Operating & monitoring the Brain

As an agency you get operator tools that clients don't - a cross-client Brain Health page and a per-client Brain ops page - so you can keep every account's knowledge sharp. This article also explains how the Brain works underneath, so the numbers on those pages mean something.

Brain Health (across all clients)

From the agency portal's More menu. The (i) beside the page title opens this article.

Four tiles across the top:

• Clients - how many accounts the page is reading.
• What the brains know - live pieces of knowledge across the whole book, with how many of them are key facts (the anchors that are always included when the Brain is consulted).
• Open questions - things the Brains are unsure about and waiting for an answer on.
• Clients needing a look - how many carry an amber status flag (below). Grey flags are not counted: they explain a figure rather than ask you to do something.

Beneath that, one row per client: Client, What it knows, Key facts, Open questions, Times used (30d) (how often an AI surface drew on this Brain in the last month), Last learned (when it last picked something up) and Status. A What it knows, by kind breakdown shows the shape of the knowledge - how much of it is facts, moments, goals, rules and so on (see "The kinds of knowledge" below).

Status flags

Four flags, in plain words, in two colours. Amber is something that should not be true and is worth acting on. Grey is the page explaining why a figure that looks alarming is not - nothing to do. A client with no flags at all reads Healthy.

Amber:

• Nothing new this month - a data refresh was due and nothing landed. Start at Connection Health: a feed has usually stopped.
• Nothing learned yet - an empty Brain. Connect the client's platforms and run the first refresh, then Rebuild from what we hold (below).
• Nothing learned in 30 days - the Brain is live but nothing new has reached it in a month. Check the feeds are moving and that the account is actually being used.

Grey:

• Not refreshed automatically - no schedule selects this client (their cadence is manual, or their account is paused), so old data is not a fault. Run a refresh by hand from their Manage page if you want the Brain fed, or move them to weekly or daily.

A client can carry both, and amber is always listed first.

Occasionally the page shows a note saying no numbers were found for any client. When that happens Nothing new this month is withheld from every row rather than shown on all of them: an empty answer across the whole book is as likely to mean the page cannot read the numbers as it is to mean none arrived. Connection Health is the place to look.

Embedding spend, confidence calibration and question-dismissal patterns are no longer on this page; they belong to the platform team's own tooling.

Brain ops (per client)

Every client has a Brain ops page, reached from the Open Brain ops link under Advanced on their settings page. Its tabs let you ask the Brain directly, read what it has been asked and what it answered with, browse what it knows, follow the change log and see the graph. The Operations tab holds the four by-hand actions, each with a "when to run" line beneath it. Most clients never need them after onboarding.

• Rebuild from what we hold - seed the Brain from the structured data already in the account (brand identity, ideal customer, tracked keywords, goals and the like). Run once after onboarding, and again if those change in bulk.
• Reflect now - make the Brain read its recent signals straight away instead of waiting for the next scheduled pass. Useful when a lot has just happened. It runs itself at most every ten minutes, so pressing it repeatedly does nothing extra.
• Rebuild the links - connect things the Brain knows that mean similar things, even when they share no words. Run after a rebuild, or once fifty or so new things have been learned. Safe to run again.
• Pull past months - fetch each past month's numbers from the connected platforms so trends have history to compare against. Run once, after the platforms are connected and the first refresh has run. Running it again just refreshes the same months.

How the Brain works

The model

Every client has its own knowledge graph: discrete pieces of knowledge, typed relationships between them, and a meaning-based index for each piece. It is the single source of context behind every AI surface - reports, dashboard insights, chat, the web and voice concierge, and the ad and copy studios.

What it watches

The Brain turns activity into observations. Sources include comments on metrics, reports, insights and recommendations; free-text notes; social posts, approved and rejected; calendar events; goal changes; brand-identity edits; report edits; monthly storylines; copywriter drafts and publishes; approved entries from uploaded documents; and the general activity feed. Data pulls from the connected platforms (analytics, ads, CRM, email) also land directly as structured knowledge.

When it learns (reflection)

When new signal arrives - at most once every ten minutes per client - and on a guaranteed weekly pass, a reflection runs: the AI reads the recent observations alongside the relevant existing knowledge and proposes new pieces of knowledge, new relationships and clarifying questions. Anything it is confident about is saved as knowledge. Anything it is less sure of becomes a question for your team or the client to answer, and answering promotes it to knowledge. Some things are written without a reflection at all: monthly metric snapshots, CRM and email absorption, and the by-hand operations above.

The kinds of knowledge

Every piece of knowledge has a kind, which is what keeps retrieval coherent: an entity (a channel, product, competitor, person or audience), a fact (a stable present truth), an event (something that happened), a goal, a constraint (a rule or limit), a hypothesis (an unproven theory), a decision (a deliberate choice and its rationale), a lesson (a proven, repeated learning), a metric snapshot (a KPI value at a date) and a note (the last-resort catch-all). Clients see five of these under plainer names - Detail, Rule, Theory, Moment and Milestone - on their own dashboard; your portal uses the names above.

Finding by meaning

When a piece of knowledge is written, its text is indexed by meaning as well as by words. That is what lets the Brain find things that are about the same subject even when they share no vocabulary, and it is what Rebuild the links works from. A weekly sweep re-indexes anything that was missed.

Linking

Relationships are typed - caused, influences, relates to, is an instance of, contradicts, mentions. Reflection proposes them explicitly, and the Brain also infers links mechanically and from meaning. Links are what let the Brain pull a coherent cluster rather than isolated facts: asking about Meta surfaces the Meta channel plus its recent moments plus the Meta spend goal. When something new is near-identical to something it already knows of the same kind, it is merged, not duplicated - confidence goes up and every source that confirmed it is recorded.

How it is used (retrieval)

When an AI surface needs context, the Brain assembles the most relevant slice: a search by meaning for the subject in hand, the knowledge linked to those hits, plus the key facts, which are always included - all within a fixed budget so the answer stays focused. Times used (30d) on the Brain Health page counts how often that happened for each client.

When to step in

• A client's Brain says Nothing learned yet - connect their platforms, run the first refresh, then Rebuild from what we hold.
• Nothing new this month - a feed has stopped; fix it in Connection Health and the next refresh will feed the Brain again.
• Open questions are piling up for one client - answer them from their Manage page (Pulse zone), or make the ones only the client can answer visible to them.
• Search on the client's dashboard is missing things you know are there - run Rebuild the links.
• A review is sitting in a client's questions - older records to retire after the client said something replaces them, or records that rested on a document point someone turned down. These are not answered in words: open it, tick what should go and confirm. Nothing in a review is retired until someone does, and Undo this review puts everything back. See Brain questions & feedback.

See Adding & managing clients for the rest of the per-client controls.