Four workflows consistently meet these criteria for established SMEs:

  • Lead-to-CRM capture. Enquiries arrive by email, web form and phone, and someone retypes them into the CRM. Records created, fields completed and time from receipt to entry are easy to count.
  • Support triage. Incoming support requests need classification, prioritisation and routing. Baseline the misroute rate and response time, then measure again.
  • Knowledge search. Staff repeatedly answer the same questions from scattered documents. Measure time to answer, escalation rate and user feedback.
  • Reporting. Weekly or monthly reports assembled by hand from systems and spreadsheets. Hours spent is the metric, and it is rarely flattering.

What to avoid

Two categories cause most of the trouble:

  • Anything safety- or compliance-critical. Decisions affecting people's safety, legal standing, health or money should not be your first automation. These can be assisted later, under tighter governance, once you have measurement discipline on lower-risk work.
  • One-off or rare tasks. A task done twice a year cannot repay the setup, testing and monitoring effort, however impressive the demo.

A useful rule: if a mistake would be expensive to unwind, or the task is too infrequent to measure, it is a second or third project, not a first.

Illustrative example: a 40-person professional services firm receives around 60 enquiry emails a week. An office manager copies each into the CRM by hand, usually the next morning, and roughly one in five records needs correction. The firm baselines this for two weeks (median time from receipt to CRM entry, and the share of records corrected), then introduces an AI-assisted draft: the model proposes CRM fields from the incoming email, the office manager reviews and approves each one, and the record is created at send time rather than the next day. Success criteria are written down beforehand: faster median entry and fewer corrected records, with the human reviewer keeping final say. The point is not the technology — it is that everything is countable and a human stays in the loop.

Baseline before you change anything

Before any change, record for two to four weeks:

  • how often the workflow runs;
  • how long it takes, per occurrence and in total;
  • the error or rework rate;
  • who does it.

You do not need precision to the minute. You need a defensible number you can compare against later. "About three hours a week, and a fifth of records get corrected" is enough to start.

Define acceptance criteria before you build

Acceptance criteria turn an ambiguous ambition ("use AI to save time") into something testable:

  • the specific metric (entry time, misroute rate, hours per report);
  • the target or direction of improvement;
  • the review period;
  • the fallback if results disappoint.

Agree these with the people who own the workflow, not just the people building it. If the process owner would not sign the criteria, the project will stall the first time a result is inconvenient.

Start with a human in the loop

For a first project, the AI should propose and a person should approve. This is slower than full automation, and that is the point: you are buying measurement and trust, not raw speed. Reviewer approvals, overrides and corrections are themselves data — they show where the system is confident and where it guesses. Only after sustained, boring reliability should any step run unattended, and even then only for the low-stakes branches of the workflow.

Constraints and trade-offs

Honesty about the costs keeps the project sane:

  • Data quality is the real bottleneck. If your documents contradict each other, an assistant built on them will contradict itself. Cleaning inputs takes longer than building the assistant.
  • Setup cost is front-loaded. A first project pays for learning: your first measurement discipline, acceptance criteria and review loop. The second project is materially cheaper.
  • Review time is not free time. A human-in-the-loop workflow shifts effort rather than eliminating it. Measure the total, not just the automated slice.
  • Scope creep is the default failure mode. "While we're in there" is how a one-workflow project becomes a six-month platform programme.

Your action checklist

  • List your candidate workflows and score each for volume, repetition and measurability.
  • Choose one. Write down why it beats the others in one sentence.
  • Baseline it for two to four weeks before changing anything.
  • Write acceptance criteria the workflow owner will sign.
  • Design it human-in-the-loop from day one.
  • Measure against the baseline for an agreed review period.
  • Decide to scale, adjust or stop — using the criteria you wrote, not the mood of the meeting.

If you want a second pair of senior eyes on your shortlist, our technical review examines candidate workflows, data readiness and measurement plans before you commit budget. If you already have a workflow in mind, our AI-assisted improvements service will baseline, build and measure it with you — one workflow at a time.