Choose a first AI workflow by measuring the current task, testing useful output and counting review time. A worked example shows how to compare options.
Start with a task whose current result you can describe and measure. Then compare a small change against that baseline. The first decision is whether the proposed workflow produces a useful result at an acceptable cost, including the effort to check it.
Write down one task from start to finish
For a sales follow-up, the task might begin when approved meeting notes become available and end when a representative sends a checked email. Generating a draft is one step within that task. It is not the same as completing it.
Record the weekly volume, active handling time, waiting time, common corrections and the systems involved. Ask the person doing the work to show recent examples. Distinguish a missing input from a slow drafting step: AI drafting will not fix a process that waits three days for the notes.
Choose a task where you can define a satisfactory result, obtain appropriate test material and identify who will review exceptions. Compare a model-based approach with an improved template, a system integration or a fixed rule. NIST’s AI RMF Playbook explicitly includes consideration of viable non-AI alternatives.
Count the work after generation
Suppose a fictional team prepares 40 follow-ups a week, taking 12 active minutes each. Its baseline is 480 minutes. In a proposed pilot, preparing and checking each draft takes six minutes. Eight difficult cases need five additional minutes of correction, and the weekly queue check takes 20 minutes.
| Work in this illustration | Calculation | Weekly minutes |
|---|---|---|
| Current method | 40 × 12 | 480 |
| Draft preparation and review | 40 × 6 | 240 |
| Additional corrections | 8 × 5 | 40 |
| Queue check | One weekly check | 20 |
| Proposed total | 240 + 40 + 20 | 300 |
The difference is 180 minutes, or three hours per week. These are invented planning figures, not a measured client result. They assume comparable output quality and that the additional correction time is not already included in the six-minute figure.
At an assumed labour cost of $60 per hour, three hours represents $180 of weekly staff capacity before software, implementation and other costs. It does not establish a $180 reduction in payroll or an increase in sales. Replace every assumption with observed figures before making a purchase decision.
Prepare a pilot brief someone can challenge
Keep the initial brief specific:
- Task boundary: approved meeting notes to a checked follow-up ready for a representative to send.
- Permitted action: prepare a draft; do not send or change the customer record automatically.
- Required output: correct recipient, accurate commitments and no invented price, date or promise.
- Test material: authorised examples covering routine, incomplete and conflicting notes.
- Owner: the person who accepts the result, investigates failures and decides whether to continue.
- Measurement: total active time, correction types, unresolved cases and operating cost.
Include a case where there is not enough information to draft the message. A useful result can be a request for clarification rather than a polished email that fills the gap with a guess.
Agree what would change the decision
Set acceptance criteria before observing the pilot. A wrong recipient or invented commercial commitment may be a release-blocking failure even if average drafting time falls. The appropriate thresholds depend on the task and its consequences.
Compare the same types of work using the current and proposed methods. Keep an untouched set of examples for a later check so that repeated prompt changes do not merely fit the demonstration cases. Record enough detail to explain which cases improved and which became harder.
Proceed when the evidence supports the intended use. Revise when a specific, testable weakness has a plausible remedy. Stop when the work is too infrequent, the review burden removes the benefit or a simpler method meets the need.
Bring the brief and baseline to an AI strategy assessment. If the main issue is data moving between tools, an integration review may be the more useful starting point.
