Automate the action
A confirmed date creates a reminder. A completed form creates a record.
Does this job need AI, automation, or a better process?
You do not need AI to copy a confirmed booking into a calendar. You might use it to turn a rambling customer message into a useful brief. The difference is what the job asks the software to do. Start there, and the technology choice becomes much clearer.
Give each job the right tool
One workflow can contain all three kinds of work.
Let AI interpret
Suggest a category or summary. Keep missing facts visible and the source close by.
Agree the process
Decide who acts and what happens next before asking software to do it.
Rules move work. AI interprets it.
Automation carries out a defined sequence: when a form arrives, create a record and assign an owner. AI can interpret less predictable material, such as an email written in someone’s own words. A useful system can combine both.
This is a practical distinction, not a strict technical border. Automation is a broad category, and AI can be one part of it. IBM describes both rule-based automation and systems that incorporate AI.
| The job | Start here | Why |
|---|---|---|
| Copy a form’s email address into the CRM | Automation | The input and destination are already defined. |
| Suggest a category for a free-text enquiry | AI within a workflow | The wording varies, so interpretation may help. |
| Remind an owner about an agreed follow-up date | Automation | A known date can trigger a known action. |
| Decide which colleague owns a new enquiry | Agree the process first | Software cannot resolve an ownership rule the team has never agreed. |
The expensive mistake is automating an unanswered question
“Can AI handle our enquiries?” is too broad to design or test. Try: “Can it identify the requested service and prepare a summary for the person who replies?” Now there is a specific input, output and owner.
Before choosing a tool, complete this sentence: When this happens, we want this result, in this place, for this person. If the team gives three different answers, resolve that disagreement first.
A worked example: an installation enquiry
Illustrative example: a small installation business receives messages asking about equipment, property type and possible dates. The following is a proposed workflow, not a client result.
One message might say, “We need two units upstairs before the office gets busy again.” AI could help identify the request and flag the missing date. It cannot infer a confirmed appointment from that sentence.
| Part of the work | Proposed approach | What to check |
|---|---|---|
| Capture the message | Form or shared-inbox connection | Keep the original message and a stable reference. |
| Identify the request | AI suggests service and missing details | Show uncertainty; do not fill gaps with guesses. |
| Assign the enquiry | A rule using service or location | Send unknown cases to a visible queue. |
| Prepare the reply | An approved template, with AI drafting if useful | A person checks quotes and commitments. |
| Set a reminder | A task with an owner and due date | Cancel or update it when the conversation changes. |
Give AI a smaller, better-informed job
Useful context includes your service descriptions, the fields you need to collect, and examples of acceptable outputs. “Be helpful” gives a model much less direction than “summarise the requested job, list missing details, and do not invent availability.”
Ask it to return unknown when information is missing. Keep the original input next to the suggestion so the person using it can judge the result quickly.
NIST identifies confident but false generated content as a risk. That is a reason to test the particular task and its consequences, rather than assume every fluent answer is correct.
Run a small test that can change your mind
Choose a batch of representative examples, including awkward ones. Agree what “good enough” means before looking at the output. An internal summary and a customer quotation need different acceptance criteria.
- Write the current task in one sentence. Include who checks it and what happens next.
- Collect examples: a clear request, a vague request, a duplicate, an irrelevant message and a request you cannot fulfil.
- Compare the proposed output with the action your team would actually take. Record missing facts and unnecessary edits.
- Check the complete handoff. A good summary is not useful if it lands in a queue nobody reads.
- Choose whether to keep, narrow or stop the idea. Include review effort, setup and ongoing ownership in the decision.
Sometimes the better answer is already in your software
A required form field can remove repeated clarification. A saved CRM view can expose overdue work. A reusable reply can be enough for a common question. These are worthwhile improvements even when no AI is involved.
Use AI where interpretation earns its place. Use straightforward rules where the answer is already known. And keep a person involved where the business judgement belongs.
Sources
Which task would you take off your team’s plate?
Tell us what happens today, which tools are involved and where the handoff gets stuck. We can help define a useful first version.
Discuss one useful taskUpdated