Start with the work itself#
For a small business, the question is rarely “Which AI tool should we buy?” It is usually “Where are we wasting time, losing consistency or making work harder than it needs to be?”
That is a better place to begin because useful AI work is tied to an existing operational problem. It has an owner, a real input, a clear output and some consequence if it goes wrong. A tool demo may be impressive, but it does not tell you whether the work will survive contact with your inbox, customer records, job sheets and team habits.
Owner-led and operations-led businesses in the UK and Australia often have the same pattern: capable people holding together a growing operation with software, spreadsheets, email, WhatsApp and institutional knowledge. Quotes take too long to prepare. Customer updates depend on one person remembering to send them. Site notes need turning into reports. A manager spends Friday afternoon trying to understand what happened during the week.
Those are the places to look. The aim is to find a workflow that can be shortened, made more consistent or made easier to review without introducing a new headache.
Build a list of operational friction#
Start with a short working session with the people closest to the work. Keep it grounded in the previous week, rather than asking for abstract ideas about AI.
Ask questions such as:
- What gets copied from one system into another?
- What do people rewrite, summarise or format over and over?
- Which tasks wait for one person because they know how to do them?
- Which reports, proposals or job notes are slow to produce?
- What information arrives in a messy form and needs sorting before anyone can act on it?
Write down the workflow, not simply the complaint. “Admin takes too long” is too vague to assess. “After every site visit, an engineer emails rough notes and photos to the office, then an administrator turns them into a client report” is something you can examine.
For each workflow, capture five things:
- Trigger: What starts the task?
- Inputs: What information does someone use?
- Steps: What actually happens between start and finish?
- Output: What does good work look like?
- Owner and risk: Who is accountable, and what happens if the output is wrong?
This small amount of detail exposes whether AI has a sensible role. It also reveals problems that are really about unclear ownership, poor data or an outdated process. Those need fixing first.
Look for work AI handles well#
AI is generally most useful when work involves language, classification, retrieval, drafting or repeated judgement within sensible boundaries. It is less reliable when it is expected to make high-stakes decisions alone, work from incomplete records or invent a process that nobody has properly defined.
A few practical examples:
Enquiries and sales follow-up#
A trade business may receive web forms, emails and calls that need qualifying before somebody prepares a quote. AI can help pull the important details into a consistent summary, flag missing information, draft a reply asking the right questions and create a follow-up task. The sales person still owns the conversation and the price. The system removes some of the chasing and re-keying.
Meeting, site and call notes#
A consultancy, construction firm or service provider may have rough notes from site visits, client calls or handovers. A controlled workflow can turn those notes into a structured summary, action list, client update or first draft of a report. Someone with context checks it before it leaves the business. That review step matters, particularly where safety, contractual terms or technical accuracy are involved.
Internal knowledge and procedures#
Small businesses commonly have good processes trapped in documents, folders or the heads of experienced staff. An internal assistant can help staff find the right SOP, checklist or policy. This works when the source material is current and has an owner.
Use a simple prioritisation test#
A long list of ideas is easy to create. Choosing one is the work.
Score each candidate workflow against a few practical questions. You do not need a complicated model. A simple red, amber and green discussion is often enough.
Frequency: Does this happen often enough to matter?
Time and drag: Does it consume attention, create delays or force skilled people into repetitive admin?
Clarity: Are the inputs and expected outputs clear enough to describe?
Data and access: Can the workflow use information the business is allowed to access, in a secure and sensible way?
Risk: Can a person review the output before it affects a customer, payment, contract, safety decision or employment matter?
Ownership: Is there one person who will make decisions, test the workflow and keep it useful after launch?
A strong first use case tends to be frequent, irritating, bounded and reviewable. It produces something visible, such as a drafted response, a structured summary, a populated record or a short list for a person to check. It does not need a major system replacement to get started.
Be wary of projects described as “an AI assistant for the whole business”. They usually hide several different jobs, sources of information and levels of risk. A better first move is to deal with one recurring bottleneck and learn from it.
Check the process before adding AI#
Some processes should be simplified before any automation is considered. If three people use different quote templates, or nobody agrees what counts as a qualified lead, an AI layer will amplify the inconsistency.
Take the workflow apart first. Remove duplicate steps. Agree the required fields. Decide who approves what. Make sure the source information exists in a form that can be used. In many cases, that piece of operational tidying is valuable on its own.
Set boundaries here. Define what the system may draft, search, classify or suggest, and what needs a person to approve. Decide which data must stay out of public tools. UK businesses should consider UK GDPR obligations. Australian businesses should consider the Privacy Act. Seek appropriate legal or privacy advice where a workflow is sensitive.
Run a small, honest pilot#
A pilot should test a real workflow with real users. Pick a defined period, a small group of users and a clear success condition. For example: can the operations team turn site notes into a consistent draft report that a manager can review faster?
Keep a baseline before changing anything. Record how the workflow is currently handled, where time is spent and where errors or delays appear. Then review examples from the pilot together. Look for quality, exceptions, staff experience and whether the output actually fits into the rest of the process.
Avoid treating usage as proof of value. The useful question is whether the workflow is now easier to run, easier to supervise or more reliable for the customer.
At the end of the pilot, make a straightforward decision: stop it, improve it, or make it part of normal operations. If it moves forward, document the workflow, train the team, nominate an owner and review it periodically. AI systems and the data around them change. A process that worked in a small trial can drift if nobody is responsible for it.
Give people a route to raise concerns#
The people doing the work will spot issues before a dashboard does. Build in a simple way for them to flag bad outputs and suggest changes. That feedback turns a promising test into a process people will actually use.
The useful question to take away#
If you are trying to find the right AI use case, choose one piece of recurring work that is visible, frustrating and contained. Describe it properly. Decide where human judgement stays in the loop. Test it with the people who have to live with it.
That approach is far more likely to produce a useful operational improvement than buying a broad platform and announcing an AI initiative.
If you would like an outside view, Aygent’s AI Opportunity Assessment helps owner-led and operations-led teams map practical workflows, prioritise the ones worth testing and identify the controls needed to introduce them sensibly. There is no obligation to start a wider project. It is a structured way to work out where AI is likely to help, and where it is better left alone for now.