What not to automate with AI: a checklist for small firms, from a studio that builds automations

What not to automate with AI

The short list of what not to automate with AI: promises to customers, decisions about people, and the conversations your clients pay you for. Hold off on any process nobody has mapped, and on work that depends on knowledge nobody wrote down. Skip tasks that happen a few times a month. Automate the frequent, document-heavy work that is left, and keep a person on whatever the machine is unsure about.

This is for owners of firms of 5 to 50 people who get a new AI pitch every week. One week it is an AI receptionist, the next an agent that promises to answer your email and give you your evenings back.

We build business automations and AI agents for a living, so we have every reason to tell you to automate more. We think the most useful thing we can do in a first conversation is the opposite: shorten the list. Here it is, with our verdict on each item.

Why the useful list is the short one

Every seller of an AI tool has a reason to make your list longer. Each item they cross off is revenue they have turned down, so almost nobody crosses anything off.

That leaves you in an awkward spot. You can't easily judge whether an AI tool will work in your firm; if you could, you wouldn't need the seller. When a buyer can't judge what they are buying, the seller gets paid for being impressive. The gap shows up a few months later, when the tool meets real work.

There's one thing you can check. You can't verify a supplier's technical judgement, but you can verify that they talked you out of spending money. A supplier who shortens your list has shown you whose interest they serve.

When we take a company through this, we start by sitting with the team. We want to see how the work actually runs, rather than how the process document says it runs. Then every task gets a verdict.

Three verdicts: automate, assist or keep

Automate means the machine does the task end to end. A person sees only the exceptions: the items it flagged as uncertain, waiting in a review queue.

Assist means the machine does the slow part, such as reading, sorting or drafting, and a person decides and sends. Assist pays when checking is much faster than doing. Say a reply takes five minutes to write and thirty seconds to check: you keep most of the time.

If checking means re-reading every source document, you've kept the work and added a step.

Keep means a person does the task. The machine may fetch information for them, but it doesn't act.

There is a fourth answer, and it comes before the other three: not yet. Some work can't get any verdict until the process is mapped or the knowledge is written down.

Promises, people and the conversations clients pay for

Promises to customers

A price, a refund, a delivery date or a contract term becomes your promise the moment it is sent. Your customer holds you to it, whichever tool wrote it.

In 2024 a Canadian tribunal ordered Air Canada to pay a customer C$812.02. Its website chatbot had described a bereavement refund that the airline's policy didn't allow. Air Canada argued the chatbot was responsible for its own words. The tribunal found it was simply part of Air Canada's website (Moffatt v Air Canada, 2024 BCCRT 149).

Verdict: assist. Let AI draft the quote or the reply, and have a person read and send it. When we build an AI agent, we put a human approval step wherever a mistake would be expensive. A wrong promise to a customer is exactly that kind of mistake.

Decisions about people

Hiring, pay, discipline, credit: anything that changes someone's life should stay with a person. The law points the same way.

In the UK, the Data (Use and Access) Act 2025 sets the rules. Since 5 February 2026, a significant decision about someone made with no meaningful human involvement needs safeguards. You must tell them, let them make their case, offer a human review and let them contest it.

In the EU, the AI Act classes AI used to recruit or evaluate staff as high-risk. Those rules apply from 2 December 2027.

AI can still help around the edges, for example by booking interviews. Verdict: keep. We aren't lawyers, so if a tool you are pitched touches these decisions, ask yours.

The conversations clients pay you for

In an owner-run firm, the relationship is often the product. The complaint, the bad news about a deadline, the call with your biggest client: these are the moments a client decides whether to stay. A polished automated reply at that moment tells them where they rank.

Verdict: keep. Use AI to prepare the person instead, with the account history and what was promised last time, so the conversation starts informed.

Processes and knowledge that aren't ready

A process nobody has mapped

Automation pointed at a bad process copies it faster, errors included. If three people do a task three different ways, an automation will pick one way and fail on the other two.

So we map the process end to end first, including the parts nobody documented. Then we delete the steps nobody needs, and automate last, on whatever survives. The deleted steps cost nothing to run. Verdict: not yet.

Judgement that depends on what isn't written down

An automation can only judge what it can see. It can't see the site visit, last Tuesday's phone call, or the fact that one client always pays late but always pays. When that knowledge lives in one person's head, a tool asked to make the call will guess or decline. Either way it looks like bad judgement, and the real gap is missing information.

Verdict: keep, until the knowledge is written down somewhere a model can read it. That can be the better first project: a company second brain, one knowledge base your team and your tools both query.

Work where a wrong answer looks like a right one

Some mistakes are loud: the automation stops and someone sees an error. Quiet mistakes do the damage: an invoice filed under the wrong client, or a supplier email marked as handled when nobody answered it. The output looks finished, so nobody checks, and the first person to notice is a customer.

Verdict: assist, until the automation reports its own failures. Ours send anything uncertain to a review queue, and failures surface to a person rather than disappearing quietly.

Work where AI costs more than it saves

Tasks that happen a few times a month

Do the arithmetic before the demo. Say a task takes 20 minutes and happens eight times a month: under three hours in total.

Even a small build takes days, and every automation needs upkeep. The other software changes, the model changes, and the person who understood it leaves. Three hours a month rarely pays for that. Verdict: keep.

Now compare a task that takes six minutes and happens 40 times a day. That is four hours of someone's time every working day, and it deserves a serious look.

Rules that don't need AI

If you can write the rule in one sentence, you don't need a model to follow it. Orders over a set amount go to the owner; invoices from these ten suppliers go to this folder. A plain rule in software you already use does this for less money, and it never has an off day.

AI earns its place in the messy middle. It can read an email a customer wrote however they liked, or pull an order number out of a scanned PDF. Ask whether the tool you are being pitched uses AI for a job a rule could do. The same question separates a real agent from a workflow with an AI step, which we cover in AI agents vs workflows. Verdict: automate, without AI.

What is worth automating in a small firm

Three things make a task worth automating. You need all three.

1. The hours are expensive, because the people doing the work could be doing something worth more.
2. The work is dense with documents and email: reading, sorting, copying from one system into another. If one of those systems is a pile of spreadsheets, read the signs you've outgrown them first.
3. You, the owner, feel the cost. In a firm where nobody feels it personally, nobody pushes the change through.

Drop any one and the numbers stop working. Cheap hours can't pay for the build, and judgement-heavy work resists automation whatever the budget.

Typical candidates include pulling order details out of emails and chasing clients for missing documents. Matching supplier invoices to purchase orders is another, and so is the weekly report nobody enjoys writing.

Six questions for the next AI pitch

Take these to the next demo. A good supplier answers each one plainly. A vague answer is an answer too.

1. What does it do when it is unsure? You want to hear that it stops and hands the case to a person, with the context attached.
2. How will I find out when it gets something wrong, and how fast?
3. What will a person still check, and how many hours a week will that take?
4. What does it cost to run each month beyond the licence, counting model usage, upkeep and my team's time? Our guide to what an AI agent costs to build and run lists every line to ask about.
5. If I stop paying, do I keep my data and the setup?
6. What would you tell me not to automate?

If the answer to the last question is "nothing", you've learned something about the supplier.

What we chose not to build

We would rather show restraint than claim it. Our public work has one example.

For Shkodra Shore, a six-bedroom villa on Lake Shkodra, we didn't build a booking engine. The site presents the property, then hands the guest to the booking channel the owners already use. There is no second system to keep in step. It isn't an AI example, but it is the same decision: the smaller system was the better one.

When this list is wrong

Treat the list as a default. Three things change it.

Volume changes the maths. If you grow, a monthly task can become a daily one, and then even careful checking pays for itself.

Evidence can move a task from assist to automate. If a person has approved months of drafts without changing a word, the check may no longer earn its time. Move one step at a time, and keep the review queue.

The tools change. This list is what we would tell a client in September 2026. Some assist verdicts will become automate as models get better at saying when they are unsure, and we will update this post when that happens.

You may also not need a supplier at all. If the job is drafting your own emails, a ChatGPT or Claude subscription will do. If your accounting software or CRM already has a built-in automation for the task, use it before you pay anyone to build one.

And if you can't yet name the task that is slow, you don't need us yet either. When you are unsure, ask for an honest read before you commit.

Start with a week of counting

This week, pick the one task your team complains about most. For five working days, count how often it happens, how long it takes and what goes wrong. Then give it a verdict: automate, assist, keep or not yet.

If it comes out as automate or assist and it happens several times a day, you've found your first project. If it doesn't, you've saved the cost of finding out the expensive way.

Questions

Do we have to tell customers they are talking to an AI?

If you serve customers in the EU, yes, unless it is obvious. From 2 August 2026, the AI Act requires a chatbot that talks to people to say it is an AI. The duty sits mainly with whoever builds the system, so ask your vendor how their tool does it. We would tell customers anyway, because a customer who finds out later trusts the rest of your service less.

Can't we do all of this with ChatGPT ourselves?

For drafting, often yes. That is the assist verdict with you as the reviewer, and it costs a subscription. It stops working when the output has to move between systems without someone copying and pasting. It also stops working when five people need the same behaviour, or when you need a record of what the machine did and why. That is when a built automation starts to pay for itself.

How we wrote this: drafted with AI, then checked and edited by the people who build our projects. Figures come from our own work or from the linked sources.