The Enterprise AI Gateway: What Salesforce, Intuit and Experian Are Actually Building

Every enterprise AI conversation of the last two years has been about bringing external intelligence in — a copilot, a chatbot, an internal agent wired to a knowledge base. A quieter and more consequential shift is running the other way: enterprises are now pushing their own capability out, into ChatGPT, Claude, Gemini and Copilot, as first-class apps and connectors rather than as content those assistants merely read about. The businesses moving first are not startups experimenting with a plugin. They are Salesforce, Intuit and Experian — a CRM platform, a financial software company, and a credit bureau — putting live systems and, in one case, regulated personal data behind a conversational front door they do not fully control.

The shape underneath these launches is consistent enough to draw as an architecture. At the top sit the assistants themselves — ChatGPT, Claude, Gemini, Copilot, or an enterprise's own internal AI — as interchangeable entry points a user might reach for on a given day. Beneath that sits what is effectively an enterprise capability gateway: an interface layer that decides what a given assistant is allowed to see and do, independent of which assistant is asking. Underneath the gateway sits the part that actually earns its keep — identity, consent, policy, permissions, audit — the machinery that turns "an AI wants your CRM record" into a governed, logged, revocable event rather than an open door. Below that sit the data products and business rules that translate raw system access into something an assistant can safely reason over, and the approved actions it is allowed to take on a user's behalf. At the bottom sit the systems of record this was all built to protect — CRM, ERP, core banking and credit platforms, marketplaces, workflow engines — the things no enterprise is willing to expose directly.

That is the theory. What is actually shipping looks less like one clean gateway and more like three different enterprises solving the same architecture problem at very different points on a risk curve.

Salesforce is building the gateway as infrastructure, not a feature. The Salesforce/Claudeforce partnership, expanded on August 26, 2026, makes Claude the default reasoning model across Agentforce, Slack AI and Slackbot, and ships a plugin called Salesforce in Claude with 37 prebuilt sales skills running on live CRM data. The commercial commitment behind it is real: Salesforce has invested five billion dollars in Anthropic and has budgeted three hundred million dollars in Anthropic token spend for 2026 alone. What makes this the clearest instance of the gateway pattern is the plumbing underneath the announcement — Headless 360, the initiative that exposed Salesforce's platform as APIs, MCP tools and CLI commands, and the Hosted MCP Servers that reached general availability in April 2026. The actual product insight, according to people close to the rollout, was administrative rather than intelligent: one admin connects an org once, and every user gets access scoped to their own existing permissions, instead of every team rebuilding authentication and governance from scratch against the same raw MCP plumbing. That is the identity-consent-policy layer of the gateway, built once and reused, rather than improvised per integration.

Intuit is running the same idea at a different altitude — read access first, actions second, and explicitly under a trust promise. TurboTax, Credit Karma, QuickBooks and Mailchimp are live as apps inside ChatGPT, and QuickBooks has a parallel connector in Claude, letting a small business ask about profitability, cash flow and benchmarks using its own real data. What is more interesting than the launch is Intuit/QuickBooks's stated trajectory: from answering questions about a business's books toward acting on them — running payroll, managing compensation changes, surfacing the business's own lending offers rather than generic peer benchmarks, and deeper creditworthiness intelligence. Intuit has been explicit that connected data stays inside Intuit's own systems, is not used to train the underlying foundation models, and that every action remains reviewable and human-verified. That is a company drawing its own line on the gateway's approved-actions layer before a regulator or a customer forces the question — read-only until governance catches up, then write access only with a human still holding the pen.

Experian is the sharpest test case, because it is the one that has moved past product comparison into someone's actual financial record. Its family of Experian apps inside ChatGPT started with insurance and loan marketplace comparisons, then credit cards — useful, but still closer to a shopping assistant than a data pipe. The step that changes the risk category is the newest one: UK consumers can now view their actual Experian Credit Score inside ChatGPT, see how it is calculated, review their score history, and check their score band, alongside current-account comparison. That is not content about a person's finances. That is a live, regulated personal record being surfaced through a third-party conversational interface Experian does not operate. It is precisely the scenario the TRUST framework and CLEAR framework were built to stress-test, applied to a channel neither framework originally had in mind.

Read against the gateway diagram, the three case studies are not three versions of the same maturity level. They sit on a spectrum. Intuit's QuickBooks app is closest to the diagram working as drawn — read access, explicit data-use commitments, human-verified actions, a company narrating its own governance posture in public before expanding scope. Salesforce has built the deepest version of the identity-and-permissions layer, but its scope so far is B2B sales workflow, not a consumer's personal financial data — an easier governance problem than Experian's, even if it is a harder engineering one. Experian is the one actually testing the CLEAR framework's lineage and accountability checks in the wild: when a consumer's credit score, history and band flow through ChatGPT, the questions that matter are not whether the feature is useful — it clearly is — but who is accountable if the assistant misstates a number, whether the conversation itself becomes a new, ungoverned data trail, and whether "OpenAI's platform" or "Experian's data" is the party a regulator calls first when something goes wrong.

That is the honest answer to whether this makes sense: yes, structurally, for the same reason retail's rush into agentic commerce made sense before anyone worked out who owns the checkout — an enterprise gets distribution inside a channel with hundreds of millions of users without handing over its systems of record, and the assistant platform gets stickier, higher-value usage instead of being a stateless chat window that answers a question and forgets it happened. The mutual benefit is real. What is not yet proven is whether the middle of the gateway diagram — identity, consent, policy, permissions, audit — is being built with the same seriousness as the top and bottom layers, or whether it is still doing more work in the press release than in the actual permission model. Salesforce's admin-scoped access model suggests the former is achievable. Experian's leap straight to personal credit data inside a general-purpose assistant is the case every CIO evaluating this pattern should study hardest, because it is the one furthest ahead of its own governance layer catching up.

For any enterprise thinking about where its own capability sits on this spectrum, the useful first exercise is not "which assistant should we integrate with." It is mapping your own data and actions against the same four tiers Intuit, Salesforce and Experian have inadvertently laid out this year: content an assistant can read about you, analysis an assistant can run on your data with your explicit connection, personal or regulated records an assistant can surface on a user's behalf, and actions an assistant can take that move money, data or commitments without a human in the loop. Each tier needs a materially different answer to consent, audit and reversibility — and building the gateway diagram correctly means refusing to let a single integration slide from tier one to tier three because the demo looked good in a board meeting.

If you're evaluating where your organisation's capability should sit on this spectrum — and what the identity, consent and audit layer actually needs to look like before you expose anything through a third-party assistant — request a strategy conversation.