How should a company start implementing AI?

Most companies I talk to are in one of two places. Either nobody has touched AI because no one knows where to start, or everyone has been experimenting on their own, with five different tools, no rules, and nothing anyone could point to as a result. Both end up in the same spot: a lot of interest and very little to show for it.

I spent more than 25 years in operations, most of it as a COO, and rolling out AI turns out to be a lot like rolling out any change in how work gets done. The technology is the easy part. The hard part is choosing the right work, setting clear rules, and getting people to change habits. This plan is built around that.

Why do most company AI rollouts stall?

The same five problems show up again and again:

• Tools before problems. Someone buys licenses because a competitor did, and then the team goes looking for something to use them on.
• No owner. AI is “everyone's project,” so it's no one's. Experiments never turn into a standard way of working.
• No training. People get a login and a lunch-and-learn. They try it twice, get a mediocre answer, and go back to the old way.
• No baseline. Nobody timed the work before, so nobody can prove it got better, and the budget gets cut at the next review.
• Fear about data. Staff don't know what they're allowed to paste in, so the careful ones avoid it and the careless ones paste in everything.

Every step below is designed to prevent one of these.

Step 1: What business goals should AI serve?

Start with two or three outcomes the leadership team already cares about, stated in business terms. “Respond to every customer inquiry within two hours.” “Close the month in five days instead of twelve.” “Give each salesperson back a day a week for selling.” Not “use AI more.”

Goals like these do two jobs. They tell you which workflows to look at first, and they give you the yardstick you'll use to decide whether the rollout worked. If a proposed AI project doesn't move one of them, it waits.

Step 2: How do you find where AI will help most?

Map the workflows behind each goal: who does what, in which tool, how often and for how long. You're looking for work that is frequent, follows rules someone could write down, takes real time, and has a clear output you can check. Ask each team lead to list their team's recurring tasks with rough times. It takes an hour, and it's the most valuable hour in the whole project.

Then use AI to help you sort the list. This prompt works well with a team lead's notes:

Step 3: What kind of AI does each workflow need?

This is where most companies overspend or underbuild. Not every process needs the same kind of AI. I match every workflow to one of three levels:

1. Level 1: Human Intelligence Tasks (HIT). A person stays in the loop. AI drafts, researches, summarizes or suggests, and a person reviews and approves. The goal is to make the work faster, better and cheaper without changing who's responsible. Best for client communication, proposals, content, hiring and anything that depends on judgment or relationships.
2. Level 2: Full automation (RPA). Software moves data between systems along fixed rules, end to end, with no one touching it. A form comes in, a record is created, an invoice goes out. Best for high-volume, predictable steps like data entry, scheduling, invoicing and routine reporting.
3. Level 3: AI-intelligent workflows (AIIW). AI agents read context, make decisions within limits you set, and hand exceptions to a person. Best for work that varies too much for fixed rules but follows a pattern, like triaging requests, qualifying leads or answering common support questions.

A simple way to choose: if the work needs judgment or touches a relationship, start at Level 1. If it's predictable and high-volume, Level 2. If it varies but you can describe what a good decision looks like, Level 3. Most companies end up with a mix, and many workflows start at Level 1 and move up once the team trusts the results.

Step 4: How should you run an AI pilot?

Pick one team with a willing lead and two or three workflows from your ranked list. Before anything changes, measure the baseline: how long each task takes, how often, and how many errors or rework loops it has today. Then run the pilot for 30 to 60 days.

Building the pilot is the practical work of creating your AI pipeline: placing the right tools with the right people and training them, designing each workflow step by step, writing the prompts and agent instructions around your business and your voice, connecting the systems involved (CRM, email, calendar, documents) so nobody is copying and pasting, and testing before anyone relies on it. Keep the first version simple. A well-written prompt that the whole team uses beats a custom app nobody opens.

At the end, compare against the baseline and decide for each workflow: keep, fix or drop. Dropping one is a good result. It's cheaper to learn that in a pilot than across the whole company.

Step 5: What AI rules does a company need?

Before the pilot starts, write a one-page AI use policy. It doesn't need a lawyer for the first version, but it does need to answer five questions clearly:

• Which tools are approved? Use business or enterprise plans that don't train on your data, and say which ones.
• What data can go in? For example: public and general business information, yes. Customer, employee, health or financial records, only in approved tools with the right settings, or not at all.
• What needs a human's approval? Anything that reaches a customer, moves money, changes records or affects an employee.
• Who owns it? One person who approves new tools and answers questions.
• What happens when something goes wrong? Who to tell, and how to roll back.

Step 6: How do you get employees to actually use AI?

Adoption is where rollouts are won or lost. Four things make the difference:

• Train on their own work. Generic training produces generic results. Show each team how AI handles the three tasks they do every week, using their real documents.
• Name a champion in each team. Someone who's a little ahead, shares what works, and is the first person colleagues ask. It costs a few hours a month and matters more than any tool.
• Make it the default. Put the prompt in the template, the agent in the shared project, the automation in the workflow. If using AI takes extra steps, people won't.
• Be honest about jobs. People hold back when they think they're automating themselves out of a role. Say plainly what the time saved will be used for, like more selling, faster service or fewer late nights, and mean it.

Then build a short feedback loop: a monthly 30-minute check-in where each champion shares what worked, what broke and what to try next.

Step 7: How do you measure AI ROI and decide what's next?

Keep the scorecard simple and tied to the goals from Step 1:

Hours saved per month = (minutes per task before − minutes per task now, including review) × times per month ÷ 60

Value of time saved = hours saved × the loaded hourly cost of the people doing the work

Net monthly return = value of time saved − monthly tool and maintenance costs

Track two more numbers next to it: adoption (how many people in the team use it weekly) and quality (errors, rework or response times compared with the baseline). Time saved with falling quality isn't a win, and great results that only one person uses aren't a rollout.

When a pilot workflow pays off, write it down as the standard way of working, then move to the next team or the next workflow on your list. Plan for regular maintenance too. AI tools change every few months, and a workflow nobody checks will quietly drift.

Should you build AI tools yourself, buy them, or hire someone?

Usually some of each. Level 1 work and simple agents can be built in-house with business plans for tools like Claude or ChatGPT, with no code. Level 2 automation often means an automation platform, set up by someone on your team or a specialist. Custom integrations, internal apps and larger agent systems are where outside developers or automation partners earn their fee. The mistake is hiring builders before you know what to build. Do Steps 1 to 3 first, and every quote you get will be cheaper and more accurate.

What does a 90-day AI rollout look like?

1. Weeks 1 to 3: Strategy. Agree on goals, map workflows with team leads, rank opportunities, match each to a level, and write the AI use policy.
2. Weeks 4 to 10: Pilot. Measure the baseline, set up tools, build two or three workflows with one team, train them, and check in weekly.
3. Weeks 11 to 13: Decide and expand. Compare results with the baseline, keep, fix or drop each workflow, document the winners, and choose the next team.

Frequently asked questions

Not ready for a company-wide rollout yet? Start with the free AI Workforce Blueprint to see how AI assistants can be set up by department, or read which tasks to hand to AI agents first. When you're ready to build, 20 AI Agents That Pay for Themselves has ready-to-paste instructions for each one.