Competing on groundedness, not features

The AI proposal category is running a feature race. I watched six competitor launches in the last 60 days. All six added the same three things: a chat-with-your-RFP interface, an agent-mode for bulk question answering, and a template gallery. All six positioned these as differentiation. None of them are differentiation. They are baseline.

This is an opinion piece. My opinion is that the feature race is a trap, and every AI-native tool that plays it will look identical by the end of 2026. The only axis that compounds is whether the outputs are grounded — tied, per claim, to a source an evaluator can check.

The feature race, in two paragraphs

A category full of startups and incumbents, all shipping adjacent features on a six-week cadence. Loopio adds an agent. Responsive adds a chat UI. A new entrant ships “one-click full response.” Sales teams track features in a matrix. RFP evaluations by prospects get filtered through that matrix. The sales motion rewards whoever shipped last.

The product motion rewards the same thing. Roadmap conversations become “what did they just ship.” Engineering teams get pulled from foundational work to match a demo. The feature surface grows; the quality of the core draft does not. Reviewers on G2 start calling this out directly — the AI feature works on simple questions and fails on the ones that matter, because nobody has been investing in the part that matters.

What actually differentiates

Grounding does. Not the word — the practice. A response where every substantive claim is tied to a specific paragraph in a specific source document, and the reviewer can click the citation and land on the paragraph. That is a different product category from “AI writes your proposal.” It is the category regulated buyers will actually buy from within three years.

The Stanford HAI study on legal RAG tools is the load-bearing citation here. Lexis+ AI, Westlaw AI, and Ask Practical Law — tools that sell themselves on grounding — hallucinate 17 to 33 percent of the time with retrieval in place. If grounding were a feature, those tools would have shipped it. They shipped the word. They did not ship the practice.

AutogenAI writes about the same failure mode in the proposal category specifically — invented case studies, fabricated compliance claims, statistics that do not exist. Teams under deadline pressure catch some of these and miss others. The ones they miss ship to the buyer.

Why the feature race can’t fix this

You cannot add grounding as a sprint. Grounding is a posture at every layer — retrieval, generation, citation rendering, claim verification, export. You can ship a citation widget that shows a link next to a sentence. That is a feature. You cannot ship — in a sprint — a guarantee that the link points at a paragraph that actually supports the sentence.

That guarantee requires evaluation harnesses that run every day. It requires per-claim verification passes that cost real money on every draft. It requires chunking strategies that preserve semantic units. It requires a content pipeline where the source-of-truth document is always accessible from the response. Each of those pieces takes months. Competitors who prioritize the next feature launch cannot also build them.

The bet

We are betting that by the end of 2028, the AI proposal category splits. On one side: tools that write fast, feature-rich, and hallucinate. On the other: tools where every sentence in the draft is checkable, and the draft takes longer to produce but ships. Regulated buyers will be on the second side. Commodity buyers will be on the first side. The commodity side is smaller than it looks — most B2B buys have a compliance review somewhere.

We wrote about this in grounded AI is not a feature and in feature parity is the wrong goal. Both posts argue the same thing from different angles. This one is the short version: if you are selling an AI proposal tool and your pitch is “we have more features than Loopio,” you are selling on the dimension the buyer is going to stop caring about.

The honest limit

We are not fully grounded yet either. Our claim-verification pass catches most fabrications, not all. Our evaluation harness runs nightly but the ground truth set is still small by any academic standard. Our citations are precise but our UI for checking them is not as fast as it should be. The difference is that these are the things we are fixing, in order, instead of shipping another agent mode. The roadmap tells you what a company believes in. Mine tells you we believe in this.

What the sales conversation actually looks like

A recent call: prospect from a federal systems integrator, shown a feature comparison across five vendors. We were behind on three features. Prospect asked why. I said: because those three features are things we could ship in six weeks, and the thing we have been shipping instead — per-claim verification — is a six-month project that none of the five vendors in the matrix actually have.

They looked at the matrix again. Then they asked what the verification does. We spent 40 minutes on mechanism. Two weeks later they were in a paid evaluation. This is not an anecdote I am scaling to a win-rate claim. It is a shape: prospects who understand mechanism do not evaluate on feature matrices. Prospects who do not understand mechanism do. Our job is to get the first kind into the room.

Takeaway

Pick your axis. If your axis is features, you will lose to a competitor that ships one more. If your axis is groundedness, you will lose sales calls where the prospect has a feature matrix. Either way, you will lose something. The question is what you want to be true about your product in three years. For us, the answer is: the outputs can be trusted. That is worth being behind on a feature matrix for.

Disagree? Write to us. The comments on these posts tell me what the market is actually hearing, and I read all of them.