The 20VC episode that explained my own business to me

I run a small Attio implementation business. For about a year I have had a working theory for why it sells, but every time I tried to say it out loud it came out fuzzy. Something about AI, something about timing, something about everyone wanting it and nobody being able to build it.

Then I listened to a 20VC episode, the one on Anthropic's raise and the Mag7, and the panel spent about two minutes on the AI services market. They said the thing I had been circling for a year, cleanly, in a way I had not managed to. So I want to write it down, both because it is the clearest framing I have heard and because my own experience says they got one part slightly wrong.

The line that landed

The framing was simple. Corporate America has decided AI is how you transform a company. Every board is telling every CEO to get on top of it. The naysayers have gone quiet. So the demand is real and it is universal.

And then the gap: almost nobody has the expertise in-house to actually do anything about it. One of the panel called it the worst gap between in-house and external expertise in our lifetimes. The dollars are going up, the mandate is going up, and the ability to execute is near zero.

That is the whole business in two sentences. The demand for AI is not the problem. The capacity to deliver it is. And whoever can deliver it gets to charge consulting prices for the gap.

The data backs the feeling. In a 2026 survey, 54% of small and mid-sized businesses named lack of in-house expertise as a top barrier to AI adoption, second only to cost (Stealth Agents, 2026). Nearly 60% of organizations said the same about agentic AI specifically. Those numbers are not a skills shortage. They are a market.

Why this is exactly my business

I did not start Craftt because I read a market report. I started it because I noticed I could do something in an afternoon that teams could not do in a quarter, and they were happy to pay for the afternoon.

The 20VC framing explained why that keeps working. The customer cannot hire the expertise, because the people who have it are rare and expensive and not looking. They cannot train it fast enough, because the tools change every few weeks. They cannot buy it as software, because the hard part is not the software, it is knowing how to shape the thing. So they rent it. From someone who already spent the months.

That is the actual product. Not the Attio workspace. The months I spent learning how to build one that an AI can operate. The workspace is just the form the expertise takes when I hand it over.

The part they got right that stung a little

One of the panelists said the quiet part about agencies. The HubSpot agencies, the Shopify dev shops, the implementation partners who deploy these tools for small businesses, most of them are gone or going, because they have no AI play. The few that build one will have, in his words, infinite demand.

That one stung, because the default version of my business is the dying one. "I set up your CRM" is the HubSpot-agency offer. If that is all Craftt is, the panel just described my own obituary.

The reason it does not apply is the part I had to learn the hard way. The defensible offer is not "I set up your Attio." It is "I build the AI agents that operate your Attio after I leave." The setup is the entry point. It proves the expertise gap is real for that specific customer, in their own data, on their own screen. The agents are the business, because they are the part the customer cannot pull off alone and cannot stop paying for once it works.

The agency layer is not dying because agencies are bad. It is dying because the old offer was a one-time build with no reason to come back. Attach something that operates, that drifts, that needs tuning, and the same channel becomes a retainer.

Where my experience says they are slightly off

The panel framed this as a consulting trade: spend six months learning, sell it for a week of your time, pocket the gap. That is true for the build. It is not the whole shape of it.

In my experience the build is the smaller half. The larger half is that AI setups rot. Prompts that worked in March miss by June. Fields go stale. Agents that were tuned for one pipeline drift when the pipeline changes. A normal CRM degrades slowly and politely. An AI-native one degrades faster, because people trust the fields more and notice less.

That rot is not a flaw in the trade. It is the recurring revenue. The consulting framing makes it sound like a clever one-time arbitrage, learn-once-sell-many. The real version is closer to managed infrastructure. You sell the build to prove it works, and then you sell the operating contract because the thing you built needs an operator. The customer is not paying for my six months anymore. They are paying so the system they now depend on does not quietly break.

So the trade is real, but the margin does not live where the panel put it. It lives in the month after the build, every month, for as long as the agents run.

What I am doing about it

Nothing dramatic. Mostly it confirmed the direction and killed my doubt about it.

I keep the one-time build cheap and honest, because its job is to prove the gap, not to be the business. I put the real weight on the agents and the monthly operating work, because that is the part that survives contact with the panel's prediction. And I let the customer own their own runtime and keys, so the retainer is a choice they keep making, not a lock-in they resent.

The episode did not tell me anything I was not already doing. It told me why it was working, in language I can now actually use when someone asks what I do. That is worth more than it sounds. For a year I had a business that ran on instinct. Now I have the sentence.

The demand for AI is not the bottleneck. The ability to deliver it is. I sell the ability. The setup proves it. The agents keep it sold.