Building PropelAI: Why I Built an AI Product After 20 Years of Selling Services

I've spent twenty years in government technology. I've been a Deputy CIO. I've run federal practices. I've led captures worth half a billion dollars in total contract value. I've sat in more proposal war rooms than I can count, in windowless conference rooms at midnight, surrounded by Red Bull cans and highlighted RFPs and the particular kind of quiet desperation that descends on a team three days before a submission deadline.

And in almost every single one of those rooms, the same scene plays out.

A senior proposal manager -- someone billing at $150 to $200 an hour, someone with fifteen years of institutional knowledge about how government evaluators actually score proposals -- is hunched over a laptop, manually reading through a 200-page RFP, highlighting every "shall," every "must," every "the contractor will provide." They're copying requirement text into an Excel spreadsheet. They're cross-referencing Section L instructions against Section M evaluation criteria against the Statement of Work in Section C. They're building what the industry calls a compliance traceability matrix.

This process takes 30 to 50 hours. It is, without exaggeration, the most expensive data entry job in professional services.

I've watched this happen hundreds of times. And every time, the same thought: why are our best people doing this?

The Moment It Clicked

The catalyst wasn't dramatic. It was a Tuesday.

We were eight days from submission on a competitive HHS recompete. Our compliance matrix had 340 requirements in it. During the Pink Team review, someone flagged a requirement from an attachment -- Attachment J-7, Exhibit B, paragraph 3.2.1 -- that wasn't in the matrix. We traced it back and found that it cross-referenced a mandatory reporting deliverable in the SOW that changed the entire staffing model for Volume II.

Eight days out. A structural requirement that should have been identified in week one, discovered by accident during a review cycle. We scrambled, rewrote two sections, pulled an all-nighter, and submitted on time. We won. But not because of our process -- in spite of it.

That night, driving home at 2 AM, I started sketching what a machine would need to understand to do this extraction correctly. Not keyword search -- anyone can ctrl-F for "shall." What a machine would need is structural comprehension. It would need to understand that a federal solicitation has an architectural backbone: Section L tells you what to write, Section M tells you what matters, and Section C tells you what you're delivering. And the relationships between those three sections -- what I started calling the Iron Triangle -- are what determine whether a proposal is evaluated as compliant or non-responsive.

That's a hard problem. It's not a "find all instances of this word" problem. It's a "read this document the way a twenty-year proposal veteran reads it" problem.

I spent the next year figuring out how to build it.

Why a Product, Not a Service

After two decades in services -- selling labor, managing programs, staffing teams -- building a product felt like crossing a line. Services people and product people think differently. Services people ask "who's the client and what do they need?" Product people ask "what problem exists at scale and can I solve it with software?"

The answer was obvious once I stopped thinking like a services executive.

The compliance matrix problem isn't unique to any one firm. Every GovCon company that responds to federal RFPs builds them. Every one does it manually. Every one burns the same 30 to 50 hours of senior labor per pursuit. And every one ships a matrix that's incomplete in ways they don't discover until review -- or worse, until evaluation.

That's not a client engagement. That's a product.

The mid-tier GovCon market -- firms doing $10 million to $500 million in revenue -- has over a thousand companies in the DC metro alone. They don't have the dedicated proposal operations teams that the Booz Allens and Deloittes run. They don't have knowledge management systems or AI-powered content libraries. They have five to fifteen people on a capture, and those people do everything: strategy, writing, compliance, reviews, production. Giving them back 40 hours per pursuit isn't incremental. It's transformational.

What I Got Wrong First

I'll be direct about this: the first version of PropelAI wasn't good enough.

The early extraction engine found requirements -- it did the keyword search part well -- but it didn't understand structure. It would pull a "shall" statement from an attachment without knowing whether that attachment was incorporated by reference into the SOW, or whether it was an informational exhibit with no contractual weight. It would identify a Section M evaluation factor but couldn't trace it back to the specific Section L instruction that told you how to organize your response.

That's the difference between a tool that saves time and a tool that changes outcomes.

I rebuilt the extraction architecture around the Iron Triangle principle: every requirement exists in relationship to an instruction and an evaluation factor. If the engine can't map that three-way relationship, it flags the requirement as an orphan for human review. If it can, it places the requirement in the compliance matrix with full traceability -- page, paragraph, bounding box coordinates in the source document.

We call this the Trust Gate. Every line item in a PropelAI compliance matrix links back to exactly where it came from. No hallucinations. No fabricated citations. An evaluator -- or your own proposal manager -- can verify any extraction in seconds.

That's when the product started working.

What PropelAI Is Now

Today, PropelAI does two things:

It generates a complete compliance traceability matrix -- 19 columns, every requirement extracted, classified by type, scored by risk, and cross-referenced across the Iron Triangle -- for $99 in minutes instead of $5,000 to $12,000 in days.

It generates an annotated proposal outline -- volumes, sections, page allocations weighted by evaluation factor importance, compliance checkpoints per section, win theme placeholders -- for another $99.

The compliance matrix tells your team what the RFP requires. The proposal outline tells your team how to structure the response so evaluators find what they're looking for. Together, they replace the first two weeks of a typical proposal effort with a foundation your writers can build on from Day 1.

The Company Library adds a third layer: RAG-powered retrieval of your organization's past proposals, past performance narratives, and capability documentation. When a new RFP drops, PropelAI doesn't just analyze the solicitation -- it connects the requirements to your existing content. Your writers know where they have strong material to adapt and where they need to develop new content.

What I'm Building Toward

PropelAI started with extraction because that's where the most time is wasted and the most risk hides. But extraction is just the foundation.

The roadmap is a proposal intelligence platform -- a system that doesn't just process the RFP in front of you, but builds cumulative intelligence about your competitive position, your content strengths, your win patterns, and the market you're pursuing. Competitive intelligence. Win probability modeling. Pricing analysis. Draft generation that's structurally aligned to evaluation criteria from the first word.

The goal hasn't changed from that 2 AM drive home: make it so a five-person proposal team at a mid-tier firm can compete on quality with a fifty-person operation at a prime. Not by working longer hours. By starting every pursuit with intelligence instead of extraction.

If you've sat in those war rooms, you know why this matters.

---

PropelAI is live. Get a free AI Opportunity Brief and see what AI-powered operations look like for your team.