There's a gold rush in AI services right now. Every consultancy, dev shop and freelancer has rebranded around "AI integration," and clients are buying. But a lot of that revenue is being booked into a structurally bad business model, and the people running these shops can feel it even if they can't name it.

The trap

The default AI-agency model is to sell implementation by the hour or by the project: we'll build you a chatbot, an automation, a RAG system over your docs. It works, briefly. Demand is high, supply of people who can actually ship is lower, and rates are good.

The problem is that the same forces creating the demand are eroding the moat. The tools that let a small team ship an AI feature in a week — better models, better frameworks, better scaffolding — are available to everyone, and they're getting better fast. What takes skill today takes a wizard tomorrow. As the work commoditises, so does the price. You end up in a race to the bottom against a global pool of people with identical tooling, where your only lever is being cheaper or faster, and both of those are temporary.

Worse, the project model is feast-or-famine. You're only as good as your next signed statement of work, you have no recurring revenue, and the moment a client's internal team levels up, they stop calling.

How the good ones escape

The agencies building something durable do three things differently.

They productise. Every time they solve a problem for one client, they ask whether that solution is a one-off or an asset. The good ones turn repeated work into reusable components — a deployment template, an eval harness for a common task, a connector to a popular system — so the marginal cost of the next similar engagement drops while the price holds. Over time the agency starts to look less like a body shop and more like a product company with services attached.

They price on outcomes, not hours. Hourly billing caps your upside at your headcount and aligns you against the client's interest in efficiency. Outcome-based or value-based pricing — tied to a reduction in handling time, a lift in conversion, a number of tickets deflected — both raises the ceiling and changes the conversation from cost to return. It's harder to sell and requires you to actually measure impact, which is precisely why most shops don't do it, and precisely why it's defensible.

They sell the ongoing job, not the one-time build. This is the big one. The build is the least defensible part of an AI engagement, because it's a snapshot that's true for a moment. The defensible work is everything that happens after launch: monitoring quality as models and data drift, re-running evals, swapping models when better ones ship, owning accountability when the system is wrong. That's a managed-service retainer, and it's worth more to the client and to you than any project fee.

The moat is operational

The uncomfortable truth is that the moat in AI services isn't the clever build — it's the boring, continuous operational accountability that clients don't want to staff for and can't easily replicate. Owning the evals, the data pipelines, the on-call, and the institutional memory of how a client's system behaves is genuinely hard to copy, because it accrues over time and is made of relationship and context, not code.

This is why the most interesting AI-services businesses are quietly converging on a model that looks like managed infrastructure: predictable recurring revenue, outcomes they're accountable for, and a productised core that keeps margins from collapsing. The agencies still selling hours by the bucket are working harder every quarter for less. The ones turning work into assets and assets into retainers are building something that compounds.