How to run an AI opportunity audit in a mid-market company

Start with records, not tools

Most AI audits begin with a list of tools and a survey asking managers where AI could help. That produces a wish list. Nobody can price a wish list, so nothing gets built.

An audit that leads to a system starts from the other end. It begins with the things your company already produces: an intake form, an order, a claim, an invoice, a support ticket. Each one has a life. It arrives, it gets touched, it waits, it moves, and eventually it is done. The cost of the workflow lives in that life.

Follow ten records end to end

Pick one workflow that matters to revenue or to the month-end close. Pull ten recent records at random, not the clean ones. Sit with the people who handled them and walk through each record step by step: who saw it, what they did, which system they did it in, and what they waited for before the next step.

Write it down as a timeline, one row per touch. For each row you want four things.

• Who touched it, and in which system
• What they re-typed, copied or looked up
• How long it sat before the next touch, and why
• Who owned the wait, if anyone

The rows that matter most are the ones that say waiting. Waiting on a callback. Waiting on a signature. Waiting for someone to notice it is theirs.

Count touches, waits and owners

After ten records you will have a pattern. Most workflows have three or four touches that add nothing but transcription, and one or two waits with no owner. Those unowned waits are where the days go. The record exists. Nobody is responsible for it moving.

This is also where you find missing information. A phone call captures half of what the next step needs, so the next step calls again. Every re-contact is a wait with a cost.

Put a price on each wait

Use your own numbers. For each touch, the wage of the person doing it, the minutes it takes, and how many times a month it happens. For each wait, what the delay costs: a lost lead, a late invoice, spoiled inventory, an audit finding. Multiply by monthly volume.

The rule is that every number in the audit is one the company confirms. If the operations lead cannot look at a line and say yes, that is what it costs us, the line comes out. An audit built on confirmed numbers survives the CFO. One built on benchmarks does not.

Pick the first system

Rank the workflows by confirmed monthly cost. Then rank them again by how many systems and people the fix depends on. The first thing to build is high on the first list and low on the second: expensive, and mostly inside one team's control.

Resist the temptation to fix everything at once. One workflow, built well and used daily, teaches you more about the company than a roadmap for twelve.

What the readout contains

• The traced timelines, with the touches and waits marked
• The confirmed cost per workflow, with the math shown
• The first system to build, what it connects to, and who reviews its output
• What it should cost, and what the company should expect to see change in the first quarter

Two or three weeks is enough for one department. Longer audits are usually surveys wearing an audit's clothes.

What to avoid

• Tool hunts: starting from a product and looking for a place to put it
• Surveys: asking people where AI could help instead of watching the work
• Strategy decks: a roadmap with no traced record behind it
• Benchmarks: industry averages standing in for your own numbers