The company that ships AI without an AI team

For most of the last three years, a company that wanted to put artificial intelligence into its own operations faced a hiring problem before it faced a technical one. The talent that could build a reliable assistant, wire a retrieval system to internal documents, or automate a workflow with a language model in the loop was scarce, expensive, and mostly employed by someone else. The result was a familiar gap: enterprises with the budget to hire an AI team did, and everyone smaller waited, watching capabilities they could not staff.

That gap is closing from an unexpected direction. Not because machine-learning engineers became cheap, but because the work stopped requiring them. As the frontier models turned into commodity infrastructure reachable through an API, the skill that mattered shifted from training models to integrating them, and integration is something a competent software team can do without a research background. The consequence is that a company can now deploy custom AI without an AI team, by commissioning it the way it would commission a website.

The generalist absorbs another specialty

We have written before about how the custom CRM and the workflow automation migrated from specialist software houses onto the menus of generalist digital agencies. The AI assistant is following the same path, and faster. Consider Devign, a remote-first agency working across Lebanon and the United States, whose business-systems division lists chatbots and business automation beside custom CRMs, quoted with a delivery window rather than an open research timeline. The AI is not sold as a moonshot. It is a line item next to the database.

That framing is the whole story. When a capability is sold with a window attached, the seller is claiming it has done the shape often enough to estimate it, which is only possible once the underlying technology has stopped being a frontier and become a component. A generalist agency putting a chatbot on its price list beside the website is a market signal that production conversational AI has crossed from the laboratory into the catalog.

When the AI assistant becomes a line item next to the database, the market has decided it is a component, not a marvel. The window on the quote is the tell.

What the commission route actually buys

For the buyer, the appeal is the same as with any productized build. A company that would never hire a machine-learning engineer, and could not evaluate one if it tried, can sign a fixed-scope engagement to get an assistant trained on its own documents or an automation that routes its own tickets. The vendor absorbs the parts the buyer cannot judge: which model, how to ground it, how to keep it from confidently inventing answers, how to make it fail safely when it does not know. What the buyer keeps is the thing it does understand, its own process and its own data.

The honest caveats are the ones that always attend de-specialization. A generalist selling web, mobile, automation, and AI from one page is making a breadth claim, and breadth trades against depth. An assistant that took four weeks is not the assistant that took four months, and the difference usually lives in the unglamorous parts: how the system handles the questions it was not built for, whether the retrieval is accurate enough to trust, what happens to the data. A buyer should ask which version the window describes before mistaking a competent wrapper for a system built around its problem.

The questions that survive the hype

The useful questions here are old ones wearing new clothes. Who owns the prompts, the configuration, and the integration when the engagement ends. What does the assistant do when it does not know the answer, and can we see it do that before we sign. Where does our data go, which model processes it, and under whose terms. How much does it cost to run each month, separate from the cost to build, because an AI feature carries an ongoing inference bill that a static website does not.

None of these require understanding transformers. They require the same procurement discipline a company would apply to any vendor holding something it depends on, and the fact that the deliverable is fashionable does not change the questions, it only makes buyers more likely to skip them.

What we would watch

The pattern to track is whether the commissioned assistant proves deep enough to keep. The bet a generalist makes is that most companies want one accountable vendor for their software and their AI rather than a separate consultancy, and that the integration work, now that models are commodities, is within a good software team’s reach. We think that bet is mostly right for the broad middle of the market, the companies whose AI needs are real but not novel, an assistant over known documents, an automation over a known process.

The frontier, the genuinely new model capability, still belongs to specialists. But most companies do not need the frontier. They need the last decade’s research, delivered reliably, wired to their own data, with someone to call when it breaks. That is a commission, not a hire, and the agencies that figured this out early are quietly building the AI layer for a market that was told it needed a team it could never afford.