An end-of-year letter to early customers
Short. What we shipped this year, what we did not, what we learned from you. No CTA, no pitch, no asking for anything. Just an accounting.
This post is a letter to the people who signed on when PursuitAgent was a product we were clearly still building. There are not many of you. You know who you are.
I wanted to write down, publicly, what you made possible in 2025.
What you got us to ship
The eight-stage pipeline document would not exist without three of you explaining, in painful detail, which stage your previous tools failed at. The grounded retrieval pillar is shaped by a single conversation with a head of proposals who said, “I don’t care if the draft is good if I can’t verify every sentence.” We heard that sentence several times afterward. We designed around it.
The DDQ response playbook is drawn from two of your DDQ libraries. Not the text — the structure. Watching how you organized answers, what metadata you kept, what went stale first, what survived three renewal cycles. That is not a thing you can get from a survey. It is a thing that comes from being allowed inside the workflow.
The bulk of the shipped posts this year trace back to a specific request from a specific customer. The diagram extraction work came from a team that responded to construction RFPs. The per-block permissions landed because a regulated-industry customer needed their legal team to see a subset of the KB the sales team did not. The freshness alerts existed because somebody told us in September that the worst thing that happens is shipping a three-year-old answer and the buyer catching it.
We built from the things you told us, and the posts that read well this year read well because the product caught up to the problems you named.
What we did not ship
I will not pretend the year was clean. Three things on our roadmap from January did not land in 2025.
We did not ship the native integration with the federal SAM.gov feed. We scoped it, built a prototype, found the data model harder than we estimated, and parked it for a Q1 2026 rebuild with clearer requirements. Two customers asked for it by name. They know it did not ship.
We did not ship the team-level win-rate reporting dashboard the way it was promised. A version exists; it is not the version we described. We chose to under-deliver visibly rather than over-deliver quietly. Both options were bad. We picked the one where the customer knows.
We did not ship a self-serve onboarding path. Every customer this year went through a human. That is fine at our stage; it will not be fine at the next stage; it is not fixed yet. Q2 2026 problem.
What we learned that we did not expect
The thing I thought would be the hardest — convincing proposal teams that grounded retrieval was worth caring about — was not the hardest. You were already convinced. You had already watched the generic-AI wave and decided it did not meet the trust bar. You did not need the sales pitch. You needed the tool to actually work.
The thing I underestimated — and I will own this publicly — was how much of the work is workflow, not retrieval. A KB that retrieves well is half the product. The other half is the ritual around the KB: who owns what, who updates what, what gets retired, what gets promoted. That half is mostly not an AI problem. It is a coordination problem, and you helped us see that.
The compounding pillar landing this week is the long form of what you taught us: proposal software earns its keep only if it accumulates knowledge across bids. A stateless drafting tool does not compound. A KB that rots does not compound. A post-mortem that does not write back to the corpus does not compound. The product only makes sense if it does all three.
No ask
This letter has no CTA. I am not asking you to introduce me to your peers, renew early, write a review, or do anything else. Customer gratitude posts that end with an ask are not gratitude; they are the thing dressed as gratitude.
The only thing I will say is: your patience in a year when a lot of things were visibly under construction was not obligated by the contract. You could have used a finished tool. You chose an unfinished one because you believed it was going somewhere. I hope we have made that bet more defensible than it was in January. Next year will be another accounting.
Happy new year. See you in 2026.
— Bo