Your team keeps reading the same kinds of emails, copying details between systems or sorting documents before the real work can begin. It’s tempting to ask whether AI could take over.
A better starting question is narrower: which part of that work could AI help with, and how would you know it had helped?
The best first candidate isn’t necessarily the most impressive demo. It’s a bounded task where you can check the output, contain mistakes and measure whether the whole process improves.
What is AI automation?
AI automation uses an AI model within a software workflow to perform a task, such as classifying an enquiry, extracting information from a document or drafting a response. Software handles the surrounding steps: collecting the input, applying checks, requesting approval and passing the result to another system.
That doesn’t have to mean a process running without people. For a first project, keeping someone in control of consequential actions is often the sensible design.
It helps to separate three approaches:
| Approach | How it works | Hypothetical example |
|---|---|---|
| Fixed-rule automation | Follows explicitly defined conditions and actions. | A form submission marked “Accounts” goes to the accounts queue. |
| AI-assisted work | A person uses AI for part of a task, then checks and uses the result. | A colleague asks AI to summarise an enquiry before writing a reply. |
| AI workflow automation | Software brings AI into a connected process, with defined checks and handovers. | An incoming enquiry receives a suggested category and summary, ready for staff approval. |
These approaches can work together. Use rules where the decision is clear; consider AI where the input varies and needs interpretation. Don’t add AI to a job that a reliable form field or straightforward rule already solves.
Automate a task, not the whole responsibility
“Handle our enquiries” is too broad for a useful first trial. It includes understanding the message, judging urgency, choosing an owner, deciding what to promise and communicating with the customer.
“Suggest an enquiry category for a colleague to approve” is much easier to evaluate.
Consider two hypothetical candidates:
Sorting enquiries. AI suggests a category from an agreed list and produces a short summary. A colleague sees the original message alongside the suggestion and approves or changes it. Nothing is sent to the customer automatically, and enquiries that cannot be processed stay visible in a manual queue.
Extracting document fields. AI proposes a supplier name, invoice number, date and total from an invoice. The system checks required fields and formats, then shows the values beside the source document for review. Missing information stays missing rather than being guessed. Approval doesn’t authorise payment.
Both examples separate useful preparation from responsibility for the final action. That boundary gives you something manageable to test.
What makes a first workflow worth shortlisting?
List a few recurring tasks, then compare them against these questions. You’re looking for a practical candidate, not trying to justify AI at any cost.
Can you get the right data safely?
Identify where the input lives, who can authorise access and how software would retrieve it. A manual export might be enough for an initial test; a live workflow may need an API, meaning a supported way for systems to exchange data.
Check whether the proposed service can receive the information involved. For personal or confidential data, confirm the relevant permissions, contractual terms, retention arrangements and UK data protection requirements before testing with live records. Use synthetic or appropriately redacted examples while those questions remain unresolved.
If data access depends on a third-party supplier, confirm its capabilities and restrictions before committing to the workflow.
Can a person check the result efficiently?
A reviewer needs the source material, the proposed output and a clear action: approve, correct or return to manual handling.
“Human review” isn’t much protection if the person has no time, context or authority to challenge the result. Nor is it much of a saving if checking takes longer than doing the original task.
Prefer work where errors are reasonably easy to spot. A short set of extracted fields is usually more straightforward to compare with a source than a lengthy answer requiring specialist judgement.
What happens when it fails?
Plan for both technical failures and plausible but incorrect answers. The service might be unavailable, a document might be unreadable, or a suggested category might simply be wrong.
Define a visible manual fallback, an owner for exceptions and a way to pause the workflow. If processing is retried, ensure it won’t create duplicate records or repeat an action. Treat incoming documents and messages as data, not as permission to change the workflow’s instructions or access other systems.
A model’s confident wording is not proof that its answer is correct.
Is there enough useful work to justify it?
Look for a recurring task with a clear input, a clear output and a meaningful burden today. Avoid starting with rare exceptions, poorly understood processes or decisions where an error could cause serious harm.
Your shortlist can stay small. Of the candidates that pass the access and safety checks, favour the one with the easiest verification and clearest benefit.
Measure the whole job, including the checking
Before a trial, record how the task works now. Then compare a representative sample, including awkward cases rather than only tidy examples.
Useful measures include:
- Staff time per item, including review and corrections.
- The proportion of outputs accepted without changes.
- Errors, separated by their consequences rather than counted as equally important.
- Items returned to manual handling.
- Running and support costs compared with the time or service improvement gained.
Agree what would justify continuing before you see the results. Faster model output means little if it creates more checking elsewhere.
Your first AI workflow should earn its place through ordinary usefulness: less copying, clearer queues or quicker preparation, without making mistakes harder to catch. Sometimes the answer will be fixed-rule automation instead. That’s a useful finding, too.
If you have a few candidates but aren’t sure where to start, talk to Atlas-James. We can help you examine the workflow, its constraints and what a worthwhile first test would need to show.