From copilot to colleague: the agentic AI maturity model

"Where are we on the agentic AI maturity curve?"

That's the question the head of ops at a 60-person SaaS company asked me last Thursday, two hours before her CFO check-in. She had everyone on Copilot. She had ChatGPT Enterprise licenses. She had a "Head of AI Strategy" she'd hired in January. And she couldn't answer her own question with anything sharper than "somewhere in the middle."

So I sent her the model we use internally to score where any company actually is. Five stages. One-sentence diagnostic per stage. The specific next move. She screenshotted it and walked into her CFO meeting with a roadmap.

Here it is.

TL;DR

Why most maturity models are useless

You've seen them. Gartner has one. McKinsey has one. Every analyst firm has one. They all say roughly the same thing: "AI-curious," then "AI-piloting," then "AI-integrated," then "AI-native." The stages are abstractions. There's no diagnostic. There's no next move. You read it and you still can't tell where you are.

A useful maturity model has three things at each stage. A one-sentence diagnostic that lets you score yourself in under a minute. A concrete example of what that stage actually looks like in a small business. And a named next move that gets you to the next stage. Without all three, the model is decoration.

So here's a maturity model with all three.

Stage 1: Curious, "we should look into this"

Diagnostic: AI use at the company is individual and informal. Someone on the team uses ChatGPT to draft an email occasionally. Nobody else knows. There's no policy, no tooling, no budget line.

What it looks like: the founder has ChatGPT Plus on their personal account. They use it twice a week. The marketing person uses Grammarly. The sales person tried Lavender once. Nobody's connected the dots.

What stage 1 costs you: mostly the value you're not capturing yet. The opportunity cost is real but invisible. You're not behind because everyone is at the same stage. You're behind once your competitors move to stage 2 and you don't notice.

The move to stage 2: stop being curious. Pick one tool, get team licenses, and make AI usage a team-wide expectation rather than an individual quirk. Boring but real. Cost: ~$20/seat/month for whatever copilot you pick.

Stage 2: Copilot, "AI helps me do my job"

Diagnostic: the team has tools. People use them. But every AI output is consumed by a human and then a human acts on it. The AI is in the chair next to the worker, not in the chair.

What it looks like: Copilot for the engineers. ChatGPT Enterprise for everyone. Maybe Lavender for sales. The marketer drafts a blog post with Claude's help. The engineer writes code with Cursor's help. The sales rep drafts emails with the AI's help. All output flows through a human before anything ships.

What stage 2 costs you: real money on licenses ($30-$60 per seat per month across all the tools) for a productivity bump that's hard to measure. Studies say 15-30% individual productivity gain. The gain is real but it's an individual gain, not a structural one. Your headcount still scales linearly with revenue.

Where most companies are stuck and why: stage 2 is the comfortable stage. The AI is helpful but it's not autonomous, so nobody worries about it doing the wrong thing. The cost is moderate. The optics are good ("we have an AI strategy"). The CEO can say the word "AI" on the board call. Nobody is forced to confront that the org chart hasn't changed.

The trap is mistaking stage 2 for the destination. It isn't. It's a stop on the way.

The move to stage 3: identify one recurring task in the company that runs on a schedule rather than on demand. Sending the weekly newsletter. Drafting the Monday morning standup. Generating the monthly client report. Wire AI to do that task on a schedule with no human prompting it each time. The human still approves the output. But the human doesn't trigger the run.

This is the single biggest perceptual shift in the whole model. The AI stops being something you summon and becomes something that runs.

Stage 3: Assistant, "AI runs work without me starting it"

Diagnostic: at least one piece of work in the company runs on a schedule or in response to an event, without a human kicking it off each time. A human still reviews and ships the output, but the cycle starts on its own.

What it looks like: the weekly client status report drafts itself on Sundays at 4pm and lands in the founder's inbox by Sunday evening. They review it Monday morning, edit if needed, send. Or: every Stripe webhook for a refund request triggers a draft response that lands in the support inbox; the support person reviews and sends. Or: every new lead in the CRM triggers an enrichment pass and a personalized draft email; sales reviews and sends.

What stage 3 costs you: real engineering time to wire the first scheduled cycle (1-2 weeks for the first one, much less for subsequent ones). Modest ongoing model spend (~$50-$200/month per scheduled cycle). The savings show up as time the human gets back. A founder who was spending 2 hours a week on status reports gets that time back. A support person who was triaging 30 refund tickets a week now triages 30 pre-drafted replies and ships them in a third of the time.

Where most companies stall here: they ship one scheduled cycle, it works, they pat themselves on the back, and then they don't ship a second one. The stack underneath is fragile (no memory layer, no orchestrator) and adding a second scheduled cycle introduces drift that nobody catches. We wrote about the 5 layers of an agentic AI stack. Most stage 3 companies are missing layers 2, 3, and 5.

The move to stage 4: give one of those scheduled agents an authority envelope. It doesn't just draft. It ships. Inside a defined boundary. The refund agent now sends the refund reply directly when the request meets criteria (under $100, within 30 days of purchase, no prior dispute). It still escalates the rest. The human stops being in the middle for 80% of the work.

Stage 4: Delegate, "AI ships work without me approving each one"

Diagnostic: at least one agent in the company ships work externally without a human in the loop on every output. There is an authority envelope (a defined scope of what the agent can do without escalating) and an escalation path for anything outside it.

What it looks like: the support agent replies to refund requests under $100 directly. The outreach agent sends up to 25 personalized prospect emails per day on a schedule, with a linter on every draft. The content agent publishes a daily blog post directly to the site after a linter pass, no human approval. The financial controller agent pays recurring vendor invoices under $500 automatically. Each of these has a defined envelope and a tripwire for anything outside it.

What stage 4 costs you: a meaningful upfront build (3-8 weeks for the first agent to reach this level of trust, plus a verification layer underneath all of them). Ongoing cost is mostly model spend (Sonnet for drafting, Haiku for verification) and the engineering time to maintain the envelopes as the business changes. We run our own venture at this stage and our monthly infrastructure cost is under $200.

Where most companies stall here: they ship one delegate-stage agent, it does the wrong thing once (sends a bad email, ships a bad post, replies oddly to a refund), and a senior person pulls the plug instead of fixing the envelope. The fix is almost always tightening the envelope, not removing the agent. But the political cost of one visible mistake is high enough that the agent gets retired.

The move to stage 5: wire agents to each other. The lead-sourcing agent's output becomes the outreach agent's input, with no human reading what came out of the first one before the second one picks it up. Same with content: the blog writer's output triggers the distributor's run automatically. The system stops needing you in the middle.

Stage 5: Colleague, "AI is a coworker the others coordinate with"

Diagnostic: agents talk to each other. One agent's output is another agent's input, without a human reading what came out of the first one in between. The team has a verification layer that catches mistakes that would otherwise compound. The founder reviews end-of-day output, not intermediate steps.

What it looks like: our venture studio. One human (Christine). 17+ specialist agents. They produce blog posts, send outreach, close deals, handle support, run the books, all on their own. The founder reviews a daily roll-up of what shipped and makes the few decisions that legally or strategically need a human. Total infrastructure cost: under $200/month. Headcount: one.

Who needs stage 5: honestly, not most companies. Stage 5 is the right target if you're running a venture studio, an indie holding company, a creator-led media business, or any company where founder output per hour matters more than headcount efficiency. For most small-to-mid businesses, stage 4 (Delegate) is the right target. Stage 3 (Assistant) covers 70% of the value with 30% of the build cost.

What stage 5 costs you: the most engineering investment upfront (3-6 months of focused build) and the most operating discipline (verification has to be airtight or the whole system goes off the rails). Once running, ongoing cost is low and output scales without headcount.

Where most companies actually are

Honest scoring. Most companies who say "we have an AI strategy" are at stage 2 (Copilot). Some have one scheduled cycle and are at stage 3 (Assistant) for that one cycle while the rest of the org is still at stage 2. Almost nobody outside venture-studio land is at stage 4 (Delegate) at scale. Stage 5 (Colleague) is rare and probably should be. It's the right answer for a few business types, not most.

The trap is calling yourself stage 4 because you have one Zapier flow that runs without you. The trap on the other side is calling yourself stage 1 because you haven't drawn the org chart yet, when in fact your team is using AI heavily and you're sitting at stage 2 by default. Score honestly. The diagnostic is the work that ships, not the tooling installed.

The move at each stage, in one sentence

You don't have to go all the way. Most companies should stop at Delegate. The move is the same either way: pick the next one, ship it, score yourself again in 90 days.

If you want this built for your business inside seven days, we ship the Assistant-to-Delegate move as a productized service. The envelope, the verification, the agent itself. See our blueprints for what we ship and what it costs.

What to read next

If you got value from this, the cornerstone post in the series is What is an agentic-AI-first business?. It is the definition that anchors every post we write on this topic. The companion infrastructure piece is the 5 layers of an agentic AI stack, which is the same model from the platform angle instead of the org-design angle.

Coming next in this series: what sales and marketing look like inside an agentic-AI-first company, with the specific roles each agent plays and how the team replaces (most of) a traditional GTM team.

If you want help scoring where your company actually is and picking the next move, email christine@operatoriq.io. Tell me what's running on a schedule today. I'll tell you what to ship next.

Cheers, Christine