What AI can and cannot be trusted with inside a business function

Sort the steps, not the job

The question is never whether AI can do finance, or sales, or quality. Those are jobs. A job is fifty steps in a trench coat. Some of those steps are clerical and repeat all day. Some of them are a judgment call that decides whether a customer stays.

So take one function and write down its steps in the order they happen. Then ask three questions of each step in turn. Each answer puts the step in one pile or the other. The two piles are what you build from.

The three questions

• Does this step repeat? A step that runs forty times a week can be watched and improved. A step that runs twice a year cannot.
• Is the rule written down anywhere? In a policy, a contract, a price list, a checklist. If the rule only lives in one person's head, write it down first. Do not automate it.
• Can a person check the answer faster than they could make it? This is the one most firms skip. A draft that takes longer to check than to write is not a saving. It is a second job.

A step that answers yes three times is safe to build. A step that answers no to any one of them is work for a person, for now.

Where it earns its keep today

The safe pile looks much the same in very different companies. That is why a firm can work by function rather than by industry.

• Reading a document and pulling out the fields that matter. The source line stays beside each one.
• Matching two lists that should agree, and showing only the rows that do not.
• Drafting a reply, a quote or a summary from records that already exist. A person sends it.
• Checking work against a written policy, and saying which clause it failed.
• Watching a queue and telling a person which item needs them now, and why.

Where it does not belong

The second pile is shorter and matters more. These are the steps where a wrong answer is costly. It may also be slow to notice, or it lands on somebody outside the company.

• Anything that leaves the company unread: a price, a promise, a filing.
• A call that turns on facts nobody wrote down, such as what this client will accept.
• A decision the person who signs it has to answer for.
• A step whose rule is still argued about inside the company. Settle that first.
• Work where no one can say afterwards whether the answer was right.

The approval is part of the build

The line between the piles is not a wall. Most useful systems cross it. The crossing is where the design work goes. The system does the reading, the matching and the drafting. Then it stops and asks a person for the call it is not allowed to make.

That pause is worth building well. The person should get the draft, the evidence behind it, and the one thing the system was unsure about. It should reach them in a place they already look. If approving takes ten seconds, it will happen. If it takes ten minutes, people route around it, and the system quietly stops being used.

How to run the sort in a week

• Pick one function and one of its workflows. Not the whole department.
• List the steps with the people who do them, in their own words.
• Mark each step against the three questions. Say who would check the output.
• Count how often each safe step runs. Price an hour of that work.
• Build the top one. Leave the rest on the list, where you can see them.

The list is worth keeping. A step that fails the second question this year often passes it next year, once somebody writes the rule down.