Which business tasks should you automate with AI first?

Most businesses don't have a shortage of things AI automation could do. They have the opposite problem. Everything looks like a candidate, a dozen tools promise to handle all of it, and the real fear is picking the wrong thing: spending money on something nobody uses, or worse, letting a bot say something daft to a customer.

The good news is that the first choice doesn't need to be clever. It needs to be boring, frequent and safe. Here's how I'd pick it.

Why the first AI automation matters more than the rest

Your first project sets the tone. If it works, people trust the next one. If it goes badly, "we tried AI and it didn't work" becomes the story for a year.

It's also where most projects go wrong. MIT's NANDA research, as reported by Virtualization Review, attributes the high failure rate of enterprise AI tools to brittle workflows, weak contextual learning and misalignment with day-to-day operations. None of those are about the model being too weak. They're about picking a process that doesn't fit, or bolting AI onto a process nobody understood properly.

Score every candidate task on four things

Write down every repetitive task you can think of. Then give each one a rough score on four questions:

• Frequency. How often does it happen? Daily beats monthly.
• Time. How long does it take a person each time? Multiply by frequency and you have hours per month.
• Error cost. If the automation gets it wrong, what does that cost you? A mislabelled internal ticket is cheap. A wrong price quoted to a customer isn't.
• Reversibility. Can a mistake be spotted and undone before it does damage? A draft can be fixed. A sent payment usually can't.

The first two tell you the value. The last two tell you the risk. Your best first target is high on frequency and time, and low on error cost, with mistakes that are easy to reverse.

A quick way to use this: give each question a score from 1 to 3, add up frequency and time for value, add up error cost and irreversibility for risk, and sort by value minus risk. It's crude on purpose. The aim is to have the argument about your tasks on one page, not to build a spreadsheet model.

Start internal, then move towards customers

The safest place for your first AI automation is somewhere only your team sees the output. If it makes a mistake, a colleague catches it and you learn something. If a customer sees the same mistake, you've got an apology to write.

So a sensible order is:

1. Internal work where a person reviews everything (drafts, summaries, sorting).
2. Internal work that runs on its own, with spot checks.
3. Customer-facing work where a person approves before anything is sent.
4. Customer-facing work that runs on its own, with clear limits and a way to hand over to a human.

Plenty of businesses should stop at step three for a long time, and that's fine.

Good first targets for AI automation in a small business

When people are asked which processes to automate first, the same list keeps coming up: follow-up emails, data entry, invoice approvals, scheduling and customer support FAQs. Here's how I'd rank the usual suspects as first projects:

• Triage. Sorting incoming emails, support requests or enquiries into categories and routing them to the right person. High volume, low risk, and a person still handles the actual reply.
• Data entry. Pulling fields out of forms, invoices or emails and putting them into your system, with anything uncertain flagged for review. This is close to the bulk-upload metadata extraction I built into SubTrada, the construction procurement platform where Claude reads uploaded documents.
• Follow-ups. Drafting reminder and chase emails for someone to approve and send. The drafting takes the time. The approving takes seconds.
• Internal Q&A. Letting staff ask questions of your own policies, price lists or process documents. If an answer is wrong, it's wrong to a colleague, not a customer.
• Summarising documents. Meeting notes, long email threads, contracts for a first read. A person still makes the decisions.

What these have in common: they happen often, they eat real hours, and a person stays in the loop at the point where a mistake would matter.

Poor first targets

Some tasks look automatable on paper and are still a bad place to start:

• Anything irreversible. Payments, refunds, deleting records, cancelling orders.
• Anything legally binding. Contracts, formal quotes, terms offered to customers, compliance filings.
• Anything where one mistake is public. Social media replies, reviews, press statements.
• Anything rare. If it happens twice a month, the build will never pay for itself.

These aren't off limits forever. They're just not where you learn. Once you've got a working system with logging, review steps and a record of how often it's right, you can have a sensible conversation about giving it more responsibility.

Measure a baseline before you build

This is the step most people skip, and it's the one that lets you prove the project worked.

Before anything is built, write down:

• How many times the task happens per week
• How long it takes on average
• How often it currently goes wrong, and what that costs
• Who does it now, and what they'd do with the time back

A fortnight of rough tallies is enough. Without it, you'll be judging the automation on feel, and feel tends to follow whoever shouts loudest in the first week.

Pick one process, then expand

The temptation is to automate five things at once. Don't. One Indie Hackers analysis of seven autonomous AI agents concluded that they work best on a very specific workflow, and that trying to use them as a general "AI worker" usually creates more complexity than value.

That matches the way I'd scope any first build. One process, one owner, one baseline, one clear measure of success. Get that running properly, watch it for a few weeks, then use what you learned to pick the second one. The second project is always cheaper, because the logging, the review steps and the connections to your systems already exist.

It's also worth checking whether you need AI at all. A lot of "AI automation" is really a well-built workflow with a model doing one step in the middle. I've written about the difference between agents, automations and chatbots if you want to test which one your task needs.

A short checklist before you commit

• The task happens at least weekly, and takes real time each time
• A mistake is cheap and easy to reverse
• Only your team sees the output, at least at first
• You've measured how it works today
• One named person owns the process and will judge the result
• You know what the automation must never do on its own

If a task ticks all six, it's a good first project. If you've got a shortlist and can't choose between them, send me the list and I'll tell you which one I'd start with.