GTM engineering · Field noteGTMhub method

A connector still needs company context

Connecting an assistant to a system makes data available. The company still needs definitions, decision rules and ownership before that access becomes useful GTM work.

The source and the change

Anthropic introduced the Model Context Protocol on 25 November 2024 as an open standard for connecting AI assistants with systems that hold data. The announcement addressed the integration problem caused by isolated sources. Original Anthropic announcement.

The decision for a B2B team

A connection can make a CRM record available without explaining which company definition or rule the assistant should use. “Find our best accounts” still needs an account unit, exclusions, a fit model, signal evidence and a named next action. Access is one component of the job, not its commercial specification.

Our recommendation is to define the smallest useful output before adding another connector. For inbound routing, that may be a valid owner and an account brief. The exceptions are equally important: an existing customer, an open deal, ambiguous company identity or an unavailable owner.

Put the definition next to the connection

LayerDecision to resolveExample output
Company contextWhat counts as fit and pipeline?Shared definitions and exclusions
Skill and lane pluginHow does this bounded job work?Input/output contract and stop conditions
ConnectorWhich records can the job read or change?Scoped access and field mapping
MemoryWhat did the team correct or accept?Dated decision and exception history

Keep one brain per company. Different lanes can use different tools while sharing the commercial definitions. A paid-media plugin and an outbound plugin should not disagree about the same account or opportunity because each was given a separate prompt.

A practical review before activation

Replay known examples with a duplicate, a missing field and a conflicting owner. Inspect the record the operator receives. A successful connector response is not the acceptance criterion; the next person must be able to use the output without reconstructing the account.

Make writes distinct from recommendations. Preserve stable event IDs, field ownership and a rollback path. An exclusion or unresolved identity should stop activation, while a temporary enrichment failure may take a defined fallback route.

What the source does not promise

The 2024 announcement establishes the protocol’s original introduction. It does not prove that every provider connection is supported today or that a connected assistant can run GTM without company rules. Check the selected implementation’s current documentation and permissions. Start with the company brain guide and use the open starter to make the context and job contract inspectable.

Publication and source record

Published on GTMhub: . Last reviewed: .

Original source: Introducing the Model Context Protocol, . This is a retrospective GTMhub analysis of that source.

How we run this

Put it to work in your company.

Our GTM engineering lane plugs into your company’s GTM brain. Start with one lane; services can run in parallel.

Book a call
  1. Week 1Your GTM brain

    Monday kickoff. A 45-minute review Friday to confirm the context.

  2. Weeks 2–3Connect the lane plugin

    Wire it into your existing tools and test the work with your team.

  3. Week 4Live, then improve

    About three hours of your time in month one, then 15 minutes a week.

The first term is three months. The brain and lane plugin are handed over in full after it. Your accounts and data stay yours.