The Exhaust Brain and the Modeled Brain

Eric Siu's thread on how his agency built a "Single Company Brain" is the clearest articulation of a pattern we have been tracking for two months. The diagrams are genuinely good. The argument is mostly right. And it contains, almost in passing, the most important admission in the whole genre: "This is where a lot of AI systems quietly fail. They look smart in a demo because the context is hand-fed. Then they fall apart in production because nobody built the retrieval layer."

He is correct. Then he doesn't build it either. The retrieval section of his architecture describes what the agent should pull — "the six pieces of context that matter for the task in front of it" — and says nothing about how. No index, no ranking, no entities, no schema, no join. Retrieval is the load-bearing layer and it is drawn as an empty box with a nice border.

That empty box is not a flaw in Siu's writeup. It is the entire question of the agentic-data era, sitting unanswered in the middle of the most-shared company-brain diagram of the month. What you put under it is the difference between two fundamentally different machines.

1. Two brains that share a diagram

Siu's stack — capture, retrieval, source-truth, permission, feedback, execution — is built over organizational exhaust: 2,862 Gong call transcripts, Slack threads, SOPs, CRM activity, daily logs. The unit of knowledge is the extracted fragment: "15 calls produced 390 insights, 470 facts, and 125 frameworks." There is no customer entity, no metric definition, no model. It is a retrieval-augmented index over everything the company has ever said.

The brain we build at an analytics shop shares the same outline and almost none of the substrate. It is architected around governed data models — a dbt semantic layer, metric definitions enriched with business meaning, canonical entities derived once and propagated, a knowledge graph over the whole thing. Its retrieval is a traversal over governed structure, and its metrics are computed, not recalled. Call the first the exhaust brain and the second the modeled brain. They look alike on a slide. They fail in opposite directions.

2. Why the empty box is load-bearing

Look at what sits directly above retrieval in Siu's stack: a Source Truth layer whose job is to decide, at runtime, which conflicting fragment wins. "Which source wins? The sales call? The CRM field? The Slack correction? The old SOP? If you don't answer that, your agents become confident liars with better formatting." That layer exists because the substrate underneath it is unstructured. When your knowledge is a pile of fragments that disagree, you need a heuristic — usually an LLM — to adjudicate them every time a question is asked.

A modeled brain does not make "which source wins" disappear — refunds, chargebacks, slowly-changing dimensions, late-arriving facts, and legal-vs-operating definitions all force some arbitration at runtime, scoped and time-bound. What the model does is resolve the structural disagreements once, at design time, so only the genuinely temporal ones remain. The readable value of a coded field is derived a single time in a shared macro and propagated downstream; the definition of revenue is committed in YAML and reviewed in a pull request. The adjudication that can be settled in advance, is. Siu's Source Truth layer is doing the same job with none of that settled — it pushes every disagreement, structural and temporal alike, to a per-query language-model judgment. That is an honest, well-engineered response to having no schema, but it is adjudicating from scratch every time. His own phrase for the failure mode it patches, "confident liars with better formatting," is the precise sound of a brain doing all its arbitration at runtime with nothing resolved underneath.

A schema is a set of disagreements you resolved in advance so nobody has to relitigate them at query time. The exhaust brain pushes every one of those disagreements to runtime and asks a language model to settle them, confidently, on the spot.

3. The grounding the exhaust brain cannot see

There is a deeper gap, and Sonal Goyal names it. A semantic layer answers a definition — "what does revenue mean?" — deterministically. It does not answer a definite value — "who is this customer?" Her example is exact: an agent told to nudge accounts dormant for 60 days fires three renewal sequences at one enterprise because that account exists as three near-duplicate records. "The semantic layer was perfect. The entity underneath it was broken."

This is the part we have been circling internally for weeks — writers mapping the analytics context layer keep landing on entity-level identity resolution as the unaddressed layer, and it surfaced in our own reading notes well before Goyal's piece arrived. Our meta-context schema starts at the metric and quietly assumes the entity beneath it is already clean. Goyal's framing sharpens the stakes: when agents only query, clean entities are a nicety; when agents act, the customer ID has to map to one real-world person across every system the agent touches, or the agent does the wrong thing repeatedly, at machine speed, before anyone notices.

Here is the asymmetry that answers Keith's original question. A modeled brain has a partial grounding problem: it has governed definitions and is missing a first-class, agent-ready entity-resolution layer (Goyal's MDM-for-agents). The exhaust brain has a total grounding problem: it has neither definitions nor entities, and — this is the crucial part — it cannot even perceive the entity-resolution gap, because nothing in a pile of call transcripts knows that "Samsung," "Samsung Pay," and "the Samsung account" are the same governed thing. The modeled brain is one layer short of complete. The exhaust brain is not on the same map.

Animesh Kumar's model-as-contract piece is the constructive other half. "The model is not a representation of the contract. The model is the contract." A model written as SQL with embedded tests and constraints is "a declarative, enforceable statement of what the data must be" — it encodes that customer means an entity with at least one completed transaction, that revenue means recognized revenue after refunds. That is exactly the meaning an exhaust brain has to guess from context every time, and exactly the meaning a modeled brain states once and computes against forever.

4. Knowledge decisions and measurement decisions

The two brains are not competitors so much as instruments tuned to different questions. The cut that matters is the kind of decision the brain is asked to support.

The exhaust brain is excellent — arguably better than ours — at the left end. Drafting an email, surfacing the objection that worked last quarter, keeping content on-voice: the output is prose, and prose needs to be on-strategy, not numerically exact. Unstructured exhaust is the right substrate for that, and Siu's results are real. The wall on the right is subtler than "the exhaust brain can't do numbers." Wired to tools, it can take a real run at measurement — query the CRM, sample documents, classify a batch of transcripts — and for provisional, diagnostic, exploratory questions ("why did pipeline risk jump last week?") that is often enough. The wall is authoritative measurement: a repeatable, audited number that will be acted on under consequences. "What is our net revenue retention by product cohort this quarter, and which accounts drive the variance?" is not a recall task and not a one-off estimate. The honest version of the limit is this: the moment the exhaust brain needs an authoritative number, it has to call a governed metric and a resolved entity — and at that instant it has stopped being an exhaust brain and become the synthesis we describe next. It can find a transcript where someone said a retention figure; to vouch for one it has to borrow the modeled brain's machinery.

This is the answer to the question that started this read — does our data-grounded approach have an advantage, given that business decisions ultimately need data? The advantage is real and it is conditional, and the condition is the decision type. For a marketing agency, whose load-bearing decisions cluster on the left, the exhaust brain is the correct build and the data structure would be overhead. For a payments business, whose load-bearing decisions — decline rates, dollar volume, disbursement timing, risk patterns — cluster hard on the right, the data structure is not a cost we carry. It is the only thing that lets the brain answer the questions that matter at all.

5. We have been watching this brain get built

Siu's thread is not the first sighting; it is the most clearly diagrammed entry in a cluster that has been forming since April. Sentra published a "company brain" manifesto diagnosing the same scattered-knowledge problem from the communications substrate. Karri Saarinen described Linear's agent as both a context store and an action layer, its prompt iterated collectively in Slack so the team's corrections fold back into the system. Tobi Lütke's account of Shopify's internal agent is the empirical receipt: its merge rate climbed from 36% to 77% in two months with no model change — the gain came entirely from people writing down what it should have known. Ramp's framing is the one-liner: treat context as the deliverable.

Read together, these point at something more precise than "structured beats unstructured." Shopify's jump is producer-maintained context layered onto an exhaust-style substrate — humans encoding meaning the corpus didn't carry. That is the same move a semantic layer makes, performed by hand on prose. The real axis underneath the two-brain picture is not structured vs. unstructured. It is grounded vs. scavenged — whether the meaning was encoded on purpose by someone who knew it, or inferred after the fact from exhaust. A schema is just the most durable, most computable way to encode meaning on purpose. The exhaust brain's best results (Siu's feedback loop, Shopify's merge rate) come from the moments it stops scavenging and starts grounding. We made the same bet earlier and pushed it all the way down to the data model, where it can be executed instead of recited. That is the producer-maintained-context bet, and the company-brain cluster is its strongest outside corroboration.

6. The binary dissolves into a three-layer stack

The two-brain picture is a useful door, not the room. Push on it and it resolves into something both camps are already moving toward — and an honest read has to say so, because treating the field's serious players as naive RAG would simply be wrong.

The enterprise "brain" vendors are not vector-search-over-Slack. Glean and Microsoft 365 Copilot both back retrieval with a permission-aware knowledge graph over people, content, and activity; Atlassian Rovo and Google's Gemini Enterprise do the same. Palantir Foundry's ontology is the closest prior art to what we keep describing — objects, properties, links, actions, and security in one modeled substrate, explicitly more than a thin semantic layer — and it is the one system we have taken apart before. These are graph-backed brains, not exhaust piles. What most of them still lack is the governed metric layer — the computable definition of revenue — which is exactly where the semantic-layer-for-agents camp lives: dbt's MCP server, Cube, Snowflake Cortex Analyst, Databricks Genie, all making one argument — an LLM should not infer business logic from raw table names; it should resolve metrics through governed semantic objects.

And Goyal's missing entity layer has a market too, freshly relabeled for agents: Tilores, Senzing, and Zingg sell runtime entity resolution, with the decades-old MDM incumbents (Reltio, Tamr, Informatica) behind them. "MDM for agents" is not a mature category yet — the buyers still think in terms of customer-360, KYC, and fraud — but the agents are making that old pain newly visible, exactly as she argues.

Stack the three findings and the frontier stops being exhaust or modeled. It is a three-layer substrate: RAG for evidence and the why, a semantic layer for governed computation, and an entity-resolution layer joining them — with Microsoft's GraphRAG and ontology-grounded retrieval as the techniques that stitch text and structure together. The cleanest test of where any system sits:

Exhaust brain: what has the organization said?
Modeled brain: what has the organization defined, computed, and authorized?
The mature brain: both — joined by resolved entities, permissions, and provenance.

There is a sharper version still, and it relocates the moat. The unit that actually has to be governed is not the metric, the document, or the retrieval result. It is the action. An agent fails when it acts on the wrong entity, with the wrong permission, from stale evidence, under an unverified definition. So the durable architecture is an action-centered graph — target entity, allowed operation, authoritative evidence, metric health, permission, lineage, feedback rule, rollback path. Read that list against what we have already built: canonical models, meta-context, a federated knowledge graph, the producer-maintained bet. We hold most of the load-bearing pieces for the right side. What the company-brain cluster adds is a standing reminder that the evidence layer — the why behind the numbers, the part that lives in calls and tickets — is a real layer we under-invest in, and that everything underneath is finally in service of a safe action, not a tidy answer. "The data structure is the moat" was too proud. The truer statement: governed operational context is the moat, and the modeled data is its load-bearing member for every decision that has to be measured.

7. Where to keep the pressure on

8. What we take from the exhaust brain

The modeled brain wins the argument it was built to win, but the exhaust brain is ahead of us on three things worth stealing outright.

Make the feedback loop a first-class layer. Siu names "corrections become rules" as its own layer; we do this in agent memory but never made it a product surface. For conversational BI, every time a stakeholder corrects an agent's reading of a metric, that correction should become meta-context — we have the schema for it and not yet the loop.

Retrieve the few pieces that matter, not the whole brain. Siu's "six pieces of context for the task in front of it" is the same finding our skillpack work reached from the other side: curated, phase-scoped context beats dumping everything in. His own warning — persistent memory grew until it ate 40% of the context window — is a lesson we learned independently and should keep learning loudly.

Treat permission as architecture. An agency's risk is leaking client context across workflows; a payments shop's is worse. We govern definitions but do not yet have workflow-level boundaries on what context an agent may see before it starts generating. That is a layer to build, not borrow.