Which business task should you try AI on first?

Start with a repeated task that has stable inputs, a clear acceptance standard and someone available to review the result. Keep the first pilot small enough to measure and recover when an output is wrong.

A task-level decision is more useful than a prediction about an entire job. One role can contain drafting, judgement, customer relationships and payments, each with different consequences when something fails.

Are the inputs stable enough?

Choose work where you can explain what information the system receives. A product-description draft based on approved specifications is easier to bound than a request to improve the whole sales operation.

Keep the source material current and identify which version is authoritative. If staff already disagree about the correct price list or process, resolve that before expecting a tool to apply it consistently.

Record the cases that do not fit the normal pattern. A useful pilot has a route back to a person for missing information or unusual requests.

Can a person recognise an acceptable result?

Write a short checklist for acceptance. For a draft enquiry reply, it could cover the customer's actual question, the correct service information, a suitable tone and the next step that the business can genuinely offer.

Choose criteria that someone can inspect without repeating the entire task from scratch. When judging the result takes longer than doing the original work, the pilot may still teach you something, but it needs a different time model.

The time-savings method records checking and rejected outputs. Use it from the first day rather than estimating the benefit from a demonstration.

Who will review and approve the work?

Name the reviewer and make sure they have time for the role. Decide which results stay as drafts and which actions require explicit approval before being carried out.

The payment example shows the boundary of a drafting pilot. Moving money introduces consequences that a rejected draft does not. Decide the approval process before any system has the ability to act.

Is the data approved for this purpose?

Use information that the firm has approved for the chosen service and purpose. Public product specifications make a simpler starting point than customer health details or a complete client file.

The ICO's governance guidance explains the written-contract requirement when a controller uses a processor. A business subscription by itself does not settle the firm's purposes, supplier roles or other obligations.

Work through the customer-data questions with the person responsible for data protection before a trial uses personal information. Record an unresolved issue as a reason to choose a different first task.

Can the business recover from a mistake?

Describe the likely error and the recovery step. An incorrect draft can be discarded. A wrong message already sent, a changed customer record or a payment may take more work to resolve.

For a UK retailer, a sensible first trial might prepare draft descriptions from its current catalogue. A person checks the result before it goes on the website. The same trial should not quietly gain permission to change prices, promise a delivery date or issue refunds.

Keep access limited to the information and actions the pilot needs. Record the exceptions that require a person and review them alongside the timing data.

Has a UK firm used a workflow-led approach?

Cast Consultancy's government case study describes training around a specific construction-document workflow: Gateway 2 submissions. Human validation remained part of the process. This self-reported example illustrates choosing a bounded use case; it does not establish a savings percentage for your business.

How do I choose between two suitable tasks?

Compare frequency, current time, likely review effort and the value of doing the work sooner. Prefer a task where the team can gather real observations without disrupting the rest of the business.

The business AI readiness score can help identify wider preparation work, while the time calculator explores planning assumptions. The five questions on this page remain a task-selection method, with the pilot supplying the evidence.

Use the ROI model after you have measured usable time and costs. A small, reviewed workflow with evidence can give the next decision a firmer basis.

Which questions come up most often?

Should the first pilot run without a person checking it?

Choose the review and approval process before the pilot starts. Drafting with review makes errors easier to inspect and recover.

Does task exposure predict that a job will disappear?

A task screen describes work that might be assisted. Staffing decisions need their own evidence and responsibilities.

What if the data has not been approved?

Choose a task using approved information while the responsible person resolves the data question. Record the restriction in the pilot scope.