The 90-Day AI Deployment Playbook for SMBs

Most AI plans are either a slide deck with no first step or a tool purchase with no destination. This is neither. It is a 90-day sequence to take a small or mid-sized business from no AI to one workflow running in production, measured, and trusted — the foundation everything else compounds on. The phases are deliberately unglamorous, because the boring path is the one that finishes.

Days 1–14: Audit and select

Do not buy anything in the first two weeks. Map how work actually flows through the business with short interviews of the people closest to it. Produce one ranked list: every recurring workflow scored by how much time it consumes and how rule-based it is. The intersection of those two numbers is your first candidate.

• Select one workflow: repetitive, rule-based, owns a real metric, tolerates a human in the loop, has documented or documentable inputs.
• Reject the exciting candidates. The first deployment should be too repetitive to be interesting and too important to ignore.
• Capture the baseline now — the metric the workflow already produces, measured for the two weeks before anything changes.

Days 15–30: Document the playbook

Write the workflow down as if training a new hire on day one: what inputs arrive, what "done" looks like, where the judgment calls are, and the rule when judgment is required. If it cannot be documented, that is the finding for this phase — fix the process before adding AI. Most of an AI project's value is created here, before any software is selected.

Days 31–45: Define boundaries and choose the tool

Only now does the tool decision happen, and it is faster because you know exactly what it must do well. Before it runs, write the boundary: what it may read, what it may write, what it must escalate. Default to read-mostly with a human approving anything irreversible.

Days 46–60: Pilot on real work, supervised

Run the workflow on live cases with a person reviewing every output before it leaves the building. Expect the first week to be rough — that is the prompt and the playbook being tuned against reality, not evidence of failure. Log every case. The pilot is finished when the outputs stop surprising the reviewer.

Days 61–75: Reduce supervision deliberately

Move from reviewing every output to reviewing a sample, but only on the parts that have earned it. Widen permissions only where the read-mostly version has proven trustworthy. This is a graduated release, not a switch. Anything irreversible keeps its human approval regardless of how well the pilot went.

Days 76–90: Measure, translate, decide

Re-baseline against the day-1 number. Then translate the result into currency before the leadership conversation: not "20% faster" but "three more files per week per person, which is the hire we did not need this quarter." Owners act on dollars, not percentages. With one workflow proven, the second is a repeat of this playbook, not a new act of faith.

The three outcomes at day 90

1. 1One workflow in production, measured against a real baseline, that the team trusts.
2. 2A repeatable playbook you can point at the next workflow without starting over.
3. 3A dollar figure for the value created, defensible in front of anyone who asks.

This is the exact sequence Integra Consulting runs with SMBs. The deliverable at day 90 is not a roadmap or a recommendation. It is a working system, a method you keep, and a number you can defend — the only starting point from which AI adoption actually compounds.