How does AI work across two companies?

Cross-company AI works by making the boundary between organizations a first-class object in the system. The agent runs inside a sanctioned shared environment with per-company identity, field-level data governance enforced from both sides, and an open wire protocol — typically Model Context Protocol (MCP) — that lets external agents invoke capabilities without a custom integration project per partner. AWS's Partner Central agents MCP server, launched March 2026, is the cleanest public example: SigV4 IAM authentication per partner, data isolated per tenant, and human-in-the-loop approval before any write.

Why don't general-purpose AI tools work for partnerships?

General-purpose AI tools assume a single tenant — one company's data, one company's users, one company's policies. Partnership motions break that assumption at every step: account mapping, co-sell execution, joint planning, and funding eligibility all involve two companies' data and two companies' permissions simultaneously. Without cross-company awareness, an AI agent either over-shares (a trust violation) or under-shares (no useful action). The architecture has to model the boundary; it is not a UI concern.

What is cross-company AI, and why single-tenant assumptions break partnership work

Q: What is cross-company AI?

Cross-company AI is the architectural property of an AI system that knows it is operating across more than one organization at once. Instead of assuming a single tenant — one company's data, users, and policies — the system tracks which side of the trust boundary it is on, what it is permitted to read from the other side, and what it is permitted to write where. It is the design pattern partnership work requires, because every partner motion involves at least two companies and often three.

General-purpose AI tools assume one tenant: one company’s data, one company’s users, one company’s policies. Partnership work fails this assumption at every step. Every motion involves at least two organizations, often three (the customer is its own party).

Cross-company awareness is the architectural property: the AI knows which side of the boundary it’s on, what it’s permitted to read from the other side, and what it can write where. This is not a UI concern; it’s a system concern. It changes how data is modeled, how identities are resolved, how actions are logged.

It’s also the property that makes the Partner Revenue Platform a category, not a feature. CRMs are single-tenant. Portals are read-only. Only an execution layer designed for two-company motions can actually do this.

What practitioners ask

“What is cross-company AI?”
“How does AI work across two companies?”
“Why don’t general-purpose AI tools work for partnerships?”

The answer

Cross-company AI is the architectural property where an AI system knows it is operating across more than one organization at once — and tracks which side of the boundary it is on, what it is permitted to read from the other side, and what it is permitted to write where. It is the inverse of the single-tenant assumption baked into general-purpose AI tools, and it is the constraint partnership work imposes on every motion.

The clearest production example is AWS Partner Central agents, launched March 16, 2026. AWS shipped a fully managed Model Context Protocol (MCP) server that lets external partner-side agents invoke AWS Partner Central capabilities — pipeline insights, opportunity progression, funding eligibility — using SigV4 IAM authentication per partner, per-tenant data isolation, and required human-in-the-loop approval before any write. That trust model is the parallel architecture: the agent is allowed to act inside the boundary because the boundary is enforced at the protocol layer, not the UI layer.

MCP itself is the wire format. The open specification defines how clients, servers, and hosts negotiate capabilities and exchange context — but it does not solve the trust boundary by itself. Cross-company awareness is the design property built on top of MCP that says: this agent is acting across two companies, and the system has to know which is which. Without that property, an MCP-connected agent either over-shares (a trust violation) or under-shares (no useful action).

This is why partner data sharing is one of the top blockers Forrester identifies in its 2025 ecosystem research — the architecture to do it safely did not exist at scale. The Partner Revenue Platform category exists because CRMs are single-tenant by design and portals are read-only. Only an execution layer purpose-built for two-company motion — like WorkSpan’s Partner Revenue Platform, which Boomi used to scale 1,200 co-sell opportunities with AWS in six months — can run the cross-company motion in production.

Use this framework

Five tests for whether an AI system is cross-company aware

1. IDENTITY PER COMPANY
   Does the agent authenticate as a different principal on each side of the
   boundary, or does it run as a single identity that pretends to act for two?

2. READ POLICY PER SIDE
   Is what the agent can read from Company A's data different from what it
   can read from Company B's? Or does the system collapse both into one tenant?

3. WRITE POLICY PER SIDE
   Where does the agent's output land? CRM-A only, CRM-B only, both with
   provenance, or a shared record? Single-tenant systems can't answer this.

4. HUMAN-IN-THE-LOOP ON CROSS-BOUNDARY WRITES
   Before the agent commits a write that crosses the boundary, does a human
   on the affected side approve? AWS's MCP server requires this. So should
   any cross-company architecture.

5. PROTOCOL OVER PRIVATE INTEGRATION
   Is the surface MCP-native (or an equivalent open protocol), or is each
   partner connection a one-off integration project? Open protocol is what
   makes the architecture buildable at ecosystem scale.

Shared Environment — the sanctioned space where cross-company agents can act
MCP / Open Protocols — the wire format that makes the motion buildable
AWS Partner Central + MCP — the parallel architecture, in production
Partner Revenue Platform — the category cross-company awareness defines
80T Ecosystem — the market force that makes single-tenant AI insufficient

What practitioners ask

The answer

Use this framework

Related concepts

Sources