AI automation for business:
where to start
Written for owners and operators, not engineers. A method for finding the work in your business that is genuinely worth automating, in what order, and how to avoid the expensive detours.
The hard part was never the technology. It's choosing where to point it.
Most businesses we speak to are not short of AI ideas. They are short of a way to decide between them. Someone read about agents, someone else saw a competitor's announcement, and the tool everyone already pays for just added an AI button nobody trusts.
The result is a familiar pattern: a pilot that impressed everyone in a demo, quietly abandoned four months later because it never fitted how the work actually happens. That is not an AI failure. It is a prioritisation failure.
This guide is the exercise we run in a first workshop, written so you can do a first pass yourself. It takes an afternoon and it will tell you whether you need to call anyone at all.
Four steps to a decision you can defend.
Follow the work, not the technology
Start from what your team does, never from what AI can do. Spend a morning with the people closest to the work and write down every task that repeats. Four questions get you most of the way.
- What do you do more than ten times a week. Volume is the single strongest signal. Anything happening daily is a candidate; anything happening monthly almost never pays back.
- Where do you type the same information twice. Re-entering data from one system into another is the most reliably automatable work in any business, and the easiest to measure.
- What do you dislike most about the week. People are accurate about this. Tedium is a good proxy for repetitive and rules-driven, which is exactly the territory where AI is dependable.
- What is a customer waiting for right now. Queues reveal bottlenecks. Work that sits waiting for a person to read something is where automation shows up as a visible business result, not just an internal saving.
Three tests each candidate has to pass
Now be ruthless with the list. Most items will fail one of these, and finding that out on paper costs nothing.
- The volume test. Multiply how long the task takes by how often it happens in a month. If the answer is not hours, stop. There is no automation cheap enough to justify saving twenty minutes a week.
- The rules test. Ask someone to explain how they decide. If the explanation is a set of conditions, good. If every sentence is 'it depends on the context', that is judgment work and should stay with a person.
- The input test. AI is strong at reading documents, forms, emails and records. It is much weaker at anything requiring a conversation, a relationship, or physical presence. Check what the task actually consumes.
Sequence by payback, not by excitement
With a shortlist that passes the tests, the order matters more than people expect. The first project sets whether anyone in your business trusts the second one.
- One workflow, one team, one clearly measurable outcome
- Something where you already know what a good result looks like
- Work where a person can easily check the AI's output
- A task whose owner actively wants it to happen
- Anything spanning several departments at once
- Customer-facing automation before you have internal confidence
- Work where being wrong carries legal or financial consequences
- Projects whose main sponsor is enthusiasm rather than a number
The four detours that waste the most money
None of these are exotic. They are the same four patterns behind almost every abandoned pilot we get called in to look at.
- Automating a broken process. If nobody can describe how the work happens today, automation will lock the confusion in and make it faster. Map and fix first, then automate.
- Starting with the most impressive idea. The ambitious cross-department project is the one most likely to stall, and its failure poisons the appetite for the boring project that would have worked.
- No human in the loop. Full automation is rarely the right target. Decide which cases run through untouched and which get flagged to a person, and start conservative.
- Buying a tool instead of solving the problem. Another subscription adds a login and leaves the manual copying between systems exactly where it was. The value is usually in the integration, not the tool.
What a sensible first project looks like
Whether you build it in-house or bring someone in, the shape of a good first project is the same.
Small enough that failure is survivable
One workflow, a few weeks, a fixed price. If a disappointing outcome would be a serious problem for you, the scope is too big for a first attempt.
Measured against what happens today
Before you start, write down the current numbers: how long the task takes, how often it is wrong, how long customers wait. Without that baseline you cannot tell success from enthusiasm.
Tested on your real history, not a demo
Run the automation over cases where you already know the right answer and compare. This is the single most useful hour in any AI project.
Owned by someone who does the work
Automation that belongs to a project sponsor rather than a practitioner drifts out of use. The person whose day it improves should be the one signing it off.
Which work automates well today.
| Type of work | How well AI handles it | Keep a person involved? |
|---|---|---|
| Reading documents and extracting fields | Strong | Review flagged cases |
| Classifying and routing incoming email | Strong | Spot checks |
| Drafting a reply from a template plus context | Strong | Always edit before sending |
| Checking records between two systems | Strong | Review mismatches only |
| Summarising long documents for a decision | Good | The decision stays human |
| Judgment calls on unusual cases | Weak | Human decides |
| Anything needing a conversation or negotiation | Weak | Human throughout |
Bring us your shortlist.
If you have run this exercise and want a second opinion, the first call is free and there is no pitch. We will tell you which item we would start with, roughly what it would take, and whether you need help at all. See how we build automation, or get in touch.