Where AI can help a business workflow, and where a person should decide

Map the task before choosing a tool

Write down what starts the workflow, what information arrives, what a person does today and what a successful outcome looks like. Separate repeatable preparation from decisions that need judgement. For an enquiry inbox, AI might identify a requested service and draft follow-up questions; a person can still decide whether the project fits and what to quote.

Choose a task with enough recurring volume to evaluate, but limited consequences if a draft is wrong. Record the current time spent and common failure points. Without that baseline, an impressive demonstration can be difficult to distinguish from a useful operational improvement.

Define approved sources and necessary information

Give the system a maintained source of truth: current service descriptions, approved policies, product information or internal procedures appropriate to its task. Decide who updates those sources. A website assistant should be able to say that a detail needs checking when the answer is absent, rather than filling the gap with an invented price, deadline or capability.

Collect and expose only the information the workflow needs. Review the provider’s data handling, access, retention and deletion settings before introducing customer or internal information. Keep credentials out of prompts and logs, and distinguish operational records from anonymous measurement.

Set boundaries around consequential actions

Document what the assistant can prepare, what it can save and what requires approval. Drafting a project summary is different from accepting a contract, changing a customer record or sending a message. The interface and connected tools should enforce those boundaries, rather than relying on a friendly instruction alone.

Provide a visible route to a person and explain what happens next. A request received by a website is not necessarily an appointment booked or a proposal accepted. Confirmation language must match the actual completed action so customers can rely on it.

Test the whole handoff with representative examples

Build a small evaluation set from approved, appropriately de-identified examples. Include ordinary requests, missing information, ambiguous wording, unsupported services and attempts to make the assistant ignore its instructions. Check whether answers stay within the available evidence and whether the assistant asks a useful question when details are missing.

Test integration failures as well as conversation quality. A disconnected service, a duplicate click or an interrupted session should not create misleading success messages or repeated records. Use a clear saved-state response, duplicate handling where appropriate and a fallback that lets the visitor continue without the assistant.

Measure the pilot and expand deliberately

Compare completion time, correction effort, missed requests and the quality of the final handoff. Track how often a person must intervene and why. Include model usage, integration maintenance and source updates in the cost of the workflow. A faster draft is not a saving if every result requires extensive repair.

Review a sample of outputs regularly and document the failure patterns. Improve the sources, narrow the task or change the process before adding more automation. Expand only when the pilot produces a repeatable benefit and the person responsible can explain its limits. AI consulting should leave the business with a manageable workflow and a way to evaluate it.

Your planning checklist

• Choose a repeatable task with a measurable baseline and a named process owner.
• Define approved sources, necessary data and retention expectations.
• Require review for commitments and provide a working human handoff.
• Evaluate realistic failures, correction time and operating costs before expanding.

Common questions