Your first AI automation should be a small, useful one.

Start with a task your team already understands, does often and can check easily. A modest automation that reliably gives time back is a better first project than an ambitious assistant with no clear job.

For a small business, the opportunity often sits in an ordinary handover. An enquiry arrives, somebody reads it, copies details into another tool, chooses a colleague and drafts a reply. None of those steps is particularly difficult. Repeating them every day is where the time goes.

The first question is not which AI product to buy. It is which part of that handover would be useful to improve, and what you would need to trust the result.

Pick a task with a clear beginning and end

Describe the job in one sentence: “When a website enquiry arrives, prepare a summary and a suggested service category for a person to review.” That is specific enough to test. “Use AI to transform sales” is not.

Choose a task with recurring inputs, a recognisable result and a named owner. Avoid making a first pilot responsible for approving payments, giving regulated advice or making consequential decisions about people. The more costly an error would be, the more carefully the system and its review process need to be designed.

Often the best solution combines these approaches. Rules handle the predictable parts; AI assists where language varies; a person handles uncertainty and commitments.

An enquiry workflow you could pilot

Here is an illustrative design, not a claim about a deployed customer system. A form submits the person’s message. The workflow prepares a short summary, suggests a category and drafts an acknowledgement using approved wording. A colleague checks it before anything is sent.

When the message is unclear, the result should say so. A request outside the supported services should be routed for review. A visitor’s instructions inside a message must not be allowed to change the workflow’s rules or grant it access to other systems.

For an early pilot, the AI does not need permission to send email or change customer records. Producing a draft may be enough to discover whether the assistance is useful.

Decide what information may enter the tool

Begin with synthetic examples or a carefully prepared set that does not expose private customer information. Before using real enquiries, establish what data is sent to the provider, who can access it, how it is retained and whether it is used for model training. Check the actual service terms and settings rather than assuming all business AI products behave alike.

NIST’s AI Risk Management Framework provides a structured way to consider risks in context, including how a system is evaluated and managed. For a small pilot, the practical application is to assign an owner, define acceptable results and keep a record of failures.

Measure the work that remains

Time the current task before the pilot. Then measure the automated version including review, corrections, exception handling and maintenance. Count the jobs completed correctly, not just the drafts produced.

For example, if a task takes six minutes manually and a reviewed draft reduces that to three, twenty tasks would release an hour. That is arithmetic for a hypothetical pilot, not a promised saving. If checking the draft takes longer than doing the task yourself, the design needs work.

A successful pilot earns the next step. It may lead to a wider workflow, a simpler rules-based automation, or a decision to leave the task with a person. All three are useful outcomes when they follow evidence from the work.