Level 1: Individual Use

One person, working alone, without telling anyone.

It’s a Tuesday, and your salesperson is writing a follow-up to a prospect who went quiet three weeks ago. She opens a second tab, pastes in the thread, asks for a few versions, picks one, rewrites the middle because the first draft sounded like a press release. Four minutes instead of twelve. She will not mention this to anyone, and if you asked her about it directly, she’d probably downplay it.

Meanwhile, someone in ops is running a vendor contract through the same tool, and someone in marketing is doing captions. None of them know about each other. Some of it is going into free consumer accounts along with customer names, which is the part that should bother you.

On paper, this costs nothing, which is exactly why it sits undisturbed for years.

You can test for it in about a minute. Ask your leadership team who at the company uses AI and what for. If you get a shrug or an estimate, you’re at Level 1. More than 80 percent of organizations have piloted something by now, so being here is ordinary. Thinking you’re past it when you aren’t is the expensive part.

Getting out of Level 1 starts with finding out what’s already happening and doing it in a way that doesn’t make people assume they’re about to be told to stop.

Level 2: Licensed Adoption

The same activity, now with the company’s name on the invoice.

There are proper accounts with the security settings someone spent a week configuring. There’s a usage policy, two pages, mostly written by legal. There was a training session in March, and roughly half the company showed up.

Time is genuinely being saved here. Your team can feel it. People are somewhat less buried than they were. But no process changed. Every workflow still runs the same steps in the same order as last year, just with a faster person at each step. The benefit is spread thin across sixty people in ten-minute increments, and it doesn’t collect anywhere you can see it.

The cost is per-seat and highly visible, which makes it the first thing questioned when budgets tighten.

The test for Level 2 is whether you can name a number. Not “people say it’s helpful.” An actual metric that moved. If your CFO asked what the AI line bought this year and the honest answer is a general feeling of improvement, you’re here.

A lot of companies stop at this level and conclude AI was overhyped. What they’ve really discovered is that buying software isn’t a strategy.

Level 3: Workflow Redesign

One process rebuilt with AI inside it.

Same Tuesday, 7:15 in the morning, nobody at their desk yet. A form came in overnight. By the time your ops lead opens her laptop, the request has been read and categorized, a draft quote is waiting with line items pulled from current pricing, and the three items that need a judgment call are flagged with a note explaining why they were flagged. She reviews it, changes the delivery window because she knows this client always asks, and approves it. Out the door at 8:04 instead of Thursday afternoon.

Something structural changed here. One process got rebuilt with AI as a component rather than an assistant. A person still approves, and that’s deliberate.

This costs more than licenses, and not in a way that shows up cleanly in a budget. Somebody had to write down how the process actually works, which almost always turns up the fact that three people were doing it three different ways and two of them were wrong. That’s the real work, and it’s where most attempts stall. The MIT researchers found the difference between the successful five percent and everyone else wasn’t the quality of the models. It was whether anyone was willing to do the unglamorous integration work instead of avoiding it.

The test is simple. If you switched the AI off, would this process break, or just slow down? At Level 2 you’d shrug. At Level 3 you’d have a problem on your hands.

Level 4: Connected Operations

Multiple steps running together, across systems, without a person in the middle.

The AI isn’t in a chat window anymore. It reads from your CRM and writes back to it. It matches invoices against purchase orders and only shows you the ones that don’t reconcile. A support ticket comes in and gets resolved end to end, refund processed, record updated, customer notified, unless it hits a condition that sends it to a person. Your team’s work shifts toward defining those conditions, handling the exceptions, and checking the output.

This is where the word “agentic” stops being conference talk. McKinsey’s most recent research puts 23 percent of organizations at the point of scaling an agentic system somewhere in the business, with another 39 percent experimenting. But the scaling is narrow. No more than 10 percent are running agents in any single business function, which tells you adoption is wide and shallow.

The cost here is integration work plus a governance conversation you can’t put off. Security and risk come up as the top barrier organizations report at this stage, reasonably, since the system is now taking actions rather than making suggestions.

You know you’re at Level 4 when something happened in your business today that no employee touched, and you’re comfortable with that, because you know precisely what would have stopped it.

Level 5: Operating Model Change

The shape of the business is different.

You’re serving clients you couldn’t afford to serve before because the cost to deliver dropped. You’ve launched something that didn’t work at your old margins. Roles that existed to carry information between systems are gone, and roles that exist to design and supervise those systems are new. What changed is your competitive position, not just your operating cost.

Achieving this status remains incredibly rare. In fact, a mere 39 percent of enterprises experience any AI-driven impact on their enterprise-level profit, a metric that even factors in organizations positioned at Levels 3 and 4. This milestone represents the cumulative outcome of sustained compounding across Levels 3 and 4 over several years.

Where Most Companies Actually Sit

In our experience, it’s around Levels 1 to 2. Scattered individual use, a few licenses, and an internal description of all this as an AI strategy.

Most of this is simply explained by what happens when the tool gets bought before the process gets written down. It’s evidence that pointing a capable tool at an undocumented process gets you faster confusion.

Moving between any two levels looks roughly the same, and it’s less exciting than most vendors let on. Pick one process. Write down how it really runs, including the parts that only exist in one person’s head. Then decide which steps a machine should handle, which ones a human has to approve, and what would need to be true before you’d trust it without watching.

If you want to work out which level you’re on, we put together a short assessment that walks through it. Most people come out a level lower than they expected, which is usually the useful part.

Quick Reference

Level 1. Individual Use. Ad hoc, undocumented, mostly invisible to leadership.

Level 2. Licensed Adoption. Company accounts and a policy. Time saved, no process change.

Level 3. Workflow Redesign. One process rebuilt around AI with human approval in the loop.

Level 4. Connected Operations. Multi-step execution across systems, exception-based human involvement.

Level 5. Operating Model Change. New economics, new offerings, different org shape.

Sources

• MIT NANDA, The GenAI Divide: State of AI in Business, 2025
• McKinsey, The State of AI, 2025 to 2026
• Structure adapted from Gartner’s AI Maturity Model (Awareness, Active, Operational, Systemic, Transformational)