The connector problem: governed data access is the bottleneck for finance AI, not the model
Published 15 May 2026
The conversation about AI in finance keeps centring on the model. Which one is best at reconciliation, which one is best at variance commentary, which one is best at audit. That conversation is the wrong one for the median finance function. The thing that determines whether AI in finance works in your business is the data the model is allowed to see, in the form it sees it, with the controls that govern it.
That is the connector problem, and it has just become the thing every finance leader needs to think about more carefully.
What changed in the last six weeks
Moody’s released an MCP app on 9 April exposing credit ratings, risk intelligence, and ownership data covering more than 600 million entities with around 2 billion ownership links across credit, compliance, and operational domains. (Source: moodys.com press release, 9 April 2026.) That was the opening move.
On 5 May, Anthropic’s finance launch shipped a set of new connectors alongside the agent templates. Dun & Bradstreet, Fiscal AI, Financial Modeling Prep, Guidepoint, IBISWorld, SS&C Intralinks, Third Bridge, and Verisk landed as new integrations. The earlier integrations with FactSet, S&P Capital IQ, MSCI, PitchBook, Morningstar, Chronograph, LSEG, and Daloopa were already in place. (Source: anthropic.com/news/finance-agents.)
That is a substantial connector library for one vendor’s agents. It is also a window into how the finance AI conversation is shifting. The vendors that win in the next two years will not be the ones with the cleverest model. They will be the ones with the broadest set of governed routes into the data the model needs.
The market read that signal in real time. On the day of the launch, FactSet shares fell as much as 8.1%, with Morningstar, S&P Global, and Moody’s all selling off in sympathy. (Source: Sherwood News, 5 May 2026.) The data vendors are not under existential pressure. The market is repricing whether the existing distribution model holds when an agent layer commoditises the user interface in front of the data. That is exactly the connector argument.
The implication for a finance function is that the data layer, not the model, is the place to start.
What “governed access” actually means
A connector is not just a pipe. It is a pipe with three things attached to it.
The first is the authentication and authorisation layer. The model can only see data the connector lets it see, scoped to the user or the role or the agent that is making the request. That matters in finance because the data that goes into an agent is often the data that, if mishandled, becomes a regulatory or commercial problem. The connector design is what stops the agent from seeing more than it should.
The second is the audit trail. The connector logs what the agent asked for, what data it returned, and what context the agent used downstream. That is the foundation of the audit defensibility argument. Without it, you have a model that produced a result and no way to reconstruct why.
The third is the data contract. The connector returns data in a known shape, at a known cadence, with known quality assumptions. The agent does not have to figure out whether a P&L is in IFRS or US GAAP. The connector has resolved that upstream. That is the difference between a model that produces a plausible answer and a model that produces a defensible one.
A vendor connector gets you all three for the data the vendor owns. Your internal data is still yours to govern.
Where the new MCP layer matters
Anthropic’s launch leans heavily on the Model Context Protocol. The Moody’s MCP app is the headline example. The pattern is general. A data owner publishes an MCP endpoint, the agent connects to it, and the data flows under the protocol’s contract. Kate Jensen, Anthropic’s Head of Americas, framed the Moody’s collaboration as designed for work where “the stakes are high and outputs need to be defensible.” That is exactly the framing this post argues for, in the vendor’s own words.
What that changes for a finance function is the integration cost. Before MCP, every connector was a bespoke build, and the cost of getting a model to talk to a new data source was a project. After MCP, the cost is closer to “install an app.” The library of data sources a finance team can reach is no longer bounded by what one vendor has bothered to build a connector for.
That is the version of the story that gets the most coverage. The version that matters more for a mid-market finance function is the internal one. Your ledger, your sales system, your warehouse, your CRM. None of those come with a published MCP endpoint by default. The work of putting one in front of them is the work that determines whether your finance function can actually use the new generation of agents on its own data.
The data quality post covers the data hygiene piece. The connector question is the operational question that sits on top of it. You can have clean data and no governed route to it. The agents will not work.
Where this leaves the mid-market
The honest position is that the businesses that benefit most from the new connector library are the ones whose source-of-truth already lives in the vendors the connectors point at.
If your business runs on FactSet, S&P, Bloomberg, and a major ledger vendor, the agents have most of what they need. The deployment question is governance, not data.
If your business runs on an internal ledger, a bespoke sales system, a warehouse that has not been touched in three years, and a CRM that nobody trusts, the agents will not work on your data until you do the connector work yourself. Or until your software vendors do it for you. The market will produce more of the latter through 2026 and 2027, but the timing is not in your control.
That is not a reason to wait. The connector work is the work that pays off regardless of which AI vendor you eventually back. It is the same data layer work that the building an AI-ready finance function post covers. The agents add a deadline to it.
What I would do with this
I would treat connector strategy as a first-tier topic at the next finance leadership meeting. Three questions to answer.
What are the systems the finance function currently relies on, and which of them have a governed route into a model? Most finance leaders cannot answer this off the top of their head. The exercise of mapping it is itself useful.
For the systems that do not have a governed route, which ones are highest-value if they did? The ledger is usually the answer for close-related agents. The CRM and the sales system are usually the answer for forecasting. The warehouse is usually the answer for anomaly detection.
For the highest-value gaps, what is the path to a governed route? An MCP endpoint, a vendor-supplied connector, a build, a managed integration. The decision is partly cost and partly control. The constraint is usually the IT relationship more than the budget.
The output of that conversation is a connector roadmap that goes alongside the model strategy. Most finance functions today do not have one. That is the gap.
Where this lands
The model is no longer the bottleneck. The model is approaching the level where, on most finance tasks, the gap between the best model and a competent senior analyst is closing. The bottleneck is the connector and the data behind it, and the governance attached to both.
The finance functions that win the next two years are the ones that treat the connector layer with the same seriousness they treat the ledger. The ones that wait for the vendor to deliver a connector for their bespoke system will spend 2027 building what they could have built in 2026.
The agents are coming. The data layer was always going to come first.
Maebh Collins is a Fellow Chartered Accountant (FCA, ICAEW) with Big 4 training and twenty years of operational experience as a founder and senior finance leader.