The 90-Day AI Rollout Plan for a 50-Person Business

Most AI rollouts we see in Phoenix start the same way. Somebody on the leadership team reads something, buys 50 licenses, sends one all-staff email with a link, and waits. Ninety days later usage sits around eleven percent, the licenses renew on autopilot, and nobody can name a single hour that got saved.

The failure is almost never the technology. It is that the rollout had no sequence. Below is the sequence we run for clients, compressed into thirteen weeks. It assumes a company of roughly 30 to 80 people with a Microsoft 365 or Google Workspace tenant and no dedicated data science staff.

Weeks 1 through 3: find out what is already happening

Before you buy anything, find out what your people are already using. In every assessment we have run this year, the answer has been "more than the owner thought." Personal ChatGPT accounts, browser extensions that summarize email, a marketing contractor running client copy through a tool nobody vetted.

Three concrete tasks:

1. Pull your identity logs and list every third-party app with an OAuth grant against your tenant. In Microsoft 365 this lives under Enterprise Applications. You are looking for anything with read access to mail, files, or calendar.
2. Run an anonymous five-question survey asking staff which AI tools they use and what for. Promise no consequences and mean it. Punishing honest answers here buys you six months of hidden usage.
3. Write down the three most expensive repetitive tasks in the business. Not the most annoying, the most expensive. Hours times loaded labor rate.

You now have a starting inventory and a target list. Everything after this points at those three tasks.

Weeks 4 through 6: write the rules before you hand out the keys

Policy first, licenses second. This order matters because the first employee who pastes a client contract into a public model does it in week five, not week twelve.

Your acceptable use policy needs to answer four questions in language a non-technical employee can follow. Which tools are approved. What categories of information may never be entered into any of them. Who reviews AI-generated work before it reaches a customer. What happens when someone finds a better tool and wants it added.

Keep it to two pages. A twelve-page policy is a policy nobody reads, which is legally the same as not having one when a regulator asks.

At the same time, decide your data boundary. For most small businesses that means committing to tools with a contractual no-training clause and enterprise data handling, then blocking or discouraging the consumer versions of the same products. The consumer and enterprise versions of these tools look identical and behave very differently with your data.

Weeks 7 through 9: one department, one workflow

Do not roll out company-wide. Pick the single department where the work is most repetitive and the output is most reviewable. In professional services that is usually proposal or intake work. In construction it is submittals and RFI responses. In healthcare administration it is prior authorization paperwork.

Give that department real licenses, two hours of live training, and a named person who owns the experiment. Then measure two things weekly: how many hours the workflow used to take, and how many it takes now. Write the numbers down in the same place every week. Memory is a terrible measurement instrument, and by week twelve everyone will remember it going better than it did.

Expect the first two weeks to be slower than the old way. That is normal and it is why company-wide rollouts fail. A pilot group absorbs the learning curve. Fifty people hitting the learning curve at once produces a mutiny.

Weeks 10 through 13: prove it, then expand

At the end of the pilot you should be able to state a real number. Something like: our intake team spent 62 hours a month building client summaries, they now spend 24, and the quality complaints have not increased.

If you cannot state that number, do not expand. Buy fewer licenses, not more. There is no shame in concluding that a workflow was not a good fit, and there is a lot of expense in pretending otherwise for a year.

If you can state the number, expand to the next department and repeat the same pilot structure. Two departments running well beats eight running badly.

The part nobody schedules

Somewhere in month two, an AI tool will produce something confidently wrong and a client will see it. Decide now who is accountable for review, and make that a named role rather than a general expectation. The businesses that get burned are not the ones whose AI made a mistake. They are the ones who could not say who was supposed to catch it.

Get the runbook

We built the week-by-week version of this as a fillable runbook: the app inventory sheet, the staff survey questions, the two-page policy skeleton, the pilot measurement log, and the go or no-go criteria for expansion.

It is free, requires no sales call, and can be completed independently. AEGITz also runs the engagement for Phoenix businesses as a fixed-scope project.

Frequently asked questions

How long does it take to roll out AI in a small business?

Plan on 13 weeks for a first workflow. Three weeks of discovery, three of policy work, three of piloting one department, and four to measure and decide. Company-wide rollouts that skip the pilot stage typically stall at around 11 percent adoption because everyone hits the learning curve at once.

Should we roll out AI to the whole company at once?

No. Pick the single department where the work is most repetitive and the output is most reviewable, then expand only after you can state a measured result. A pilot group absorbs the learning curve. Fifty people hitting it simultaneously produces a mutiny.

What should we do first when implementing AI?

Find out what your staff already use. In every assessment we run, the answer is more than the owner expected: personal accounts, browser extensions, and contractor tools nobody vetted. Pull your OAuth grants and run an anonymous survey with a genuine promise of no consequences.

Do we need an AI policy before buying licenses?

Yes. The first employee who pastes a client contract into a public model does it in week five, not week twelve. A two-page policy covering approved tools, prohibited data, human review, and a request path is enough for most small businesses.

Related reading

• Writing an AI Acceptable Use Policy Your Employees Will Actually Follow
• Copilot, ChatGPT Enterprise, or Gemini: What an Arizona SMB Should Actually Buy
• The Questions to Ask an AI Vendor Before You Sign
• What Your Data Has to Look Like Before AI Is Worth Buying
• Measuring AI ROI Without Lying to Yourself