An AI adoption playbook: from the first workflow to the next department

Step 1: map the department with the people who run it

Start with one department, not the whole company. Sit with the team and trace how work arrives, which systems it touches and where a person retypes, checks or chases something. One question sorts most of it: how many systems does one transaction touch, and where does someone key it in again?

The output is a list of workflows with the client's own numbers on each: volume, minutes per item and the loaded cost of the person doing it. Numbers the business states itself are numbers its leaders will believe later.

Step 2: pick the first workflow

Rank the list by hours per week times loaded cost, then weigh the cost of an error. A workflow where each item is an order, a claim or a payment is worth more than its hours suggest. The strongest candidates share a shape: information arrives in a format you don't control, and a person reviews it, validates it, chases what's missing and resolves it.

Then apply a buildability screen. The inputs must be capturable, the shape of a correct output must be known, and there must be a ground truth to test against. At USCAPE, order entry passed all three: forwarded emails, a FileMaker order with a known structure, and past orders already in FileMaker to check against. It went from kickoff to production in five weeks.

Step 3: write down the agent's authority

Before building, list every action the agent can take and put each in one of four classes: automatic, prepared for a person to send, internal record only, or an external write that always needs approval. Sending to a customer, committing an order and moving money never happen without a person.

Start every switch off and enable actions one at a time as the reviewers gain confidence. TIE decides endorsements automatically only for change types where the rules settle the case, and refers the rest.

Step 4: make evals the release gate

Build a test set from real historical cases with known right answers, and run it before release and after every change to a prompt, model or rule. Then sample live output each week and add every disagreement to the set.

At go-live, compare the first production records field by field against the same work keyed by hand. Differences are expected; they show the defaults and judgement calls the manual process applied without anyone writing them down, and they become the next work list.

Step 5: train reviewers and champions

The person who did the work by hand becomes its reviewer, and their checklist becomes the agent's rules. Train them on the review queue, on what the agent can and can't decide, and on how a rejection with a written reason improves the next draft.

Pick a champion in each team who knows the tooling well enough to help colleagues. At Avalara, 40 GTM leaders were trained in person, and optional platform training reached a 65% completion rate. Training is part of the build plan, with its own owner and its own numbers.

Step 6: measure adoption and impact separately

Adoption tells you whether people use the system: share of items going through the agent, review times, override rates. Impact tells you whether it matters: hours saved per year, cycle time, decisions made without a person, errors caught before they reach a customer.

Avalara scored every initiative on hours saved per year, which gave comparable numbers across 8 departments and set the order of work. Its governance scanner also checked 720 workflows and found issues in 84% of them before they reached production. Report both kinds of number to the workflow's owner every month.

Step 7: take the next department

• Reuse the intake. Every new idea comes in through the same form and gets the same impact score.
• Reuse the plumbing. Integrations, review queues and logging built for the first department serve the next one.
• Bring the first department's numbers. Reported impact makes the case for the second department faster than any forecast.
• Keep ownership in house. Code, data and credentials stay in your name, and each department's team is trained to run and extend its agents.
• Hold the pace. Pilots live in weeks, whole departments AI native within months.