Connecting AI to your legacy systems: the real timeline

Ask a vendor how long it takes to add AI to your product and you will hear a number that assumes the model is the whole job. Ask anyone who has actually shipped it into a decade-old system and the number changes shape entirely, because the legacy platform gets a vote, and it votes slowly.

The gap is not small. When AI has to integrate with heavy enterprise systems like an ERP, industry timelines for a full production integration run something like 26 to 32 months. That is not the model taking two years. That is everything around the model catching up.

Adding AI without putting the old system to rest

The CTO of a dental telehealth startup described the exact move most teams face. His team did not get to build on a clean slate. They had a legacy platform running the business, and they had to add AI into it without breaking what already worked. His description was precise: “We did not want to build it as a monolith, we built it as microservice-based services. Now all our AI agents are also microservices. We are adding that into the legacy system and making sure the legacy system is not put to rest.”

That last clause is the whole timeline. “Making sure the legacy system is not put to rest” means every new AI service has to coexist with code that customers already depend on, data that already has meaning, and workflows people already trust. You cannot rip and replace. You bolt on, carefully, and you test against the reality that there is a patient on the other side.

He was blunt about what that does to velocity. The old startup joke about shipping a feature in two weeks stopped being funny. Testing and publication took longer than the coding, because a mistake in a healthcare workflow is not a bug ticket, it is a safety event. And the scenarios never end. As he put it, you might have worked through 30 different cases, “but once you give it to someone, they’ll come up with another 30 to fix.”

The device that would not cooperate

The same team learned the timeline lesson again with hardware. Their intraoral camera was, in his words, “one of the major, major challenges.” It went through three hardware versions and an SDK, and the Wi-Fi camera fought the phone for the connection, so that if a patient connected to the camera their phone lost internet. Streaming a live image into the patient chart in real time was not a plug-in. It was a six to seven step designed workflow, built and rebuilt until it survived a real appointment.

None of that showed up in an early estimate, because early estimates price the exciting part. The camera was supposed to be a component. It became a quarter of the roadmap.

Why the calendar stretches

Legacy systems were built for a purpose that predates your AI feature. They hold data in shapes nobody documented, enforce rules buried in old code, and connect to other systems through habits rather than contracts. Every one of those has to be discovered, respected, and worked around. The model is ready in a sprint. The integration is ready when the old system stops surprising you, and old systems have a lot of surprises.

How we approach it at Density Labs

In the AI Readiness Assessment ($2,500), we timeline the integration separately from the model, because they run on different clocks. We inventory what the legacy system actually enforces, where the AI has to read and write, and which coexistence rules cannot be violated while the old code keeps running. Then we give a range that includes the surprises, not just the happy path. A two-year integration is not a failure. Calling it a two-month project is.

The model is the sprint. The legacy system is the marathon. Plan the calendar around the one that is longer.