There is a gap at the centre of workplace AI that everyone has noticed and few have named.
Adoption is essentially settled. Nearly everyone doing knowledge work now uses AI daily, and most of them will tell you it makes them faster. Ask the same people whether their company is measurably better off, and the answer thins out immediately. The hours are being saved. They are not arriving anywhere.
Watch where they go and the mechanism is obvious. People re-prompt, re-check and re-do. They keep three or four assistants open because none of them knows what the others know. And a great many of them ship work they have not fully verified — not out of carelessness, but because verifying it properly would cost more than the draft saved.
That last part is the tell. It is not a story about models being bad. Models are extraordinary. It is a story about what happens when you give an extraordinary model no reliable context, no rules it must obey, and no way to prove where its answer came from. You get output that is fluent, plausible, and unaccountable — so a human has to sit next to it, forever, checking. The saved hours never leave the building. They just change hands.
The architecture that answers this has been converging quietly for about eighteen months, and it now has a name.
The definition
An Agentic Workspace is a system with three layers, in this order:
1 · Context. The team's real working material — email, meetings, documents, and the content in connected apps — captured into one graph, with its metadata, its relationships, and crucially its permissions. Not a folder of files. A map of who said what to whom, attached to what, about which piece of work.
2 · Rules. The skills and policy the team writes once — how we pitch, what we never discount, which proof point closes which buyer, which claims legal has approved — that govern every output the system produces. Written by the customer, not inferred by the vendor.
3 · Agents. Software that acts on both and produces the work itself: the drafted reply, the meeting summary, the deck, the invoice. Not an answer about the work. The work.
The order is the whole argument. Most AI products are built in reverse — agent first, context bolted on later through whatever connectors are available — and that is precisely the shape that produces work nobody can sign off on. An agent with borrowed context has to be checked line by line, because nobody can say what it saw.
A copilot answers questions about your tools. An Agentic Workspace is where the work gets done.
Why the order is the whole argument
Most AI products are built backwards — agent first, context bolted on afterwards through whatever connectors happen to be available. That is precisely the shape that produces output nobody can stand behind, because nobody can say what the agent actually saw.
Built in the right order, each layer earns the next. Context indexed with its metadata and its permissions — rather than dumped into a flat pile of vectors — is what makes retrieval both safe and citable. Rules written by the team are what make an output predictable enough to trust without reading every line. Only then does an agent have something to be accountable to.
It also settles which part is worth building. The agent layer is the visible one, and it is the replaceable one: models improve monthly and everybody gets the same ones. The permissioned, relationship-aware corpus underneath, and the rules a customer wrote themselves, are what cannot be reproduced in a quarter.
Horizontal finds. Vertical finishes.
The difference between the two is not size. It is what the system is able to complete.
A horizontal workspace is optimized to find and answer across everything. That is genuinely valuable, and for a 40,000-person company with a thousand internal tools it may be the only tractable approach. But its agents automate generic processes, because generic is all a system can be when it must serve legal, engineering, support and finance with the same primitives.
A vertical Agentic Workspace owns one workflow end to end. That constraint is the feature. Because it knows what the finished artifact looks like, what governs it, and whether it worked, its agents can produce the artifact rather than describe where to find one. A horizontal system tells you where the answer is. A vertical one ships the follow-up.
Why sales is the right vertical
Three reasons sales is unusually well suited to this architecture — and, we would argue, the most under-served function of the three.
The output is artifact-shaped. A rep does not primarily need to find things. They need things produced: the follow-up that references what the buyer said on Tuesday, the pre-call brief, the tailored deck, the invoice, the renewal summary. Those are finite, repeatable artifacts with a known shape — exactly what an agent layer can be held accountable for.
The feedback signal is unusually clean. Most knowledge work has no ground truth. Sales does: the deal closed or it did not. That means a sales workspace can actually learn which objection answer preceded a win and which preceded silence, instead of optimizing for a proxy like open rate. Very few domains hand you a label that honest.
The context is fragmented by construction. The unit of work — the deal — spans the inbox, the calendar, the call recording, the shared drive, the contract and the CRM record. Six systems, six memories, and a human paid to be the integration layer between them. We have argued this at length elsewhere: no single-surface tool has ever seen the whole deal, which is why none of them can answer the simplest question a manager asks.
What separates one from a chatbot with a search index
The term will be diluted within a year — every vendor with a text box will claim it. Five questions cut through, and each maps to one of the failures described at the top of this piece:
1 · Is the context captured natively, or borrowed? A system that reads your email through a connector it does not own has a partial, delayed, revocable view. Native capture is the difference between a graph and a cache.
2 · Are permissions applied at retrieval? Not filtered after generation — enforced at the point the context is fetched. If a model can see a document the user cannot, the product is one prompt away from a data incident.
3 · Can every claim be traced to its source? This is the direct antidote to shipping work nobody can stand behind. If a generated line cannot be clicked back to the sentence in the contract or the moment in the call that produced it, verification stays a human job, and the hours stay lost.
4 · Who writes the rules? Horizontal tools infer preferences from behavior. An Agentic Workspace lets the team author the discount floor, the approved claims, the objection answers that actually work — and then governs every output by them. That authored layer is also the part that does not transfer to a competitor.
5 · Does the system hold the outcome? A tool that drafts an email and never learns whether the deal advanced writes a first draft forever. Holding the reply, the call and the closed-won result is what turns usage into compounding.
A product that fails questions three through five is a chat interface over a search index. It will save an individual real hours and hand every one of them back as checking.
Where Agent07 sits
Agent07 is the Agentic Workspace for sales. The same three layers, pointed at one function.
Context: every email, meeting and document captured natively into one deal-scoped, permissions-aware context layer — plus a growing map of the content living in connected apps. Rules: the Brain, where your team writes the skills and policy once and every output obeys them. Agents: they prep the call, draft the reply in your voice, build the deck, send the invoice — grounded in your real context, referenced back to source, and yours to approve before anything leaves the building.
That Workspace is the core product. Auto-CRM and Outbound are extensions that run on the same graph — and the reason the CRM can write itself is simply that the Workspace already holds everything the record was asking a human to re-type. The pipeline stops being a chore and becomes an output. Leadership gets the same benefit for free: reporting assembles itself out of context that was captured anyway, rather than out of a Friday request to eight reps.
We are deliberate about what we do not claim. We are earlier than the horizontal platforms, and our map of third-party apps is growing, not complete — we say growing, because the alternative is the one claim we could not defend. Nothing an agent learns activates on its own; every learning lands as a proposal a human approves or rejects. And the honest promise on time is the one we can show: reps get their week back from admin, not a magic multiplier.
The category does not need another AI feature. It needs a place where the context, the rules and the work finally live together — and for sales, that place has to be built for the deal.
See the Agentic Workspace →