Reef TRHThe model was the easy part, and almost nobody believes that until they try to ship it. The prototype that gets the nods in the room, the one that reads the document or answers the question, now takes a good engineer days, sometimes an afternoon. The system that survives contact with your actual business takes months. The distance between those two things is not the intelligence. It is everything you have to connect the intelligence to.
This is the quiet inversion of the last two years. The part that used to be hard, building something that behaves intelligently, got radically easier. The part that was always hard, making it work inside a real organisation, did not move at all. And because the clever-looking half now comes cheap, teams keep mistaking it for the whole job.
The model is the part everyone can do now
A capable model is a call away, and a demo on sample data is a weekend. That is genuinely new, and worth being excited about. But it plays a trick on the people paying for it: the intelligence arrived so easily that the rest feels like a formality, a bit of wiring to finish off. It isn't a formality. The demo is the first tenth of the work, and it is the tenth that was always going to succeed.
The project is the plumbing
The real work starts when the model has to live inside the systems your business already runs on, and data has to move both ways without anything breaking. The AI has to read from the record system, the billing tool, the scheduler, the portal a supplier logs into, and the fifteen-year-old database nobody wants to touch, and then write back into them. Around that sit the parts no one demos: access control, so the model only ever sees the record it should; formats that never quite agree; a sync that has to stay honest when two systems disagree about the same customer. None of it is glamorous. All of it is the project. It is the same reason almost anyone can demo AI now and almost no one can ship it, the execution gap we keep coming back to.
The unglamorous nine-tenths
The edge cases that are rare on a slide and constant on a Tuesday. The human who has to sign off anything that carries real risk. The evaluation and monitoring that catch the model drifting a month after launch. This is the work that decides whether the thing still runs in six months, and it is precisely the work a proof of concept is designed not to show, which is why most pilots never reach production.
This is also where the timeline actually sits. For one property manager we worked with, the entire project was a bridge into a fifteen-year-old system that had no sane way in; the AI that read the documents on top of it was the quick part, finished long before the integration was trusted. The intelligence was never the bottleneck. The plumbing was, as it nearly always is, the reason a promising pilot stalls on the way to production.
So when a vendor sells you "the AI", notice what is actually on the table: the tenth that was always going to be easy. The question that decides whether you end up with a system or a very good demo is the one that never makes the slide, who is building the other nine-tenths, and do they know your systems well enough to survive them. Ask that first. The model will be fine. The project is everything after it.
Common questions
If the model is the easy part, why does AI cost so much to put in production?
Because the cost was never the model. It is the integration into your existing systems, the two-way data flow, the edge cases and the monitoring after launch. The intelligence is a small, cheap layer on top of a large amount of engineering.
Can't we just buy an off-the-shelf AI tool and skip the integration?
For a generic, standalone task, sometimes. But the value usually lives in your own data and your own systems, and connecting the AI to those safely is exactly the work a shrink-wrapped tool leaves for you to do.
How long does the integration actually take?
Far longer than the prototype. Plan the timeline around the plumbing, the permissions and the exceptions, not the demo. The model working is the start of the project, not the end of it.