Which process to automate first

The question arrives reliably around minute twenty of every audit: "where do we start?" Almost always the expectation is that we'll point at the most complex process, the one that would land well in a board meeting. Almost always we answer the opposite.

The first automated process isn't there to impress anyone. It does two things: it proves the method works inside that particular company, and it teaches everyone how working with an agent actually feels. If it also saves hours, better — but that isn't the main criterion.

The four criteria

During the audit every candidate process is scored on four axes. No elaborate scoring needed: an honest yes or no will do.

1. Can it be measured today? If nobody knows how many hours it costs now, no saving can be demonstrated later. A process without a baseline is a process people will argue about for months.
2. Does it have volume? Something that happens three times a year doesn't repay the work of automating it, however annoying it is.
3. Is the data reachable? If the information lives in a PDF, a mailbox or a system with an API, it can be done. If it lives in someone's head, it has to be written down first.
4. Is the person who does it today on board? An agent built against the will of whoever knows the process won't survive week three.

A process you can't measure beforehand is one you can't defend afterwards.

Why it usually ends up being quoting

In manufacturing the first candidate is almost always the same: reading incoming requests for quotation and preparing the offer. Not because it's glamorous, but because it clears all four criteria almost on its own.

• The volume is there: dozens of requests a week.
• Measuring is easy: time how long one person takes to produce one today.
• The data sits in emails, supplier PDFs and price lists, all reachable.
• Whoever prepares them today is the first to be glad to hand it over.

The typical outcome isn't "the computer writes the quotes". It's that a structured draft arrives in three minutes instead of two hours, and a person checks it, corrects what needs correcting and signs it. Response time to the customer drops from days to hours. That's where the commercial difference shows up, not in the hours saved.

What not to put first

Three categories that, from experience, are better left to the second or third round:

• Anything touching the customer directly without human review. Not out of technical timidity: because a mistake there costs reputation, and on a first project internal trust is still fragile.
• Processes crossing five departments. Every organisational boundary adds a negotiation, and the first project needs to close quickly.
• Strategic decisions. AI is good at preparing the material a decision is made on, much less so at making it.

The second process gets chosen far faster than the first, because by then the company has learned to recognise what works. Which is exactly why the first one deserves the care.