Why your systems integrator can't build you an agent

You asked your systems integrator for an AI agent. They came back with a fourteen-person team, an eighteen-month roadmap, a discovery phase, and a steering committee with its own logo. Somewhere around slide thirty you felt the familiar gravity of a Large Programme forming around what you'd hoped was a focused build.

Here's the uncomfortable explanation, and the position of this essay: your integrator can't build you an agent — not because their people lack talent, but because their business model sells hours, and agents exist to destroy hours. That conflict is structural, not moral. No amount of goodwill on either side of the table repeals it. And once you see the mechanism, a decade of stalled corporate AI programmes stops being mysterious.

The physics of the billable hour

A systems-integration business is an equation: revenue = heads × hours × rate. Everything about how such firms sell, staff, and scope flows from maximizing those three factors. Utilization is the god-metric; the bench must stay billable; big teams on long timelines are not a failure mode of the model — they are the model.

Now place an agent build inside that equation and watch it misbehave twice:

First: the deliverable attacks the unit of sale. An agent that eliminates ten thousand hours a year of back-office work is, to an hours-selling firm, a strange object — a machine whose explicit purpose is to make hours disappear. They can sell it once. But their instincts, pricing sheets, and promotion criteria were all forged by selling the presence of effort, and an agent is the absence of effort, productized.

Second: done well, it's too small to matter. The honest version of most agent projects — the version we described in The demo-to-production chasm — is a handful of senior people for a couple of months, then a long tail of operation and improvement. To a body shop's revenue plan, that's a rounding error. So the proposal inflates until it fits the business model: enter the discovery phase, the fourteen chairs, the committee. Nobody is lying to you. The shape of the deal is doing the talking.

The partner-stack trap

There's a second incentive layer, quieter and just as binding. Integrators live inside partner programmes: certifications, reseller margins, co-selling agreements with the big platforms. Those partnerships pay — in licence margin and in lead flow — which means the integrator's answer space is bounded before you ask the question. The diagnosis will, with remarkable reliability, be that you need the platform they're certified in, plus its new AI add-on, plus implementation hours on top.

Agents want the opposite architecture: thin, composable, model-agnostic — a small system that orchestrates across your existing tools through APIs, swaps models as the frontier moves, and accumulates your knowledge rather than a vendor's licence count. Nobody gets certified in "thin and composable". There's no margin in it — for them. The margin in it is yours.

Agents are a product discipline, not a project discipline

Here's the deepest mismatch, the one that would remain even if incentives were pure. A project has a defining ritual: handover. Requirements in, deliverable out, sign-off, exit. The integrator's entire operating system — contracts, staffing, QA, the champagne — is built around that ending.

An agent doesn't end. It begins at handover:

• Evals need running on every prompt, model, and retrieval change — forever.
• Memory accumulates and needs curating; playbooks update when the agent is corrected (the compounding we described in Agents need memory, not bigger models).
• Models leapfrog each other; someone must re-benchmark and swap engines without breaking behaviour.
• The business drifts — prices, policies, people — and the agent's ground truth must drift with it.
• Telemetry must be watched by someone empowered to act on what it shows.

That's not maintenance in the "keep the lights on" sense; it's operation — closer to running a small product than to owning a delivered system. Projects are handed over; products are lived with. An agent handed over like a project doesn't fail dramatically — it just quietly rots: evals stale, memory unpruned, prompts tuned for a model two generations gone. Handover is where agents go to die.

You cannot buy an agent from a business model that sells hours. The incentives don't survive the invoice.

What to demand instead

The fix isn't "hire a better integrator". It's contracting for product-engineering incentives — from whoever you engage, including us:

1. Outcome scope, not FTE counts. The contract names the workflow, the done-state, and the measurable target — not the number of consultants attending your meetings. (The purchasing philosophy is Don't buy AI, buy outcomes.)
2. Working software in week two. A thin slice of the real workflow, in production contact, in days — not an architecture document in month three. The speed isn't bravado; it's the only honest way to discover where the real problems live.
3. The eval suite is a named deliverable. Golden cases, accuracy thresholds, regression runs — versioned, handed to you, yours. A supplier who won't be measured is telling you their forecast of the measurement.
4. An operate-and-improve arrangement, or a real internal owner. Someone runs the agent after launch: watches the error lane, curates memory, re-benchmarks models. If it's you, the deal must include training your owner — not a PDF, an apprenticeship.
5. A team you can name. Two to four senior engineers, whose names you know, who attend zero steering committees. If the org chart on the proposal has more layers than the architecture diagram, believe the org chart.
6. IP and exit rights. Your prompts, your playbooks, your data, your evals — portable on day one. Anything else is dependency with extra steps.

Notice what this list does: it makes the engagement small, senior, iterative, and permanent — the exact quadrant the hours model can't inhabit, which is why demanding it works as a filter. The firms that can't say yes will excuse themselves politely.

What this looks like in practice

Our own answer to the incentive problem is structural: we build like a product company that happens to serve clients, from a place where senior engineering is findable and the economics let small stay small — that's part of why we build in Portugal.

And we operate what we ship, on ourselves first. The assistant in the corner of this page — retrieval over the whole site, persistent memory — is ours, in production, watched and improved. The publishing hub that turns one brief into blog, LinkedIn, and Facebook output runs our own content operation. The brandbook-to-website engine renders real multi-tenant sites from structured brand data, this one included. None of that proves we're right for your problem — but it proves the operating discipline exists, because we'd feel its absence before any client did. Dogfooding is the only honesty test in this industry that can't be faked in a slide. It's also why our AI systems page describes running systems, not capabilities.

A clearly illustrative archetype for contrast: picture a mid-market distributor that wants inbound orders drafted into the ERP automatically. The integrator quote: fourteen people, discovery, a platform migration "while we're at it", go-live in Q4 next year. The product-shaped alternative: three seniors; week one, the agent reads real historical emails and drafts orders in a sandbox; week three, it runs live behind an approval gate; week ten, the eval suite says which order types it may file alone. Total elapsed time: less than the integrator's discovery phase. The difference wasn't talent. It was what each team's business model needed to be true.

The honest caveat

Integrators aren't villains, and this isn't a category execution. Large ERP rollouts, infrastructure migrations, scaled staff augmentation, twenty-country system harmonization — genuinely project-shaped work with real handovers — is what that model was built for, and it does it well. The claim is narrower and sharper: agents are not project-shaped. They're small, senior, iterative, and permanent — and no amount of talent inside an hours-driven structure can hold that shape for long, because the structure always wins.

So when the fourteen-person proposal lands on your desk, you don't need to argue with it. Just ask the six questions above and watch what happens to the room.

If you'd like a second read on an agent proposal you've received — including "this part is actually fine, keep it" when that's true — talk to us. Worst case, you'll walk away knowing exactly which incentives you're buying.