Sales needs a context layer. Here is the architecture.

Run one test on whatever your sales team uses today. Ask it: what did this prospect object to on our last call, and did we follow up?

Almost nothing can answer. Not because the software is bad — most of it is excellent at its own job — but because the question spans the inbox, the call recording, the shared drive and the record, and no single system holds all four. The unit of work in sales is the deal. Every tool that touches it holds one slice and its own private memory of that slice. The stitching between them is performed by a human, usually your best-paid one.

Then AI arrived and inherited the fragmentation rather than fixing it. Each surface got an assistant that could see only that surface. The question in 2026 is no longer who has AI. Everyone has AI. The question is what the AI can see — because an agent grounded in a partial picture produces partial work, and partial work has to be checked by a human, which is where the promised time goes.

The order that works

Systems that hold up under this converge on the same shape, and the convergence is not a coincidence — it falls out of what an agent needs in order to be accountable.

Content has to be indexed together with its metadata, its relationships and its access controls, and queried through that structure rather than a flat pile of vectors. That is what makes an answer citable and permission-safe instead of merely plausible. Then the rules have to be written down somewhere the system reads, so an output is predictable without a human checking every line. Only then is there anything for an agent to be held to.

Which gives one order, every time:

Context, then rules, then agents. Never the reverse.

That order is the entire lesson. Most AI products are built backwards — agent first, context bolted on afterwards through whatever connectors are available — and that is precisely the shape that produces output nobody can stand behind. Neither of these companies leads with "we have agents." They lead with "we hold the truth of your work." The agents are the payoff, not the pitch.

What changes when you point it at sales

We took the anatomy and changed two things.

The unit of context is the deal, not the enterprise. A horizontal graph is optimized to find and answer across everything a company knows. Sales does not primarily need to find things. It needs a specific, bounded object — this deal, with its thread, its calls, its documents, its commitments and its objections — assembled completely.

The agent's job is to finish an artifact, not to answer. Sales output is artifact-shaped and finite: the follow-up, the pre-call brief, the deck, the invoice, the diagnosis. Because the system knows what each finished artifact looks like, what governs it, and whether it worked, the agent can produce it rather than describe where to find one.

The three layers, as we built them

1 · Context. Email, meetings and documents captured natively — not read through a connector we do not control — parsed, chunked and embedded into one deal-scoped graph, with a growing map of the content in connected apps alongside it. Permissions are enforced at retrieval, not filtered after generation, so a model never assembles context a user could not open themselves. Every generated claim carries a reference back to the sentence in the contract or the moment in the call that produced it.

2 · Rules. The Brain: the skills and policy your team writes once — how you pitch, what you never discount, which proof point closes which buyer, which claims are approved — governing every output the system produces. Authored by the customer, scoped to a team or the org, and reviewed before it governs anything. A skill still in draft governs nothing.

3 · Agents. They act on both and produce the work itself: the call prepped, the reply drafted in your voice, the deck built, the invoice sent. Each run is a durable session — what triggered it, the context it read, the rules that governed it, what it produced, and who approved it. Nothing leaves the building on its own.

Why the order is the moat

The agent layer is the visible one. It is also the replaceable one — models improve monthly and everyone gets the same ones. What does not transfer is underneath: a permissioned, relationship-aware corpus of your deals, and the rules your team wrote themselves.

One graph buys a second thing that compounds. Because the system holds the reply, the call and the closed-won outcome, it can learn which objection answer preceded a win rather than optimizing for a proxy like open rate. Most knowledge work has no ground truth. Sales does: the deal closed or it did not. That is a rare and honest label, and it is the customer's own data doing the work — not our roadmap.

Nothing activates on its own. Every learning lands as a proposal a human approves or rejects; a rejected one is never re-proposed.

The honest limits

A manifesto that overclaims is just marketing. We are earlier than the horizontal players, and our map of third-party apps is growing, not complete — we say growing because completeness is the one claim we could not defend. The record updates itself with high-confidence changes applied automatically and the rep approves the rest; nobody should promise you "no manual entry, ever." And the superiority is not in any single surface. It is in the one thing every surface shares.

Context, then rules, then agents — built for the deal. That is the whole architecture, and the order is the part that matters.

See how the Workspace is built →