The question we get most often is not whether to use AI. It is where to start, and the honest answer is that the starting point matters more than the tool. Businesses that begin by buying licences for everyone tend to end up with a subscription nobody uses. Businesses that begin with one workflow, measured, tend to end up with a second workflow.
This is the 90-day sequence we use with small teams. It assumes five to fifty people, no in-house developer, and no appetite for a science project. It is deliberately slow at the start and fast at the end.
Key Takeaway
Ninety days is long enough to ship one real automation, train the people who use it, and have the data to decide on the next one. It is short enough that nobody loses interest.
Before day one: pick the workflow, not the tool
Statistics Canada's 2026 survey of AI use shows where businesses actually get value: data analytics (36.6 percent of adopters), text analytics (34.5 percent), and virtual agents or chatbots (28.2 percent). In plain terms, that is reporting, documents, and customer questions. Your first workflow almost certainly lives in one of those three.
The selection rule is simple. List the five tasks that eat the most senior time each week. Cross off anything that requires judgment a client is paying for. From what is left, pick the one that is most repetitive and most annoying. Typical winners: first drafts of proposals, weekly reporting, inbound email triage, meeting notes and follow-ups, or job-posting and screening. If you want a structured version of this exercise, it is the first hour of an AI readiness assessment.
Weeks 1 to 2: map it and measure it
Write down, step by step, how the workflow happens today. Who starts it, what information they need, where it comes from, what they produce, who checks it. Then measure the baseline: how many times a week, how long each time, and where the errors show up. This feels bureaucratic and it is the single most skipped step. Without a baseline you will never know whether the project worked, and you will have no argument for the next one.
Two things usually surface here. The information the workflow needs is scattered across tools, which tells you what has to be connected. And the workflow has an unofficial version that is different from the official one, which tells you what to actually automate.
Weeks 3 to 4: choose the tool and write the data rules
For most first projects the choice is between the general assistants (ChatGPT, Claude, Gemini) on a business plan, and a small automation platform (Make, n8n, or Zapier) that connects those assistants to the systems you already use. Start with the assistant. Add the automation layer only when the workflow needs to run without a person pressing a button. We compared the three assistants for business use in ChatGPT vs Claude vs Gemini.
Pay for the business tier, not the consumer one, and write down why. On the business plans, both OpenAI and Anthropic state that customer content is not used to train their models by default; on consumer accounts that is a setting you have to manage. That distinction is the foundation of your data rules.
Source: OpenAI: Enterprise privacy (business data is not used for training by default) and Anthropic Commercial Terms of Service, Customer Content
The data rules themselves fit on one page. What can go into the tool: internal documents, your own marketing, anonymized examples. What cannot: client personal information, anything covered by a confidentiality agreement, financial account details, health information. Who approves exceptions. This is not legal advice and it does not replace a privacy review; it is the operating rule that lets your team use the tools without guessing. If your business handles regulated information, involve whoever owns compliance before week five.
Weeks 5 to 8: build the first version, with a person in the loop
Now build. For a drafting workflow, this means a saved prompt or a custom assistant loaded with your best past examples and your house style, producing a first draft that a person edits. For a triage workflow, it means the assistant reading inbound messages, sorting them, and drafting replies that a person approves. For reporting, it means an automation pulling from your systems and the assistant writing the summary.
The design principle for the first version is that a human stays in the loop at the point of judgment. The AI drafts; a person decides. This is not caution for its own sake. It is what makes the rollout safe enough to start this month instead of next year, and it produces the corrections that make version two better. This is the shape of most of the automation builds we do, and the first one rarely needs to take more than two to three weeks.
The rollout that stalls Licences for everyone on day one No workflow chosen No baseline measured AI output goes straight out the door Training is a shared video link Nobody owns it after launch
The rollout that sticks One workflow, mapped and timed Business-tier tool with written data rules AI drafts, a person decides Half-day training for the people who use it Measured at 30 and 90 days One owner, one next project
Weeks 9 to 12: train, measure, decide
Train the people who touch the workflow, in their context, on their tasks. A half-day session on the actual prompts and the actual approval step does more than a week of general AI literacy. In British Columbia, employer-sponsored training can be reimbursed at 80 percent of eligible costs up to $10,000 per participant per year through the B.C. Employer Training Grant, and other provinces run similar programs; check the current rules before you book anything.
Source: WorkBC: B.C. Employer Training Grant, 2026/27
Then measure against the baseline you took in week two. Hours per week, turnaround time, error rate, and one qualitative question to the people doing the work: would you go back? If the answer is no, you have your case for project two. If the answer is yes, you have learned something cheap, and the most common reason is that the workflow was chosen for being interesting rather than for being painful.
At day 90, decide one of three things: expand this workflow to the rest of the team, move to the next workflow on your list, or stop. All three are legitimate outcomes. The failure mode is not stopping; it is drifting, where the tool stays paid for and half used.
The guardrails that let you move fast
- One owner. Someone, even at four hours a week, who is accountable for the rollout and the next step.
- Written data rules. One page, shared with everyone, reviewed by whoever owns compliance.
- A person at the judgment point. For the first version of anything client-facing.
- A baseline and a 90-day number. Without them, you are running on vibes.
- A way to switch tools. Keep your prompts, examples, and process documents in your own files, not locked inside one vendor. The tool landscape is changing quarterly.
What this costs
For a small team, a realistic first quarter is a few hundred dollars a month in business-tier subscriptions, a modest automation platform plan if you need one, and either your own time or a short engagement to map, build, and train. If you want help, our engagements are scoped to exactly this shape: a short assessment to choose the workflow, a build, and team training so it sticks. We are honest on the first call about whether you need us at all; some teams can run this plan themselves with our small business AI tools guide and a free afternoon.
Want a second opinion on which workflow to start with? Book a free 30-minute discovery call. We will map the first 90 days with you, no pitch attached.
Get Started