Building AI so each project makes the next one cheaper

Health system leaders funding their first AI projects want them to add up to something more than a set of unrelated pilots. So far, AI pilots have a poor track record. MIT's 2025 report on the state of AI in business found that about 95% of enterprise generative AI pilots deliver little or no measurable effect on profit and loss. Gartner predicted that at least 30% of generative AI projects would be abandoned after proof of concept by the end of 2025, citing poor data quality, weak risk controls, rising costs, and unclear business value. MIT's researchers wrote that the split between the few organizations getting value and the rest does not seem to come from the quality of the AI models, but from how the organizations go about it. Most of the custom and vendor-sold tools in their study failed because the workflows were brittle, the tools did not learn from context, and they did not fit the way staff work day to day.

How pilots pile up

Imagine a department that finds a workflow AI can help with, hires a vendor or contractor, and ships a pilot. The pilot may work well enough to keep, work badly enough that people argue about it, or get shelved when the person who pushed for it changes jobs. Six months later, another department might do the same thing from scratch, with a different vendor, a different data pull, and none of what the first team learned. In that pattern, no pilot makes the next one cheaper or faster.

A lesson from interoperability

Healthcare has spent about two decades moving systems that were never designed to talk to each other onto shared formats, because one-off connections between every EHR and every downstream tool don't scale. The federal government created a National Coordinator for Health Information Technology in 2004 to coordinate that work. The Congressional Budget Office estimated that Medicare and Medicaid would pay $30 billion from 2011 through 2019 in incentives for electronic health records, which had to be able to exchange health information. That work isn't finished, but the reasoning behind it still holds: with one shared standard, each new system needs one connection instead of a separate one to every system already in place.

When each AI pilot gets its own custom integration, it repeats the old problem of one-off connections between systems, this time with AI tools.

Four practices

A shared data format. Agree on the shape of each AI tool's output so the next tool, person, or system can read it without a custom integration. The format only has to cover the handoffs you already have.

A quality check. Every output is checked against a source of truth before it ships or anyone acts on it, and the model never grades its own work.

A shared knowledge layer. This is one place where all your AI tools keep what they learn. Every tool can read from it and, where appropriate, write to it, so what one tool learns stays available after that tool is retired.

No dead ends. Before a new build ships, it either connects to something that exists or defines the format the next build will connect to.

Why each build costs less than the last

With these in place, each new build can reuse the connections and formats the earlier ones left behind, so it needs less new work than a build that starts from scratch. Disconnected pilots each start over.

An example

Tyler Kent built four connected systems for a public healthcare data platform. Tessera, a research agent, scans the public web daily and drafts possible findings, each with a claim, evidence, sources and a suggested visual, and a person reviews each one before anything is published. Emblema reads the same finding, in the same format, and turns it into a branded infographic that follows the brand rules automatically. Every morning, Sentinel compares the live platform with the production code and the published data to catch errors, such as a broken link or a number that no longer matches its source, and sends each team its own report. Opus, built last, writes a one-page weekly brief for leadership from the other three tools' output and the records the team already keeps. Three of the tools work from another tool's output: Emblema reads Tessera's findings in the same file format, Sentinel checks both tools' work once it is live, and Opus reads what all three did. The case studies have the details.

Where to start

The AI Readiness Sprint is a two-week, $9,500 engagement that comes before the first build. It produces a costed roadmap, a ranked list of opportunities, and one working prototype, so the first project you pick is one that later projects can build on. Then that first project is built at a fixed price, as a Custom AI Tool or a Data Warehouse and Reporting project, and is tested and handed off to your staff with training.