How to choose your first AI workflow in a small service business

Describe three tasks at the same level of detail

Ask the people doing the work to name tasks they repeat and the step that consumes time. For each candidate, write one sentence containing a trigger, an output and an owner. “When a website enquiry arrives, prepare a source-linked summary and reply draft for the sales owner” is a usable starting point.

Keep the candidates similarly narrow. Comparing one email-drafting step with the whole customer service function will distort the choice. Record what happens before and after each task, so you can see whether improving it would move the work forward.

Also ask whether the task needs AI. If it only copies an exact field, sends a fixed reminder or routes by a clear rule, a simpler automation may be sufficient. OpenAI’s practical agent guide recommends validating the need for an agent rather than assuming one is required; it highlights work involving unstructured information or decisions that are difficult to express as fixed rules.

Rule out poor starting points before scoring

Some conditions should pause a candidate regardless of its apparent value. If nobody owns the process, appoint an owner first. If access to the necessary data is not approved, resolve that before a pilot. If success cannot be defined, narrow the task until it can.

• Pause tasks that require unapproved access or unnecessary sensitive data.
• Pause tasks with no reliable source for the facts the output needs.
• Pause tasks whose mistakes could create commitments your team cannot review safely.
• Pause tasks where the proposed output does not solve the actual bottleneck.

An internal summary is often easier to test than an unsupervised external action. That does not make every internal task low-risk; a report used for major decisions still needs appropriate checking.

Score six practical dimensions

Rate each dimension from 0 to 2 using evidence from recent work. Zero means the condition is weak or unknown, 1 means partly ready and 2 means reasonably well established. The total helps compare candidates, but it should never override a blocking issue.

• Frequency: 0 for rare or unknown demand; 1 for recurring but limited demand; 2 for enough regular cases to observe performance during the pilot.
• Effort: 0 when little measurable work is removed; 1 for some repeatable preparation; 2 when a clearly time-consuming step can be isolated.
• Input readiness: 0 for missing or unreliable information; 1 for sources needing cleanup; 2 for approved, accessible and reasonably current sources.
• Output reviewability: 0 when errors are hard to detect; 1 when review needs specialist effort; 2 when a named reviewer can check the result against evidence.
• Containment: 0 when errors trigger difficult-to-reverse consequences; 1 when boundaries need work; 2 when the pilot can remain draft-only or otherwise safely limited.
• Ownership and measurement: 0 for neither; 1 for partial arrangements; 2 for an accountable owner, a baseline and agreed acceptance measures.

Write the evidence beside the score. “Two points for frequency” is less useful than “We received 180 eligible enquiries last month and can access representative examples.” Use your own volume; there is no universal minimum that makes a workflow profitable.

An illustrative comparison

Imagine a small agency considering three candidates. Enquiry reply preparation happens regularly, draws on a service brief and has a sales owner available to review drafts. An end-of-month report takes considerable time but depends on inconsistent spreadsheet definitions. Automatic discount negotiation would affect commercial terms and lacks a clear approval policy.

Using the six dimensions in order, the agency might rate reply preparation 2, 2, 2, 2, 2, 2 for a total of 12. Reporting might receive 1, 2, 1, 1, 2, 1 for a total of 8. These invented scores illustrate the method, not the typical value of either task.

The discount candidate should be paused before scoring because authority and review boundaries are unresolved. A high potential revenue figure would not repair that gap. For this example, enquiry drafting is the clearest first pilot, while reporting needs source cleanup before a fair test.

Check whether the likely benefit can cover the effort

A well-contained workflow can still be too small to justify its cost. Estimate eligible volume, current active handling time and the human review time likely to remain. Add setup, software, internal maintenance and ongoing management costs.

Do not rely on a dramatic percentage reduction from a tiny task. Saving half of a two-minute activity once a week is different from releasing several minutes on hundreds of monthly cases. Conversely, a low-volume task may matter because it reduces costly errors, but that value needs its own evidence.

Use the AI workflow ROI calculation to build a conservative scenario. Keep potential sales upside separate until there is evidence to attribute it. If the economics are weak, choose a smaller intervention or a different task.

Write a pilot brief before building

Once you choose a candidate, document the scope in a page your team can understand. The brief should make it possible to judge a result without moving the goalposts after the demonstration.

• Trigger and scope: which cases enter, and which are excluded.
• Output: the exact draft, summary or proposed update to be produced.
• Sources and access: approved information and the minimum required permissions.
• People: process owner, reviewer, backup and technical support responsibility.
• Exceptions: what must stop, escalate or return to the manual process.
• Measures: baseline, handling time, quality, review effort and operating cost.
• Decision: the review point and conditions for continuing, changing or stopping.

Choose a pilot period that includes enough representative work and the variations you need to test. Time alone is not evidence. A quiet fortnight with three straightforward enquiries will tell you less than a period containing routine requests, duplicates, missing information and exceptions.

Expand only after the first task is understood

Review the actual outputs with the people using them. Look for repeated edits, delayed approvals, unresolved exceptions and unexpected costs. Decide whether the remaining problems are specific enough to fix. Keep a working manual route while the pilot is being evaluated.

Being AI follows diagnosis, pilot, deployment and ongoing management. The first diagnosis is a place to compare candidates, define boundaries and assess tool access, rather than assume that every business needs the same system. Request a workflow diagnosis with your three candidate tasks and the one your team most wants to improve.