A new enquiry arrives. Someone copies it into a spreadsheet, decides who should handle it and writes an internal summary. The same sequence happens again tomorrow. This is a more useful place to explore automation than a vague goal to “put AI into the business”.
Map the work before selecting a tool
Write down what starts the process, what information arrives, what happens to it and who uses the result. Include the awkward cases: a missing email address, a duplicate message or a request that does not fit any service.
Watch someone perform the task. Steps that sound simple in a meeting often involve decisions that have never been documented. Clarify those decisions before asking a system to repeat them.
Separate fixed rules from language tasks
Use ordinary rules when the answer is explicit. Checking whether a required field is present or routing a selected service to its owner does not require a language model. Clear rules are easier to inspect and test.
AI may help when the input is varied text: summarising a long enquiry, suggesting a category or drafting a reply from approved information. Keep the result labelled as a draft until you have established a suitable review process.
A first project: prepare an enquiry for the team
Consider this example workflow. It is a proposed design, not a claim about a deployed customer system.
- Receive: save the enquiry and issue a reference number.
- Validate: check required fields, apply abuse controls and identify duplicates.
- Prepare: produce an internal summary of the request, with missing details clearly marked.
- Review: let the assigned team member check the original message and suggested next step.
- Respond: send the approved reply and record the status.
The automatic receipt email can use a fixed template. The AI-generated internal summary helps the team prepare, while a person retains responsibility for promises about price, scope and delivery.
Design for failure, not just the happy path
What happens if the AI service is unavailable? The enquiry should still be saved and visible to the team. What happens if a process retries? The customer should not receive the same message three times.
Use stable record identifiers, track processing status and make repeated actions safe. Set a limit on retries and route unresolved items to a person. Keep enough logging to investigate a failure without copying unnecessary personal data into every system.
Be careful with text that arrives from customers or external documents. It may include instructions that conflict with the task. The controls discussed in OWASP’s prompt injection guidance are relevant when AI is connected to a workflow, especially one that can change records or send messages.
Run a small pilot with a baseline
Before changing the process, record how long the work takes, how often it needs correction and which steps are commonly missed. During the pilot, compare similar tasks and include the time spent checking AI outputs.
Track exceptions as well as successful runs. A workflow that saves a minute on most enquiries but creates occasional serious mistakes may need a narrower scope. Give one person responsibility for reviewing the pilot and deciding whether to extend, change or stop it.
Write the brief in one paragraph
A useful brief might read: “Prepare an internal summary for each valid enquiry, keep a link to the original, flag missing information and leave sending the reply to the team. If summarisation fails, create the normal enquiry record anyway.”
That is specific enough to design and test. Once it works, consider the next task. For an assistant that speaks directly to visitors, read our chatbot planning guide; the responsibilities are different when the output is customer-facing.
Have a project in mind?
Let’s talk it through