Our take on the AI deployment gap and where organisations can win

There's a good essay doing the rounds at the moment by Andrew Fisher, called "What happens if this is as good as it gets for AI?" The headline oversells it a little, but the argument underneath is solid: even if model capability plateaus for a while, most organisations still haven't realised the value of the AI they already have access to. Fisher calls the space between what's technically possible and what's actually deployed the "deployment gap".

It's a useful frame. For a lot of routine commercial work, the model isn't the constraint any more. The constraint is whether an organisation has built the harness around it: the context, the workflow integration, the review loops, the judgement about where AI can be trusted and where it can't. Fisher points to forward-deployed engineering, teams embedded inside a business unit, building for that unit's specific systems and workflows, as the pattern that actually closes the gap. Not a platform rollout. Not a chatbot bolted onto everyone's desktop. People who sit inside the work and build for it.

The timing backs the argument up too. Back in May, Anthropic, Blackstone, Hellman & Friedman and Goldman Sachs committed $1.5 billion to build an enterprise AI services firm focused on helping midsize businesses deploy AI across their core operations, then acquired Fractional AI to give it a running start. That venture launched officially under its own brand, Ode with Anthropic, in mid-July, just days after Fisher's essay went up. A foundation model company deciding the deployment layer is worth its own standalone firm, and committing that much capital to it, is a fairly strong signal that this isn't a passing trend.

What we're seeing in our own work

We've been living a version of this argument for a while now, not as a thesis, but as the day-to-day shape of our agent builds with clients. A recent build for a long-standing recruitment technology client is a good example. The first phase, eight agents shipped over eleven weeks, was the kind of thing that looks impressive from the outside: candidate summary agents, reporting tools, a proposal support agent that logs its own GitHub issues and teaches itself from corrections.

What's more interesting is what came after. The next quarter of work isn't about adding more agents. It's about the operating model around them: feedback loops so the candidate summary agent learns from reviewer corrections instead of staying static, recovery functions for the edge cases that break automation, a shared integration layer so new agents don't each reinvent access to the same systems. None of that is a model upgrade. It's the unglamorous work of making agents reliable enough that a business can actually depend on them.

The pattern we keep coming back to isn't just "forward-deployed engineer embedded in the business", though that's part of it. It's more of a triad: a business analyst who understands how the work actually flows (not how the org chart says it flows), an engineer who can build and iterate inside that context, and a subject matter expert who can tell the difference between an agent that's technically working and one that's producing something the business can trust. Take any one of those three out and the gap doesn't close — it just moves.

Where this leaves organisations

Fisher's essay makes a point we agree with: the quiet winners of this next phase probably won't be the AI-native startups or the largest enterprises. They'll be small and medium businesses with real domain depth and a loyal customer base — the kind of proprietary advantage that's easy to have and hard to use well. Most of these businesses won't close their own deployment gap alone, and they shouldn't need to. Their edge is domain knowledge and customer trust built over years, not the ability to stand up an AI engineering team.

Helping businesses like that put that edge to work, without needing to become an AI engineering shop themselves, is the part of this whole shift we find most rewarding.