Which Process to Automate First with AI

The first process you automate with AI should not be the one that hurts most or looks most impressive. It should have an owner, repeat with little variation, and fail in ways you can see in time. Those three filters turn the first attempt from a bet into a decision.

Carlos Andrés Ramírez · 2026-10-09

The first process you automate with AI should not be the one that hurts most or looks most impressive. It should have an owner, repeat with little variation, and fail in ways you can see in time. Those three filters turn the first attempt from a bet into a decision.

Start with how the choice is usually made. A committee gathers a list of candidate processes and ranks them by pain: the one that eats the most hours, draws the most complaints, costs the most. It is a logical criterion, and it is the one that most often sends a first attempt to failure. The processes that hurt most tend to be the most tangled, with the most exceptions and the most dependencies across functions.

The short answer is that the first process is chosen for what it lets you learn, not for what it saves. A first attempt that goes well, even a small one, fixes how the work is supervised, how it is measured and who answers for it. One that goes badly on a critical process teaches the same lessons, but with leadership watching and less patience for a second try.

The usual mistake

Why does the first AI automation so often fail?

Think of supplier invoice reconciliation. It hurts: six people do it, it takes days and it produces discrepancies. It looks like the ideal candidate. But every supplier invoices in its own format, every country has its own tax rule, and half the discrepancies get settled with a phone call. An agent trying to cover all of that from day one meets hundreds of special cases, and each one needs somebody to decide.

The usual result is a pilot that handles the easy cases and forces a manual check on the rest, so the team ends up working harder than before. The conclusion people then draw is that AI is not ready yet. What was not ready was the choice of process.

The framework

A good first process passes three filters, not one

Three filters, in this order. The first two rule out most of what does not fit, and the third decides among what remains.

1. It has an owner who knows it and can change it

One named person understands the process end to end and has the authority to modify it. Without that person, nobody decides what to do when the agent hits an odd case, and the project sits waiting for a meeting.

2. It repeats with little variation

The same type of case arrives many times a week, and the rules that resolve it can be written down. Volume with bounded variation gives the agent enough to work with and gives you enough evidence to measure. A process with few cases, all different, gives neither.

3. Its errors show up in time and cost little

If the agent gets it wrong, someone spots it the same day and the damage is fixed with no consequence for a customer, a regulator or a reported figure. It is the filter most often forgotten and the one that protects most: it lets you be wrong on the first attempt without it being expensive.

The mechanism

The third filter decides whether there is a second process

Imagine two candidates. The first is sorting incoming emails in a customer service team: hundreds a week, a handful of categories, and if the agent sorts one wrong, a supervisor sees it when reviewing the queue that same afternoon. The second is approving high-value refunds: few cases, a lot of judgment, and an error discovered only once the customer has already complained.

Both can be automated, but only the first lets you learn safely. When errors surface fast, the team adjusts the rules, counts how many corrections it needs and builds confidence on its own data. When errors surface late, every adjustment is made blind, and the organisation responds with the only tool it has: adding human reviews until the saving disappears.

The first automated process is not chosen for how much it saves, but for how cheaply you can be wrong in it.

BECOME

The consequence

A modest first process opens the way to the hard ones

Choosing a modest process often feels like giving up on ambition. In practice the opposite happens. The first case that works leaves three things the next one can use: a way to supervise the agent's work, an agreed measure of what working means, and someone in the organisation who already knows how to answer for it. The big processes get tackled afterwards, with those three pieces already in place.

Take your list of candidates and drop any process without someone who knows it end to end. Of the ones left, keep the one that receives the most similar cases each week and whose errors can be seen and corrected the same day. That is your first process, even if it is not the one that hurts most.

Let's pick the first process worth automating