How to Choose Which Process to Automate First With AI

My first AI automation in my own business was nothing to show off. It answered inbound sales inquiries. Those emails asked the same things, in the same structure, day after day. Writing a reply took me around 30 minutes. Once it was automated, around 3.

The secret wasn't a brilliant model. The work was simple, it came back constantly, and nobody liked doing it, me included. Those three things together made it succeed. They are also pretty much the reverse of what most companies go for when they start with AI.

Why does the impressive pick keep failing?

The story barely changes from company to company. A senior person gets enthusiastic about AI. There's a meeting. The team goes for the biggest, shiniest process in the room: the complicated one that runs through five departments and would make a nice slide in a board update.

Fast forward six weeks. There's a demo that nearly works. The team has stopped believing in it, though nobody says so out loud. And the verdict lands: AI just isn't there yet for a company like ours.

Except AI was fine. They chose the wrong process.

I've seen this play out over and over. I've done it myself. When I began automating my own business, I wanted to build something clever. What paid off was almost embarrassingly dull.

There's a reason the flashy option fails. A process usually looks impressive because it's messy underneath. Exceptions. Judgment calls. Handoffs between people and systems that nobody ever wrote down properly. On day one, that is precisely the work an AI system has the hardest time with. It's also the kind of project that drags on until people give up on it.

What is a first AI process actually for?

Picking the boring option sounds like settling for less. It's not.

Your first process has a single task: show your team that AI works here. In your company, on your data, inside the way you already work. A plain process can prove that within weeks. A grand one almost never does.

Plain doesn't mean low value either. Take my inquiry replies. They sit at the very front of every deal, which is where revenue starts. Saving 27 minutes a reply is just the mechanism. The value is where that time goes: back into selling and delivering, instead of typing out the same answers yet again.

Not every dull process qualifies, though. Three questions tell you which ones do.

How do you score a candidate?

Score each process from 1 to 5 for each of the three questions below, then add them together. Fifteen is the top.

Does it run often?

Look for work that comes around every day or every week and looks alike each time.

This matters more than it seems, because frequency is how the system gets better. Each run is a chance to spot a mistake and correct it. A quarterly process hands you that chance four times a year. A daily process hands it to you about twenty times a month. Three weeks of daily runs teach you what a quarterly process would need a full year to show.

It also means you see results quickly. Nobody has to sit through a quarter wondering if it paid off.

Is it clear what goes in and what comes out?

Inquiry arrives, quote leaves. A document lands, its data ends up in your system. A call wraps up, the CRM record gets filled in.

Here's the check. Can you explain the process in one sentence, naming what goes in and what comes out? If not, it shouldn't be first. It could become number three or four, after your team has had some practice.

A sharp in and out also makes the work easy to check. When you know what the result should look like, you can tell straight away whether the system nailed it or missed. Vague processes give vague results. Then your meetings turn into arguments about whether it works, instead of sessions on making it better.

Would your team be glad to hand it off?

Nagging clients about late invoices. Retyping data from one tool into the next. Sending the tenth near-identical follow-up of the week.

Give that kind of work to AI and nobody pushes back. You're removing the bit of the job people grumble about. Try passing AI the work people actually like, and the pushback comes fast. Pushback ends AI projects quicker than any technical issue.

This question gets skipped because it isn't technical. Please don't skip it. Whether your first AI process makes it through month one is up to your team.

What do the totals tell you?

Honest cost: parking things means turning down ideas people are excited about. Sometimes the idea is yours. That stings in the meeting. Still, it's far cheaper than spending six weeks on a demo that nearly works.

Which processes tend to win?

Across nearly every kind of company, these ten are the ones I see coming out on top:

All ten pass the same three questions. They come back constantly. Each fits in one sentence with an in and an out. And nobody ever applied for a job hoping to get more of them.

Look at how close they sit to the money, too. Inquiries, quotes and unpaid invoices are revenue and cash flow. Briefings and CRM updates decide how well prepared your people are when they face a client. Strip the admin out of this work and your senior people get time back for selling and delivering.

What stays off the first list?

Some things are missing from that table on purpose: strategy, pricing decisions, creative work, and anything that reaches a customer before a person has checked it.

They come later. Some never. Both are fine.

The three questions show why:

That last row doesn't ban customer-facing work. Answering inquiries is right there in the table above. What makes the difference is a person reviewing the output before it's sent.

Honest cost: your first project won't go near the work that feels most strategic. It will feel small. That's what you trade for something that actually goes live.

Which three traps stall most AI starts?

I run into these all the time.

The cure is identical for all three. Take one process all the way through. Only then start on the next. That really is the entire method.

Honest cost: one at a time feels slow, especially with a long wishlist and everyone pushing their favourite. But a single process that genuinely runs teaches your team more than six pilots that never got finished.

How do you pick yours in ten minutes?

Open a blank note. List the three processes in your company that repeat the most. Rate each one 1 to 5 per question. Add up the numbers. The top score is where you start.

Copy this worksheet straight into your note:

Candidate:                        ____________________
Runs often (1 to 5):              __
Clear in and out (1 to 5):        __
Team glad to hand off (1 to 5):   __
Score:                            __ out of 15

Fill it in three times, once per candidate. Ten minutes, give or take. Whatever lands on top is where everything else begins.

Four rules keep you honest:

• Rate the process as it works right now, not the tidy version you wish you had.
• Needed a second sentence to explain the in and the out? Then that score is low.
• Let the people who do the work answer the hand-off question. Whether they'd miss it is their call, not yours.
• Two processes tied? Let frequency break the tie. More runs, faster lessons.

What does choosing boring really cost you?

About ten minutes. Plus a small dent in your ego.

Your first AI process won't be the one you show off over dinner. What it will do is give a few hours a week back to your team, without anyone making a fuss. Those hours aren't the point by themselves. What counts is how your people spend them: more time with clients, more time selling, less time on work nobody wanted to do anyway.

That's the deal. You get to brag around process number five.

What happens once you have a winner?

You map it. Before an AI system can run a process, you have to describe it so precisely that nothing is left to guess. What comes in. What happens to it. Which decisions get made along the way. What goes out. Most do-it-yourself attempts fall apart right here. It's also why, at Arcgent, we map every process before we build a thing.

Choose the boring one. Map it. Get it live. Then move on to the next.