The most impressive AI idea in the room is often the worst place to start.
A sales forecast may sound more strategic than sorting document intake. A broad assistant may sound more ambitious than routing customer requests. But if the data is unreliable, nobody owns the result or “done” cannot be defined, the impressive idea is still a weak first workflow.
The better starting point is usually a smaller piece of real work that already repeats, has an owner and ends in an observable outcome.
The aim is not to prove how capable AI is. It is to learn whether one workflow can become faster, clearer or more reliable without creating a larger problem around it.
Use six filters before choosing a tool.
A strong candidate is repeated, grounded in real inputs, observable, checkable, contained and measurable.
Use a recent example to expose the real friction.
Include the source information used to resolve them.
Connect the output to an action, record or status.
Make examples and acceptance criteria visible.
Keep approval around consequential actions.
Pair it with a quality guardrail.
Select one bounded workflow that can produce trustworthy evidence.
Start with a workflow, not a tool
“Use AI in the business” is not a useful project brief.
A workflow is more concrete. It begins with a recognisable trigger, moves information or a decision through a series of steps and ends in a business outcome.
For example:
a client request arrives and reaches the right person
supporting documents are checked before a file moves forward
a site update becomes an invoice-ready record
meeting notes become assigned follow-up actions
a recurring report highlights the cases that need attention
This distinction matters because an AI output is not necessarily a completed workflow. A summary can be accurate while the task remains unassigned. A draft reply can sound good while the client still receives nothing. Extracted invoice details can be correct while finance cannot see that approval is pending.
Choose the business journey first. Then decide whether AI belongs in one part of it.
Use six filters to compare candidate workflows
A good first workflow does not need to score perfectly on every filter. It should be strong enough that the team can run a bounded test, see what happened and make a sensible decision.
The work repeats often enough to matter
Look for work that appears daily or weekly, or less frequent work with a meaningful consequence when it is delayed or missed.
Frequency alone is not enough. A five-minute task can still create disproportionate friction if it interrupts several people, crosses multiple tools or regularly waits for someone to remember the next step.
Ask for a recent example. If nobody can show what happened last time, the opportunity may still be too vague.
The inputs are available and representative
AI needs more than a polished demonstration file.
The team should be able to provide ordinary examples, incomplete examples, unusual cases and the source information used to resolve them. That might include emails, forms, documents, CRM records, policies or previous approved outputs.
If the necessary context lives only in one experienced employee's memory, the first task is to make that knowledge visible. Adding AI before doing this can make guesses faster without making the process more reliable.
The finish line is observable
Define what “done” means in business terms.
For an enquiry workflow, done may mean that the request is classified, assigned and visible with a response deadline. For document intake, it may mean that required fields are captured, missing items are flagged and the file is ready for review.
“AI produced an answer” is not a finish line. The result should connect to an action, decision, record or status that the team already understands.
The output can be checked
A first workflow should make it reasonably easy to judge whether the output was useful.
Can a person compare the extracted data with the source? Can a manager confirm that the right cases were escalated? Can an approved example show whether a draft follows the company's actual standard?
If quality depends on an entirely subjective judgement with no examples or acceptance criteria, the team will struggle to learn from the test. Start by making the standard clearer, or choose a more checkable task.
The consequence of a mistake can be contained
The first workflow should have a safe operating boundary.
AI may prepare, classify, summarise or recommend while a person reviews consequential actions. Deterministic automation can carry agreed statuses, reminders and handovers. Human approval should remain where the work creates a commitment, changes a sensitive record, affects money or requires accountable judgement.
A reversible internal draft is a safer starting point than an automatic external promise. Authority can expand later if the evidence supports it.
The value can be measured
Choose one business outcome before implementation begins.
That might be turnaround time, backlog, avoidable rework, staff handling time, missed follow-up or the number of cases waiting without an owner. Pair it with a quality guardrail so speed does not hide worse work.
The purpose is not to build a large reporting exercise. It is to know whether the new workflow made ordinary work better enough to justify its cost and maintenance.
Separate AI work from fixed workflow rules
Once a candidate is selected, divide the job carefully.
AI is useful where the input is messy or language-heavy: reading an email, extracting details from varied documents, classifying a request, drafting a response or summarising a case.
Fixed automation is usually better for agreed actions: creating a task, updating a status, checking a required field, sending a reminder after a set period or recording who approved the next step.
People remain responsible for judgement, exceptions and important commitments.
This division makes the workflow easier to understand and maintain. It also prevents the AI component from quietly becoming responsible for the whole business process.
Interpret
Messy inputs, language and varied documents.
Execute
Agreed statuses, tasks, reminders and records.
Decide
Judgement, exceptions and commitments.
Begin in assisted mode
The first version does not need permission to act everywhere.
Let it prepare work for review. Keep the original source visible beside the proposed output. Record corrections and the reasons behind them. Make uncertain cases stop in one visible place rather than disappearing into inboxes or private messages.
This assisted period should answer practical questions:
- ✓Does the workflow handle normal cases consistently?
- ✓Which exceptions appear repeatedly?
- ✓Is the source information good enough?
- ✓Does review save time, or does it create a second job?
- ✓Does the work reach the real finish line?
The decision after this stage may be to expand, simplify, change or stop. All four are useful outcomes when they prevent a weak idea from becoming a larger commitment.
Warning signs that a workflow is not ready
Pause before implementation when:
- •the problem is described only as “we need AI”
- •the work is rare and low-consequence
- •nobody owns the current process
- •no representative examples are available
- •the rules change constantly but are not documented
- •success cannot be distinguished from a plausible-looking output
- •the proposed first step gives AI authority that the team cannot safely review or reverse
- •the existing system already solves the problem and the remaining gap is unclear
In these cases, the useful next move may be process clarification, better source data, a simpler automation or no project at all.
A practical selection exercise
List three repeated workflows that currently create delay, manual transfer, checking or chasing.
For each one, write down:
the trigger
the current owner
the information required
the real finish line
the common exceptions
the consequence of a mistake
one outcome and one quality guardrail
The strongest candidate is not automatically the one with the most manual hours. Prefer the one where the pain matters, the boundary is understandable and a small test can produce trustworthy evidence.
That is a better first AI workflow than a broad assistant looking for a purpose.
Start with one piece of work. Keep the human decision where it matters. Measure the complete outcome. Expand only when the workflow has earned more responsibility.
Related reading
How should an SME keep an AI workflow reliable after launch?
Clear ownership, visible exceptions, tested changes and a safe fallback for live AI workflows.
MEASUREMENTHow should an SME measure the value of an AI workflow?
A scorecard for outcomes, quality, human effort and full operating cost.