Why the mid-market is the real AI services gap.
If you run a fifty-to-five-hundred person firm and you have tried to buy AI services in the last eighteen months, you already know the shape of the problem. The calls you have taken fall into two camps, and neither is quite what you want.
At one end, the Big 4 and the SI majors will not quote under a quarter of a million. Below that number their cost structures do not work and their partners have better things to do. When they do take the call, the engagement arrives as a strategy workshop, a maturity framework and a roadmap. You will get a deck. You will not get a working piece of software inside eight weeks.
At the other end, a flourishing ecosystem of three-person studios will happily sell you a chatbot wrapper for £4,000. It will be fast, it will look smart in the demo, and it will spend the next six months failing to integrate with practice management. Your paralegals will open it twice, decide it is slower than the current process, and go back to the browser tab.
The middle is missing. The middle is where a senior engineer spends two weeks inside your operation, builds the specific thing that takes hours out of one workflow, and sends you an invoice in the mid five figures. We only take a Build on if the Discovery arithmetic says the firm will have recovered that fee inside six months; if the arithmetic turns out wrong on our side, the Build fee is refundable. And the conversation never once descends into a slide titled AI Maturity Journey.
Why the middle got thinned out
It is not a coincidence. The obvious firms that should have occupied this band have spent the last fifteen months being absorbed. Faculty went to Accenture. Peak AI went to UiPath. Mind Foundry's consulting arm went to Aioi. The survivors are either scaling toward enterprise (and losing mid-market instincts) or niching very hard into a single vertical. The supply side is genuinely thin.
Why the demand side is finally present
Three things landed in the same year. Employer NIC and minimum-wage rises put a measurable line on every finance director's P&L: a back office of any size got materially more expensive in April 2025. Retrieval and tool-use finally became reliable enough that a fixed-price Build is not a research project. And the stack your firm is on - Microsoft 365, Sage, a practice management system, a document store - finally exposes the APIs that make wiring the work in plausible.
The result is that the single question a mid-market FD now asks - "which workflow, what does it cost, when does it pay back" - has answers that were not available two years ago. The firms that learn to ask it in 2026 will spend two years being ahead of the firms that do not.
What it is not
It is not a Copilot licence. Your MSP is not lying to you when they say Microsoft has shipped useful things; they are just answering a different question. A horizontal assistant does not prepare your HMRC investigation brief. It does not run an IR35 determination. It does not read a lease. The work the partner floor actually cares about is vertical, and it needs software with opinions - opinions about which clauses matter, which questions HMRC tends to ask, which evidence is likely to be missing.
What it is
One workflow, picked on the numbers. One Build, fixed price, shipped into your existing stack. One managed retainer, named engineer, answering the phone. Paid back inside six months against the targets we set in Discovery together, or the Build fee goes back.
It turns out this is not a new business shape. It is the shape the regional system integrators had in the 1990s and early 2000s, updated for a world where specialist software has better economics than it used to. The surprise is that it stopped being available for a decade, not that it is coming back.