How to pick the first process

Which process should you automate first?

Start with the pile that already has a name. 'Tier-1 tickets about billing dates.' 'First-pass CV screens for this role family.' 'Lead notes that should exist before a human calls.' Those sentences have a verb and a boundary. 'Make us more AI-native' does not. I will not spend day 1 translating a slogan into a workflow. That is how you burn the week before anyone writes a connection.

I score a candidate process on four things: weekly hours, how alike the items are, whether a miss is recoverable, and whether the tools already hold the data. High hours and high sameness win. A recoverable miss means a person can catch it without a legal event. Data already in the tools means we are not waiting on a migration.

At SmartComply the work that paid off was the same document, again. Pull the fields. Flag the gaps. Leave the weird cases for a person. That is a first process. 'Understand all of compliance' is not.

What makes a process a bad first pick?

Anything that needs a new source of truth. Anything that is political. Anything where two senior people disagree on the correct outcome and both are in the room. Anything you cannot show me ten recent examples of. If you cannot produce last week's pile, we are inventing the work, and invented work does not survive contact with Monday.

Brand-new products are a bad first pick. So is 'replace the team.' An AI employee takes a slice. If the slice is not obvious, you do not have a process. You have a department. Departments do not ship in seven days.

• No recent examples, only a story about how it usually goes.
• The 'process' changes by who is on shift.
• A miss would be a regulatory event, and you have no reviewer.
• The data lives in inboxes nobody can search.

How do you measure hours before you build?

Ask for a week of the actual queue, not a guess. Count items. Ask how long a clean one takes and how long a messy one takes. Multiply. People undercount because they do not include the search, the re-reading, and the 'let me just check with.' I include those. That is the work.

If the team has no numbers, we sit with the queue for an hour and sample. That is still day 1, not a six-week baseline project. You need enough to know the week is worth $5,000 or $7,500. You do not need a consulting model of the firm.

Should the first job be customer-facing?

It can be, if the exits are boring. A support employee that answers the questions you already answer, and hands off the rest, is a good first job. A support employee that 'handles customers' is not a job. It is a wish. The difference is whether you can list the intents and the hard gates before we build.

Hiring screens and sales follow-up sit in the middle. They touch people, but the first output is a shortlist or a note, not a promise. I like those as first jobs when the team already has a bar. If they do not have a bar, the model will invent one, and you will spend week two arguing with a ranking you never agreed to.

What do you do with the leftover messy cases?

You name them on day 1 and write them on the map on days 2 and 3. They are not a later phase. They are the reason the AI employee is allowed to exist. If 10 to 15 percent of cases should escalate, that is a healthy first process, not a broken one. If everything must escalate, you picked theater. If nothing may escalate, you picked a landmine.

The leftover cases need an owner with a name. Not 'the team.' A person who will see a working state and a decision brief. If you cannot name that person, pick a different process. An orphan queue is how these builds become another thing the founder checks at midnight.

How do you get the team to name the real pile?

Ask what they would hand a competent new hire on week one. Not the exciting work. The work they apologise for. That is usually the process. People protect their interesting work and give away the copying. Good. Give the copying to the AI employee.

Then ask what they would be embarrassed to still be doing in six months. If the answers match, we have a first process. If they point at a transformation programme, I bring them back to the queue. The week is $7,500 because it is a week, not because it will reorganise the company. Several workflows at $15,000+ come after the first pile is gone and nobody misses it.