Value in AI is moving to the context layer. Generating code is cheap. Interfaces on top of data are cheap. What's left is the layer that tells a model what to do, in what order, and whether it's allowed — the institutional knowledge that used to live only in people. When one layer commoditizes, the adjacent one captures the margin.
That's the frame we've been building on for two years, and it's the reason our own stack is a context layer with a model on top rather than an agent with a database underneath. But most of what gets written about the context layer describes the half of it that can be harvested. The half that matters most for go-to-market can't be.
The missing half: harvest vs. capture
Read almost any description of where context comes from and it's all exhaust: email threads, wikis, Slack, onboarding docs, and now the execution traces of agents running workflows. Context accumulates because the work produces it. A system that watches thousands of runs learns that large deals stall at legal review.
That's true for process context. It's true for permissions. It's true for anything with a forcing function: you cannot ship code without touching the repo, so the repo fills. You cannot close a deal without touching the CRM, so the CRM fills.
Now try it on the context that decides whether the deal closes at all. Why did the last ten customers actually buy? What did they call the problem before they found you? Which alternative did they almost pick? What claim on your homepage is true, and which one is the founder's hope? None of that has a forcing function. You can run a company for three years and never once record why a customer bought. Nothing in the work emits it. There is no exhaust to harvest, because the exhaust was never produced.
You cannot harvest what was never emitted.
This half has to be captured: deliberately, from the people who hold it, on the record, with consent. That is a different motion from mining, and it produces a different kind of asset. A harvested fact is whatever the trace happened to contain. A captured claim is something a named person said, knowing it would be used, in a session that existed for that purpose.
And for commercial truth, "learned from traces" is a liability rather than a feature. Traces encode what happened, not what's true or approved. The failure mode is already visible: companies pointing a model at a year of call recordings and treating the output as positioning. Those recordings were captured inside a sales cycle. The prospect was negotiating, playing hard to get, or being polite. The rep was performing for a coaching tool. Mining that for "what buyers believe" is mining a negotiation for a testimonial. Whether the context was captured with the contributor's knowledge and consent is the first question I'd ask of any proof layer, and it's the first question on our own buyer's checklist.
Not all context is equal
Two distinctions fall out of this that the harvest framing flattens.
Provenance. Context carries the situation it was captured in, and the situation changes what it means. The Gong library is sales-cycle context: useful for coaching, usable for pipeline, and dangerous as a messaging source unless you discount for the negotiation it came from — which is the quiet limitation of the whole voice-of-customer software category, built to mine exhaust after the fact. An on-the-record interview, where the person chose to participate and knows the words will be used, is a different grade. The same sentence means different things depending on which room it was said in, and a context layer that doesn't index by provenance will average a negotiation with an endorsement and call it insight.
Audience. Some context is internal: you harness it, distill it, and by the time anything reaches the public it's a case study or a line on a landing page. But buyers increasingly send their AI to read you before they ever read you themselves, and that reader wants the context, not the distillation. How much more useful is your proof if the buyer's model can reach the full account — who said it, in what role, about which outcome — and skip the parts that don't apply to them? That is the liquid content argument from the other direction: the atoms have to be solid, because machines are now the ones recombining them.
Which is where the journalism analogy earns its place. A story built on anonymous tips can still be a story. A story with a named source, on the record, who has put their credibility behind the claim, is a different story — readers weight it differently, and so do editors deciding what runs. That discount now operates twice. Once on the human reading what their AI distilled from your marketing. And once, earlier, on the AI deciding which sources to retrieve and trust in the first place. Attributed, on-the-record context is more likely to be brought forward, and more likely to be believed when it is.
Provenance now determines both what gets retrieved and how much it's believed.
Startups have no exhaust — and that's the advantage
The harvest model assumes volume. The enterprise version of context is statistical: a year of recorded calls across the whole sales org, product analytics across thousands of accounts, adoption curves you can average over. You harness it by averaging over scale.
A startup under $10M ARR has none of that, for two reasons. First, there's just less of it — two or three dynamic reps closing enterprise deals on osmosis, in the room together, not diligently recording every call so someone can study it later. The collaboration happens in conversation and leaves no trace. Second, even where the recordings exist, you can't do statistics on thirty calls. The method that works at a thousand doesn't work at thirty; it produces confident averages of noise.
So a startup's context exists in three states, and almost none of it is in the one a harvest model can reach:
- Latent. In the founder's head and in the customers' heads. Real, specific, and never written down — because nothing forced it.
- Undecided. Positioning is a decision, not a fact. Who you're for, what you'll never claim, what you call the thing — these don't exist until someone rules on them. Most of a startup's context has to be decided into existence before it can be captured at all.
- Emitted. The exhaust. At this scale, barely exists, and what exists is too thin to average.
That looks like a disadvantage. I think it's the opposite, for one reason: a startup's context is small enough to be complete. Every claim the company makes can carry a proof state. Every number can trace to a source. Every customer quote can be on the record. An incumbent has volume without provenance — a decade of exhaust nobody adjudicated. A startup that installs the forcing function now, so that nothing ships without compiling from the canon and everything produced lands back in it, has provenance from interview one.
And the reason completeness beats coverage is the same reason I keep returning to: a human carrying a slightly wrong framework hedged in real time. A rep heard the line not landing and adjusted mid-sentence. A model does not hedge. It takes whatever context it was given and renders it fluently, consistently, across every channel, at volume. At that point an incomplete context layer is a reliably wrong one — and a complete one, however small, is a reliably right one.
The markdown snapshot: the fair objection
The sharpest objection to all of this is that a written process document is a snapshot — real context is learned from thousands of workflows and compounds; a markdown file can't. I run a company on markdown files, so I should answer this directly.
Half of it is wrong. A governed canon is not a snapshot. It has a change ledger, it carries a proof state on every claim, it marks retired claims as retired, and the proof layer underneath it is served live rather than pasted in. A deck is a snapshot. A versioned, adjudicated body of claims with its sources attached is closer to a repo, and repos compound.
Half of it is right, and I'd rather say so than pretend. What we don't yet have is the outcome trace loop — which proof was used in which deal, what it moved, what converted. That is the feedback execution traces provide for process context, and it's the thing that would let captured context compound the way harvested context does. We're building toward it. It isn't built. The credibility of everything else in this piece depends on not overclaiming that part.
The investor version: where "Context is King" lands
The clearest investor-side statement of the context-layer thesis is Evan Armstrong's "Context is King" at The Leverage, from February. His version: software is splitting into three layers — systems of record, point applications, and a new context layer between them — and the first two are commoditizing. He reaches for Christensen's conservation of attractive profits to explain why the margin moves to the middle.
His cleanest line is that systems of record store data, not meaning. Ours is the operational version: the CRM records the transaction, not the reasoning. Closed-Won tells you what happened. It doesn't tell you what the buyer was afraid of, what they'd tried before, or the phrase they used for the problem that your product team has never said out loud. I've made that argument at length in the context engineering piece; Armstrong arrives at the same place from the investor side, which is useful confirmation that it isn't just a practitioner's complaint.
His second strong point is that the context layer takes its money from payroll, not from IT budgets. Agreed, and we already price that way: the comparison for a fractional go-to-market partner isn't another tool line, it's the product marketing hire a company under $10M ARR can't justify yet.
Where I'd move his boundary is the source. Every example of context in his piece is exhaust — threads, docs, and the traces agents leave behind — which is the half described above. The go-to-market half was never emitted, so it isn't in his layer at all until someone captures it.
Two years from now: do you have anything to harness?
Project forward. Every quarter brings another model release. The tooling gets easier, migration gets easier, and the ability to harness context — to point a model at a pile of records and get something coherent out — becomes a feature of everything, available to everyone. That's where the value is going. But harnessing is itself going to commoditize, the same way generating the interface did. In specialized domains it'll stay hard for a while. For go-to-market, it won't.
At that point the question isn't whether you can harness context. It's whether you have anything to harness. Whether the reasons your customers bought exist anywhere outside their heads. Whether the claims your company runs on were ever decided, let alone proved. Whether a buyer's AI, reading you cold, finds named people on the record or finds marketing copy.
When harnessing context is free, having context is the moat.
The term is being claimed by the data stack right now — search "context layer" and you'll find semantic layers for data agents, which is a real thing and not what I'm describing. The go-to-market half is unclaimed. It is the half with no exhaust, no forcing function, and no vendor whose product fills it as a byproduct. It has to be captured on purpose, by someone whose job it is, from a founder who can't self-extract and customers who will only say the sharp version to someone who isn't their account manager. That is the work. The practitioner pieces on how to do it — the founder session, and voice-of-customer research as the product marketer's actual job — are the next two in this series.
I write about building an AI-native company at Proofmap. The diagram above is free to use with attribution.
