Responsive Builds Governed Knowledge Into the Model Context Protocol (MCP)
Responsive has expanded its MCP, Model Context Protocol capabilities so that teams can reach and act on the knowledge it manages from inside ChatGPT, Microsoft Copilot and Claude rather than switching to a separate application. The company says customers have made more than 198,000 tool calls since launching its MCP server in May 2026, and that on more than 38,000 occasions teams moved work from an AI assistant into an active pursuit. The technical claim worth examining is narrower than the marketing around it: every action connects back to Responsive and the assistant enforces each user’s existing permissions.
What the Model Context Protocol is
The Model Context Protocol deserves a brief explanation because the release assumes familiarity. It is an open standard for connecting an AI assistant to external systems, so the assistant can call tools and retrieve context from those systems rather than relying only on what a user pastes into a prompt. In practice it lets an assistant query a content management system or an application without the vendor building a bespoke integration for each assistant.
That standardisation is why three assistants can be supported at once. Before the protocol, a vendor wanting ChatGPT, Copilot and Claude support would build three integrations with three authentication models and three maintenance burdens. With a protocol, the vendor builds one server and each assistant connects to it, which is the difference between a product decision and a platform decision, and between maintaining three integrations and maintaining one server.
Why the trust problem is real
The argument the release makes is that AI has made answers fast while trusted answers remain hard, and in the specific domain it serves that is not marketing.
Bid and proposal work produces documents that become contractual commitments. A buyer’s CISO may rely on a security questionnaire response, a regulator may audit a due diligence answer, and a flawed public-sector bid submission may disqualify a supplier outright. A general-purpose model produces a fluent answer in seconds and has no way of knowing whether that answer matches the company’s current security posture, its actual product capabilities or its agreed contractual position.
That makes this a different problem from general enterprise search. The failure mode is not a slow answer; it is a confident wrong one that becomes binding. The release puts it precisely: getting it ninety per cent right is not good enough.
What the permission enforcement actually means
The substantive technical claim is that the MCP server applies user-level permissions, so what someone can see and do in ChatGPT, Copilot or Claude matches what they can see and do in Responsive.
That matters more than it might appear. The common way to give a language model access to enterprise content is to index that content into a vector store and retrieve against it, and in that architecture the permission model is whatever the index enforces. If the index does not mirror the source system’s permissions, a user can retrieve a document they should not see, and the model will summarise it without knowing it has crossed a boundary.
Enforcing permissions at the source, through a protocol the assistant calls rather than an index it reads, is the architecturally correct answer. It also keeps governance where the content lives, which is where an auditor would expect to find it rather than inside a copy made for retrieval, and it means a permission change in Responsive takes effect in the assistant without a reindexing job.
The release describes the mechanism at the level of principle and does not say how the system resolves a conflict between a user’s Responsive role and their assistant session, which is the detail a security architect would want first.
The shadow knowledge base
The CIO argument deserves isolating because it targets the buyer rather than the user. Responsive says the model extends existing AI investments without creating a shadow knowledge base.
That phrase names a real problem. When teams paste enterprise content into an assistant to get a better answer, the content leaves the governed system and the answer that comes back has no provenance. Multiply that across an organisation and the company ends up with knowledge scattered across chat histories, none of it approved and none of it auditable. A protocol that reads from the source rather than copying into a chat answers that directly.
The usage numbers and what they don’t say
Two figures anchor the adoption claim: more than 198,000 tool calls since the server launched in May 2026, and more than 38,000 instances of teams moving work from an assistant into an active pursuit.
Both are useful and neither carries a denominator. The release gives no user count, so a reader cannot tell whether 198,000 calls across six months represents deep usage by a few hundred people or shallow usage by many thousands. It gives no definition of what counts as a tool call, which could mean a search, a write-back, a permission check or all three counted together.
The 38,000 figure is the more interesting, because it describes an action rather than a retrieval. Retrieving content into a chat is a convenience. Writing an answer back into a pursuit, where it becomes part of a governed record, is the behaviour that makes the platform sticky. Roughly one action for every five calls suggests the product is being used for real work rather than experimentation, though the release does not make that argument itself.
Intelligent content management
The second half of the release concerns keeping the knowledge accurate, and the problem it describes is the one that erodes every content library.
Enterprise information changes constantly: a new SOC 2 report, an updated data retention policy, a product rename. Each change can invalidate answers approved against the previous version, and most teams discover the problem during a periodic manual audit, by which point stale content has reached proposals and AI-generated responses.
Responsive says its content management flags duplicate and conflicting information, assesses writing quality and metadata completeness, and recommends whether to accept, reject or review a piece of content. Teams can automate straightforward decisions with configurable rules while keeping human judgment for cases where accuracy and accountability matter.
The design choice worth noting is the recommendation rather than the automatic decision. A system that silently resolves a conflict between two documents is making an editorial judgement on the company’s behalf, and in a domain where the output becomes contractual, a recommendation with a human approver is the safer architecture.
The category position
The category matters for context. Strategic response management is a specific software market serving bid, proposal, security questionnaire and due diligence work, and it has historically been unglamorous infrastructure bought by a proposals team rather than a CIO.
The Model Context Protocol changes that position. If the knowledge that wins bids can be reached from inside whatever assistant a seller already uses, the content library becomes the asset and the interface becomes interchangeable. That is a stronger position for a content platform than for a workflow tool, and it explains why this release leads with the protocol rather than with a new interface, which is what the category would have announced two years ago before the protocol existed at all in any usable form, and before a single server could serve three assistants without three sets of credentials to maintain.
The customer and the market
BlackRock appears as a customer, with a quotation from Kristen Carloni, Head of Aladdin Business Proposals and Sales Optimization. That is a credible reference for this category, because Aladdin proposals are exactly the high-stakes, knowledge-intensive documents the product targets, and a named customer at that scale is more than most releases in this genre provide.
The market context the release supplies is a statistic from Responsive and APMP’s 2025 State of Strategic Response Management Report: 75 per cent of organisations say buyers demand faster response times, more personalisation and tighter budgets. The figure comes from a survey the vendor co-produced, which does not make it wrong and does mean a reader should treat it as an industry estimate rather than an independent finding.
What is missing
No pricing appears and no availability date beyond the Fall 2026 label. The release does not describe how conflict detection identifies a conflict, what threshold triggers a review, or whether the recommendations leave an audit trail. It gives no user count behind the 198,000 tool calls and no definition of a tool call. The release cites the Gartner prediction from a February 2025 press release without a link in the body. And the release does not separate which MCP capabilities are new in this release from those carried over from May.

What to watch
Whether Responsive publishes a user count and a definition of a tool call, because that would convert an adoption claim into an adoption measurement. Whether Responsive documents the permission enforcement in enough detail for a security team to assess, since it is the strongest technical claim in the release. And, whether the conflict detection produces an audit trail, because a recommendation without a record is a judgement nobody can review. And whether the MCP approach spreads across the category, because if it does, governed access to enterprise knowledge stops being a differentiator and becomes a baseline.

