Cross-agent positioning consistency: govern what your AI says

A typical revenue team in 2026 runs five to ten AI agents at once — an outbound SDR, a CRM assistant, a renewal agent, a content generator. Each was bought separately, and each carries its own copy of what the company says about pricing, competitors, and product. The RevOps lead is left holding the question nobody wants on the renewal call: why is our AI SDR sending different positioning than our website?

The Commercial Truth manifesto argues that marketing has never had infrastructure — no shared layer the way engineering has a repository and finance has a ledger. The AI revenue stack inherited the same gap. Every agent vendor ships its own knowledge layer; none of them is the neutral source the others read from.

The cost is not hypothetical. The SDR quotes a discount the deal desk never approved; the renewal agent cites a feature deprecated last quarter; five tools give five versions of the same answer. Each is a chance to say the wrong thing at machine speed.

Why agent governance tools don’t fix this

The instinct is to reach for the agent-ops tooling — the observability and orchestration layer that has matured fast. It is the wrong layer for this problem. Agent observability tells you what the agent said; it does not enforce what the agent should have said.

That distinction is the whole story. Observability governs whether the agent behaves — latency, structure, tool calls, guardrail trips. It is silent on whether the pricing the agent just quoted is the pricing the company actually charges.

Walk the three failures and the pattern holds. The SDR quotes an unapproved discount and the deal stalls; the renewal agent cites a feature deprecated last quarter and trust erodes on the call; the CRM agent surfaces an outdated competitive line and the loss gets filed under “positioning.” Observability can tell you each happened. None of it could have stopped the claim, because none of it knows what the claim was supposed to be.

The retrieval workaround fails for the same reason. Give each vendor its own index of your decks and the indexes diverge the moment positioning changes — now you are hand-syncing five stores, and the drift between them is exactly the inconsistency you were trying to remove.

What changes when claims are the substrate

The fix is not a better agent. It is a layer beneath all of them. When pricing, positioning, and claims live in one typed, source-attributed graph, every agent reads the same governed truth — and a change to a claim reaches all of them on their next query.

This is the difference between governing behavior and governing substance. The agent vendors compete on how the agent acts; the substrate governs what is true. One is about the engine; the other is about the fuel.

It also closes the question Counsel keeps asking — can you certify what every AI tool in revenue told prospects last quarter. When every agent grounds in one audited source, that answer is a query, not a fire drill.

Why no agent vendor ships this

There is a structural reason the fix has not come from the agent vendors. Each one is built to be the intelligence for revenue, which means each is incentivized to own its knowledge layer, not to defer to a neutral one. A shared substrate beneath them is an admission none of them wants to make.

The observability and orchestration tools sit one layer below the problem — they route and trace; they do not hold canonical commercial knowledge. The CRM platforms have the customer base but not the claim-level data model, so they are likelier to consume the substrate than rebuild around it. The gap stays open because no incumbent’s incentives point at filling it.

The concrete capability

Concretely, the agents query the canon through one Model Context Protocol layer — the Orchestration surface that already exists. The integration is vendor-neutral: a custom agent or an embedded CRM assistant connects to the same endpoints and receives typed, source-attributed, confidence-scored claims rather than a scraped page.

Three things sit on top of that bus. A policy layer can refuse a claim before it ships — no agent quotes pricing without querying the canon, no agent emits a claim flagged high-risk in the EU without human pre-approval. A coverage view shows every connected agent and where each agrees or diverges from the canon, with a kill switch to pause one agent or all of them at once.

For RevOps, this is the part that matters: the agents stop being black boxes. The segment definitions stay yours, every deployment is visible, every claim is logged — and the agents agree because they are reading the same source.

For Counsel, the same layer produces what procurement keeps asking for: a record, on demand, of what every AI agent in revenue said over the last year, why it said it, and which canonical claim it was grounded in. The control plane is where that artifact comes from.

Closes / opens

The question is shifting from how many agents can we deploy to how do we govern what they say. Agent-ops tooling answers the first. Only a shared substrate answers the second.

The methodology Assay is developing for the Commercial Truth Index scores this directly — cross-surface consistency, propagation lag, and version skew across every surface a vendor publishes to. A revenue stack where ten agents give one answer scores well. Ten agents and ten answers do not.

Closes the multi-agent cluster’s consistency seat (LSO §F.18): governing what agents say, not just whether they behave. Opens the ownership question — does the substrate belong to RevOps, who runs the stack, or to the positioning owner, who authors the canon?

This essay is grounded in Assay’s agent control plane spec and the Apps alignment sub-pillar, with buyer context from the brand canon. Methodology for the Commercial Truth Index is in development.