The useful question is not “what can AI agents do?” — they can do a great deal. It is “which of those repays a five-figure build at the volume my business actually has?”
Every use case below carries the volume threshold at which it starts paying for itself. Hit the threshold and the payback is usually faster than people expect.
The arithmetic you apply to all of them
Before any use case, run this:
Annual cost of the process = hours per week × 52 × fully loaded hourly costFully loaded means salary plus employer’s NI, pension and overhead — typically 1.25–1.35× salary. Someone on £28,000 costs roughly £18–£19 per hour fully loaded.
A process consuming 12 hours a week costs about £11,800 a year.
Against a £30,000 build plus £500/month running: £36,000 in year one, £6,000 after. Payback on labour alone: roughly 2.6 years.
That is the test. If the process consumes four hours a week, the arithmetic does not work and no amount of enthusiasm fixes it.
Where it pays
1. Supplier invoice processing
Threshold: 300+ invoices/month
Reads invoices in any layout, extracts totals, VAT, dates and references, matches to purchase orders, flags variances outside tolerance, posts the clean ones.
Why it works: high volume, varied formats that break conventional automation, clear right answer, errors caught by the PO match.
Typical result: 70–85% processed without a human. The rest escalate with a reason attached.
2. Inbound email triage and response
Threshold: 200+ emails/week needing more than a template
Reads the email, identifies the customer and what they are asking, retrieves the relevant information from your systems, drafts a reply with the actual answer. Sends or queues for approval depending on the boundaries you set.
Why it works: customers write in unstructured prose, which is exactly where automation fails and agents succeed.
Start in draft mode. A human approves every reply for the first month. Confidence builds on evidence.
3. Quote and tender preparation
Threshold: 15+ quotes/month with real variation
Reads an enquiry or tender document, extracts requirements, pulls pricing and availability from your systems, assembles a draft, flags anything requiring a judgement call.
Why it works: the input is unstructured, the assembly is repetitive, and the value per quote is high enough that even modest time savings repay quickly.
4. Goods-in and delivery note reconciliation
Threshold: 200+ deliveries/month
Matches delivery notes against orders, identifies shortages and substitutions, updates stock, raises queries with suppliers.
Why it works: paper-heavy, format-varied, and errors here are expensive later.
5. Document-heavy compliance checking
Threshold: 100+ documents/month
Checks certificates, insurance documents, licences or qualifications against requirements. Flags expiries, gaps and mismatches.
Why it works: tedious, error-prone when done by humans under time pressure, and the cost of a miss is high. Common in construction, logistics, recruitment and facilities.
6. Lead qualification and enrichment
Threshold: 150+ enquiries/month
Researches an inbound enquiry across public sources, assesses fit against your criteria, enriches the CRM record, routes or rejects.
Why it works: sales time spent on unqualified leads is the most expensive waste in most SMEs.
Caution: check your lawful basis for processing personal data. UK GDPR applies.
7. Stock reordering and exception management
Threshold: 500+ SKUs with variable demand
Monitors stock, considers lead times, seasonality and open orders, drafts purchase orders, escalates anything unusual.
Why it works: a judgement task done many times daily, where a human applies rules of thumb inconsistently.
8. Internal knowledge retrieval
Threshold: a team of 20+ asking the same questions repeatedly
Answers staff questions from your documentation, policies and historical records — with citations back to the source.
Why it works: removes a tax on senior people’s time. Requires citation so answers can be verified, or it becomes a confident misinformation machine.
9. Data cleanup and deduplication
Threshold: 20,000+ records with known quality problems
Identifies duplicates, inconsistent formats and contradictions, proposes merges for human approval.
Why it works: genuinely hard to script, genuinely tedious for humans.
Note: this is often the prerequisite project. An agent reading bad data produces confident bad output.
What each one looks like in month three
Volume thresholds tell you whether to start. These are what the good ones actually settle at, so expectations are calibrated before rather than after.
Supplier invoice processing
Settles at: 70–85% processed without a human. The remainder escalates with a reason attached — usually a price variance outside tolerance, an unmatched PO, or a document too poor to read.
What surprises people: the exception rate does not go to zero and should not. An agent that escalates an ambiguous invoice is working correctly.
Where the saving actually comes from: not only the processing time but the chasing. Exceptions surface the day the invoice arrives rather than at month-end, when the context has gone.
Inbound email triage
Settles at: 65–80% drafted accurately. Most businesses keep a human approving sends for the first few months, then release the clearest categories.
What surprises people: how many emails need information from two or three systems to answer. That is why automation never handled them and why the agent earns its cost.
Quote and tender preparation
Settles at: a usable draft in most cases, with a human adjusting commercial judgement.
What surprises people: the real value is turnaround time rather than labour. Quoting in two hours instead of two days changes win rates, and that is worth more than the hours saved.
Goods-in reconciliation
Settles at: 80–90% matched automatically.
What surprises people: the discrepancies it surfaces in the first month. Most businesses discover shortages and substitutions they had been absorbing without noticing.
Compliance document checking
Settles at: 85–95%, because the task is narrow and the right answer is checkable.
What surprises people: expiry tracking turns out to be worth more than the checking. Knowing what lapses next month is the thing nobody was doing reliably.
How to sequence several of them
Most businesses with one candidate process have three. Doing them in the right order matters.
First: the one with the highest annual cost and the cleanest data. Not the most interesting. You are buying evidence as much as capability, and a first project that works makes the second one fundable.
Second: something that shares the plumbing. If the first agent integrated with your accounting system, the second one that uses the same integration is substantially cheaper. Sequence by shared infrastructure, not by department.
Third: the one that needed the first two to exist. Reporting and exception dashboards make sense once there is data flowing.
What not to do: run three pilots simultaneously. They compete for the same scarce resource — your team’s attention — and none gets the shadow-running discipline it needs.
A realistic schedule: first agent live at month three, second at month six, third at month nine. Each funded by the one before.
Where it usually does not pay
Social media content generation. Low value, high volume of mediocre output, and the brand cost of generic content is real. A cheap subscription tool does this adequately.
“An AI chatbot for our website.” Most SME sites do not have the traffic. A good FAQ page and a contact form outperform a chatbot that annoys people below a few thousand monthly visitors.
Meeting notes and summaries. Genuinely useful and already solved by off-the-shelf products at £10–£20/user/month. Do not commission a build.
Anything with fewer than 50 occurrences a month. The arithmetic does not work. It will not start working because the technology is interesting.
The prerequisite project nobody budgets for
A pattern worth naming, because it catches a significant share of businesses at the feasibility stage.
The agent works. The data does not.
Feasibility runs, accuracy comes back at 61%, and the failures are not the model misunderstanding the task — they are the model being given contradictory information. The same supplier under three names. A product code that means two different things depending on which system wrote it. A date field containing four formats.
What happens next, usually: the project is judged a failure and shelved. That is the wrong conclusion.
The right conclusion: data cleanup is the project, and it comes first. It typically takes four to eight weeks, your own team can do most of it, and it has value whether or not the agent follows — reporting improves, duplicates stop confusing people, and the migration you will eventually do for some other reason gets cheaper.
How to avoid being surprised: export a sample and look at it before commissioning anything. Count duplicates. Check null rates on the fields the agent would depend on. Ten minutes of looking prevents a £10,000 feasibility sprint reaching a conclusion you could have had for free.
What all the good ones share
Look at the nine that pay:
- High volume — hundreds of occurrences a month
- Unstructured input — documents, emails, varied formats
- Judgement, but not expertise — the kind a trained person makes in seconds
- A checkable right answer — a PO to match against, a rule to apply
- Errors are recoverable — nothing irreversible
- Currently consuming real hours — measurable today
If your candidate process has fewer than four of these, be sceptical. If it has all six, it is probably worth a feasibility sprint.
Sector patterns we see
Manufacturing and wholesale — goods-in reconciliation, stock exception management, supplier invoice processing.
Professional services — document review, compliance checking, proposal assembly.
Construction and trades — certificate and compliance tracking, subcontractor documentation, quote assembly.
Logistics — delivery reconciliation, exception handling, customer status enquiries.
Healthcare administration — document triage and records workflow. Special category data — take proper advice first.
Recruitment — document checking and candidate enrichment. Caution on automated decisions about individuals under Article 22.
How to pick yours
- List every process consuming more than five hours a week.
- Score each against the six shared characteristics above.
- Calculate the annual cost of the top three.
- Pick the one with the highest cost and the most characteristics — not the most interesting one.
- Run a feasibility sprint on it. Here is how to structure that.
The most common mistake is choosing the process that sounds most impressive rather than the one with the strongest arithmetic.
Working out your own candidate list
An hour’s work that replaces a lot of speculation.
Step 1 — list every process consuming more than five hours a week. Across the whole business, not just the obvious department. Ask people; do not estimate on their behalf.
Step 2 — score each against the six characteristics. High volume. Unstructured input. Judgement but not expertise. A checkable right answer. Recoverable errors. Measurable hours today.
Step 3 — calculate the annual cost of the top three. Hours per week × 52 × fully loaded hourly cost.
Step 4 — check the data for each. Where does the information live, and how consistent is it? A high-scoring process sitting on inconsistent data is a data project first.
Step 5 — pick the highest cost with the cleanest data, not the highest score. A slightly lower-scoring process with good data will deliver faster and prove the case.
Step 6 — run a feasibility sprint on it. £6,000–£12,000, two to three weeks, a measured accuracy number on your real records. How to structure it.
The common mistake at step 5: choosing the process that would be most impressive to automate. Impressive and economic are different, and only one of them survives a budget review.
What it means for the people doing the work
The question everyone is thinking about, and worth addressing directly rather than leaving to rumour.
What usually happens: an agent removes the repetitive, judgement-light portion of a role rather than the role. Someone still handles exceptions, and the exceptions are the part that needed a person anyway.
What that looks like in practice: a person spending 22 hours a week on invoice entry spends six on exceptions and sixteen on work that was being neglected — supplier queries, cost checking, chasing credits. Most businesses find the freed capacity absorbed immediately by things that were not getting done.
Where headcount genuinely is the objective, say so openly. People work it out, and discovering it late poisons the project — including the cooperation you need from the people who know where the edge cases are.
What makes adoption work: involving the people doing the job during feasibility. They supply the awkward real-world examples, they spot where the agent is subtly wrong, and they become the people who trust the output rather than quietly working around it.
What makes it fail: a system designed from the sponsor’s description of the process, imposed on people who were not consulted and who know fifteen things about it that were never written down.
Frequently asked questions
What can AI agents actually do for a small business?
Process documents in varied formats, triage and answer inbound email, prepare quotes, reconcile deliveries, check compliance documents, qualify leads, manage stock exceptions, answer internal questions from your documentation, and clean up data. All require meaningful volume to repay a build.
What volume do I need to justify an AI agent?
As a rule of thumb, a process consuming at least 10–12 hours a week. Below that, payback on a five-figure build exceeds three years on labour alone.
What is the best first AI project for an SME?
Usually supplier invoice processing or email triage — high volume, unstructured input, checkable answers, recoverable errors. Pick based on which consumes the most hours in your business, not which sounds most impressive.
Is AI worth it for a business with 20 staff?
It depends entirely on process volume, not headcount. A 20-person business processing 800 invoices a month has a strong case. A 200-person business with no high-volume document process may not.
What should I not use AI agents for?
Social media content, low-traffic website chatbots, meeting notes (buy a product), and anything occurring fewer than 50 times a month.
How do I calculate the payback?
Hours per week × 52 × fully loaded hourly cost gives the annual cost. Compare against build cost plus annual running cost. Two to three years on labour alone is typical; faster if error reduction or freed capacity has real value.
Written by Jas Singh, Founder. Last updated 8 October 2026. Volume thresholds are based on typical payback arithmetic and should be tested against your own figures.
Related:
Agentic AI systems · How to pilot an AI agent · What it costs
