The GTM AI adoption curve: four stages, two traps

Work Musings · GTM AI

Is your company AI-native?

Now a harder one. Is that mostly because every rep has a Claude or Codex license?

That was us too. Two years ago, AI in go-to-market at Canva was one person in ops with a Clay table.

I presented this at Sculpt this year. I thought I’d post it here too, for anyone who couldn’t be in the room.

It’s the story of how we got from there to here, drawn as a curve with four stages.

By the end, you should be able to point at the curve, say “that’s us,” and know what your next stage needs.

The curve tracks how a go-to-market organization adopts AI, not how one clever rep does.

We’ve been working through these seriously for about a year. It didn’t happen overnight.

At each stage, I’ll cover what worked for us and the trap I’d watch for.

Stage one is really an assessment. You can’t roll out an AI strategy until you know where your reps are, because your reps are your customers.

You’ll see a huge range. Some reps barely open the tool, and others do almost everything in AI.

That second group is also the one costing you a lot of money.

The trap at this stage is the one I opened with: calling yourself AI-native because every rep has a license.

Stage two is a shared AI capability, owned by RevOps or a GTM engineer, so reps stop figuring it out alone. That means skills, connectors and company knowledge, built once for everyone.

A skill is a reusable set of instructions the AI follows for a task. We started with data access, then centralized the skills:

It’s baked into onboarding too. New reps get it as a core part of their first 90 days.

Salesforce data is messy. We have seven different ARR fields.

So alongside the Salesforce MCP, we wrote a skill that tells the agent which ARR field to use, and which field to use when you pull a region. MCP is the connection that lets an AI tool work with Salesforce.

It was one of our most successful launches. Make it dead simple, and give reps direction.

Once the agent understands Salesforce, you plug in Gong and Snowflake. Now you can summarize every call, every account and every QBR.

At some point you’re summarizing the summaries. In the talk, this is the slide with Admiral Ackbar on it.

This is the aggregation trap: summaries everywhere, long docs nobody asked for, and no work actually done.

The way out is to stop aggregating data and start accomplishing tasks. Take the follow-up: you have the transcript and the summary, and the follow-up is still a blank page.

So the skill saves a follow-up draft in Gmail, and the rep reviews it and hits send. No copying and pasting out of Claude.

I walked through that setup in From helpful answers to finished work.

Where we’ve had the most success finishing work for reps is the Canva MCP. I’m contractually obligated to call it my favorite feature. It’s also true.

It goes beyond decks: digital sales rooms, customer leave-behinds and social content.

I wrote about why deliverables get adopted in The AI workflow works. Why is nobody using it?

Stage three is agentic programs, and building them takes real lifting. The first move almost everyone makes is to hire an agentic lead.

You go find your knight in shining armor. Sales background, ops chops, knows every tool, startup on the résumé.

Then you wait for the millions in pipeline, because you’ve seen the LinkedIn posts. It’s a trap. Again.

And this one’s more expensive, because it comes with a salary.

Here’s the job you actually hired them into:

And drive pipeline in month one.

Before you hire an agentic lead, check that three foundations are in place. Or hire someone to build them, and give it time.

Agents can scale a motion. They can’t agree on one for you.

Then put the right people around the lead. We think of it as a pod: someone on systems and data, a BDR or prospecting lead with enough pull to define the motion, and an executive sponsor above.

Turns out you need more than one knight.

Stage four is custom tools and continuous improvement. The signal I’d point to is hiring: companies are putting software engineers inside go-to-market teams.

When I built the talk, Anthropic, OpenAI and Vanta all had roles like that open. The pattern is agents, evals and feedback loops, built inside the revenue team.

Evals, short for evaluations, are repeatable checks that hold the work to a standard.

Our first project with real engineers on it is a custom GTM harness, built on our Claude and Codex workflows. A harness is the software around an AI model: its context, tools, permissions and actions.

It’s still in development. I wrote about it in What is a harness, actually?

Why build our own? Three reasons:

That last one matters more than it sounds. AI spend is going to hit your P&L sooner or later.

The trap at this stage is chasing the hot new tool before the foundation and evals are ready.

Because we own the harness, we have the tracking. That’s what could turn agentic prospecting into a real growth experiment: launch variants, check they’re on the right play, measure what moved.

We’re not there yet. But we’re a lot closer than one person with a Clay table.

If you keep one thing from this post, keep this table.

You don’t need to be at stage four. What matters is knowing what your next stage needs before you jump to it.

So point at the curve and be honest about where you are. Then take the step that stage is asking for:

Stop asking what an agent can do. By stage four, you’ll know.

Robert