The short answer for buyers
The best automation candidates are frequent, rules-led and measurable, with usable data and reversible mistakes. The worst candidates depend on missing context, affect legal or employment rights, carry severe error costs or cannot be checked economically.
A use case is worth automating when it passes five tests
A use case is worth automating when it passes five tests: volume, value, repeatability, data readiness and risk control. Volume means the task happens often enough to matter. Value means the task costs real money, slows revenue, creates errors, or consumes senior attention. Repeatability means there is a recognisable pattern. Data readiness means the information needed to do the work exists and can be accessed safely. Risk control means a wrong output can be reviewed, corrected or contained.
The mistake many businesses make is starting with the most exciting AI idea rather than the most economically sensible one. A chatbot that answers every customer question sounds attractive. But if support volume is low, knowledge is messy and one bad answer creates reputational risk, it may be a poor first project. A less glamorous workflow, such as drafting quote follow-ups from CRM notes, may pay back faster because the inputs are clearer and the value is easier to measure.
Use this blunt rule: if you cannot describe the current manual process, estimate the monthly cost of that process, name the person who owns it, and define what good output looks like, do not automate it yet. Discovery work is fine. A production build is premature.
Source: ONS Business insights and impact on the UK economy, 2 October 2025.
Score every candidate use case out of 25
Score every candidate use case out of 25. Give 0 to 5 points for each of these areas:
- Frequency: How often does the task happen?
- Cost or value: How much time, money or revenue is attached to it?
- Clarity: Are the inputs, rules and desired outputs clear?
- Data access: Can the AI safely reach the documents, systems or records it needs?
- Risk: Can a human review the output before harm is done?
A score of 20 to 25 is a strong automation candidate. Build a pilot, measure it, and consider production if the numbers hold. A score of 15 to 19 is worth a short discovery or prototype. A score of 10 to 14 usually needs process clean-up first. Anything below 10 should stay manual for now unless there is a strategic reason to learn.
The fastest returns usually come from tasks that are repetitive
The fastest returns usually come from tasks that are repetitive, text-heavy and already semi-structured. Good examples include:
- Sales follow-up: Drafting personalised emails from call notes and CRM history.
- Customer support triage: Classifying queries, suggesting replies and escalating risky cases.
- Document processing: Extracting key facts from invoices, forms, contracts or applications.
- Management reporting: Turning system exports into weekly summaries and exception lists.
- Compliance checks: Flagging missing fields, policy gaps or unusual patterns for human review.
These use cases work because the AI is not being asked to run the company. It is being asked to prepare work, summarise information, spot likely issues, and speed up human decisions. That is where most SMEs should start.
By contrast, fully autonomous customer-facing agents, pricing decisions, hiring decisions and regulated advice need much stronger controls. They may still be worth automating, but they should not be treated as quick wins. In the UK, data protection, fairness, accountability and auditability matter. The ICO provides an AI and data protection risk toolkit designed to help organisations reduce risks to individuals' rights and freedoms caused by their AI systems. If your use case touches personal data, special category data, vulnerable customers or meaningful decisions about people, build governance into the business case from day one.
Source: ICO AI and data protection risk toolkit.
The strongest first project is normally one where the AI prepares, checks or routes work, while a person remains accountable for the final decision. That balance gives you measurable efficiency without pretending the model is infallible.
Before approving an AI automation project
Before approving an AI automation project, calculate six numbers. First, the current monthly volume. Second, the average handling time. Third, the fully loaded hourly cost of the people doing the work. Fourth, the expected percentage reduction in manual effort. Fifth, the cost to build and operate the automation. Sixth, the cost of errors, including rework, refunds, churn, complaints and compliance risk.
That may be worthwhile. But only if the screening rules are fair, documented and checked. If the data is biased, the role requirements are unclear, or the tool makes unexplained recommendations, the risk may outweigh the saving. In a people-related workflow, the business case needs to include governance, not just time saved.
Source: DSIT AI Adoption Research.
Not every automation needs AI
Not every automation needs AI. If the task is rule-based, predictable and does not require judgement, use simpler automation first. Zapier, Make, Microsoft Power Automate, CRM workflows, spreadsheet scripts and database rules are often cheaper, easier to maintain and more predictable than an AI agent.
Use rule-based automation when the instruction looks like this: if X happens, do Y. Send a Slack alert when a deal moves stage. Create an invoice when an order is approved. Add a tag when a form field equals a certain value. These are not AI problems. They are workflow problems.
Use AI when the instruction looks more like this: read this messy information, understand what it means, classify it, summarise it, draft a response, compare it to a policy, or suggest the next best action. AI earns its place when language, ambiguity or judgement is involved.
A simple internal rule helps: automate with rules where the process is deterministic, assist with AI where the process is interpretive, and only delegate to AI where the risk is low or the oversight is strong. That framework prevents expensive over-engineering.
This framework does not apply cleanly when the project is mainly strategic research
This framework does not apply cleanly when the project is mainly strategic research, product experimentation or capability building. Sometimes a business funds an AI prototype because it wants to learn, not because the first version will pay back immediately. That can be valid, but call it what it is. It is research and development, not a guaranteed ROI project.
Finally, do not use this framework to justify replacing human judgement where trust is the product. If clients pay you for expert advice, relationship quality, negotiation, safeguarding or accountability, AI should support the professional, not impersonate them. The most valuable automation may be the one that gives your people more time for the parts of the job customers actually care about.
If you want a wider view on project risks before choosing a use case, read Why Do Most AI Projects Fail?. The short version is simple: bad use case selection is one of the fastest ways to waste AI budget.
Related buying questions
See how this works in practice in How can AI help reduce missed appointments and last-minute cancellations?.
See how this works in practice in Can AI help me forecast whether my team has capacity to take on more work?.
Is This Right for You?
This approach is right for you if you have several AI ideas but no clear way to rank them. It is especially useful for UK SMEs where budget, staff time and operational risk matter more than impressive demos.
It is not right for you if you are trying to automate a broken process before anyone has agreed how that process should work. In that case, fix the process first. AI will only make a bad workflow faster, louder and harder to manage.
