The Day the Ship Date Stopped Moving

Thread AI Editorial · November 10, 2026

The first note mentions a question about records FDA inspectors might ask to see. The second points to a security review. The third cites a finding from the client’s audit team.

The agent is designed to speed up the consultancy’s core work, supplier qualification for mid-sized pharmaceutical and medical device companies. Before a client works with a new supplier, the firm checks supplier audit reports, certifications, and quality records against the client’s requirements. The firm’s reviewers then recommend to the client whether to work with the supplier.

The review work had been piling up for the longtime client, which asked whether AI could help clear the backlog. The consultancy built an agent that reads each supplier’s documents against that client’s requirements and drafts a recommendation for a reviewer from the firm to check. The prototype ran on a capable open-source framework and was working within a few weeks.

Early results were promising enough that the firm saw a bigger opportunity to turn its supplier review playbook into a product. Every client needs some version of this review, so the same product could serve all of them, with the firm’s reviewers still making the final calls.

The work fell to experts in supplier quality who had never built or run an AI product. The prototype came together in a few weeks, so 12 weeks to launch seemed like plenty. But a product the client would use every day, and that the firm hoped to sell to others, had to answer to the client’s quality team, security reviewers, and auditors. Critical questions, like what records the agent should keep or how a reviewer would sign off on a recommendation, dragged on for weeks.

Three Notes, One Question

As the timeline slipped, notes started appearing on the plan. After the third slip, a managing partner read the notes side by side to understand what went wrong.

The first note came from the client’s quality team, which wanted to know what approval records would show FDA inspectors about the agent’s role in each recommendation. The prototype kept no history, so no one could tell the agent’s draft from the reviewer’s edits, and the ship date moved for the first time.

The second note came from the client’s security team, which wanted proof that the documents the agent read didn’t come from another client, since several of the firm’s clients compete. With no log of the documents the agent pulled, the firm had nothing to show, and the date moved again.

The third note came from the client’s audit team, which asked the firm to prove each recommendation used the requirements in effect at the time. The prototype didn’t track which version of the requirements the agent used, so the date had to move a third time.

Read one at a time, each note looked like a contained fix. Read together, they showed three different groups asking the same thing: “What did this agent do, and why?”

The List Between You and the Date

To answer the three notes, the firm had to build what the prototype skipped, including a record of every draft and edit, document logs, data isolation, and version tracking. This work required an engineering budget, a hiring plan, and specialized engineers the firm had never needed before.

All told, building the agent took weeks. Building the infrastructure around it took quarters.

The open-source framework behind the prototype did its job. Tools like it give teams what they need to build an agent, and they leave the record-keeping, data separation, and version tracking for each team to handle on its own.

The consultancy knows exactly what the agent needs to do, which makes it easy to plan. The launch date slips when the list of unplanned work keeps getting longer. That list grows fastest around governance, because each client brings its own quality team, security team, and auditors, and any of them can ask what the agent did and why.

This pattern reaches well beyond life sciences. Thread AI’s analysis of enterprise AI buying conversations, published as The AI Ship Date Dilemma, found that only 5% had a dated, external commitment to ship with AI. The research traces that hesitation to the work of getting to production, which runs larger and more complex than teams first expect.

One Date, Nothing Beside It

The life sciences consultancy eventually moved the agent onto infrastructure built to run it as a true product for multiple clients, without staffing up like a technology company. Now every recommendation carries its own record. The client’s quality team can show an FDA inspector what the agent drafted and what the reviewer changed, security reviewers can see the documents behind it, and auditors can see which requirements it was checked against.

Without the move, every client the firm added would have brought notes of its own, and each launch date would have moved the way the first one did. The plan for the firm’s next client has one launch date and nothing written beside it, because the only work left is setting up that client’s requirements. This time, the list stopped growing, so the date stopped moving.

The Series

This is Post 6 of six, and the second of two about productizing the playbook. Read the thinking behind this series in the manifesto, and our research about product development timelines in the AI Ship Date Dilemma report.