The automation audit, on one page.

Do this for one process, not the whole business. Pick the one your team complains about most. Answer every question honestly. A “no” is not a failure, it is the thing to fix before any AI goes near the work.

Step one: pick the process

• What starts it? A form, an email, a call ending, a date on the calendar.
• What are the steps, in order, as they really happen today?
• Who touches it, and where does it wait for someone?
• How often does it run each week, and how long does one run take?
• What does it cost when it goes wrong, and who notices first?

Step two: the data

• Which fields does this process actually depend on? List them. Usually it is fewer than ten.
• Where does each one live, and which copy is the true one?
• Does every team mean the same thing by each field? Ask sales and finance separately.
• The test: can two people from different teams produce the same number from the same question?

Step three: the rules

• Which decisions in this process always go the same way? Write each one as “if this, then that”.
• Who owns each decision when the rule does not fit?
• What happens when nobody replies? Write the time limit down.
• The test: is every decision that should never vary a written rule, not a judgement call?

Step four: the agent

• Which steps need judgement: reading messy text, summarising a call, drafting a reply, sorting a request into one of a few types?
• What documents should it answer from, and can it say which one it used?
• Below what confidence should it hand the work to a person?
• Do you have ten to twenty real past examples with known good answers, to test it against on every change?
• The test: does every answer cite a source?

Step five: the surface

• Where does your team already work: Slack, the CRM, the inbox?
• Where should the result appear so nobody has to remember to go and look?
• Anything touching money, a contract or a customer: where is the human approval step?
• The test: does it appear where they already work?

Reading your answers

Fix the lowest layer with a “no” first. If the data fails, clean up those few fields and reporting becomes trustworthy on its own. If the rules fail, writing them down removes most of the waiting. Only when data and rules hold is an agent worth building, and by then it is doing the small, genuinely hard part.

The reasoning behind the order is in The GTM automation stack, explained.