Connectors or Foundations: The Two Kinds of AI for Finance

A question comes up on nearly every sales call: how does this compare to the other “AI for finance” tools out there? The honest answer is that there are two fundamentally different kinds of software sold under that same label, and confusing them costs buyers real money. Call them connectors and foundations.

What is a connector?

A connector is a frontier model sitting in the middle of your stack, reaching out through MCP to your CRM, your data warehouse, your email, your document store, and synthesising an answer from whatever it pulls back. It is a chat box wired onto your existing systems rather than a new architecture underneath them.

This is what most “AI for finance” platforms pitched today actually are. Take a general-purpose model, put a finance-shaped label on it, and connect it out to the tools a firm already runs. It is the cheap, fast, vendor-friendly way to claim an AI strategy, and it demos beautifully, because demos use clean data and simple questions. The trouble starts once real deal data and real questions arrive.

Why do connectors struggle in practice?

Connectors struggle because the model never has a single coherent view of the firm. It only ever sees fragments, reconstructed fresh, from systems that do not agree with each other.

Open an MCP connection to a CRM and the model gets a slice of fields, and has to guess what they mean. Open a connection to the data warehouse next and it meets a different schema, with a different idea of what “revenue,” “customer,” and “deal” mean. Add documents, emails, and call notes and the model is now reconstructing context from scratch across five separately modelled worlds that do not agree, with no coherent map anywhere in the loop. The integrations are brittle on top of that: they break whenever an upstream system renames a field, changes a schema, or shifts a permission.

Being a chat box on top of your data is a taken seat, because the moment frontier assistants ship the same connections natively, “we have a chat interface on your systems” stops being a product.

Is connecting to your CRM and data warehouse a real moat?

No. MCP is a public, open protocol, which means the ability to connect to a CRM, a warehouse, or an inbox is not something one vendor can own. Anyone can build the same connection, including the model providers themselves as they ship native connectors of their own. “We integrate with your CRM and your data warehouse” describes plumbing that commoditises in real time, not a defensible position.

Why does connector cost and quality break down on real data?

Real financial documents are large enough that pulling them into a model’s context window directly starts to erode the quality of the answer, not just the cost of getting it. A single Excel model can run to roughly a million tokens, which puts it near the ceiling of even the largest context windows available today. Faced with that, a connector either processes one file at a time and loses cross-file reasoning, or floods its context and starts dropping pieces. The tidy demo answer gets noisy the moment it meets a real data room.

Cost moves the same direction. Feeding an entire data room into context “just in case” turns what should be a manageable token budget into a large, recurring overrun. Managing that well requires real routing, a platform choosing the right model for each step in the background, plus disciplined control over what actually enters context. Bolting a single frontier model onto MCP connections has neither.

Won’t better models eventually fix this?

It is a fair challenge, and worth taking seriously, but three things stand in the way. Model capability has visibly plateaued, and the labs’ own spending, buying services firms and forming consulting joint ventures, is a signal that they see the next bottleneck as something other than raw model quality. Even a genuine step-change in capability would not close the gap, because reading coherently across five disjoint systems and reconstructing a single trustworthy answer is orders of magnitude harder than answering a question about one document, not a modest step up. And more capable models generally cost more per token with larger context windows, so the economics of pulling raw data into context get worse as models improve, not better.

What is a foundation, and why is it harder to build?

A foundation consolidates and structures a firm’s information first, before a model ever answers a question. CRM, documents, financial data, call notes, deal pipeline, and portfolio history all sit on one connected, coherent layer. The model stops being a frontier API with plugins bolted on and becomes part of an architecture that already understands what the firm does, because it is reading from the structured data layer rather than reassembling five schemas on the fly.

This is genuinely harder and slower to build. The migration of a firm’s existing information onto a coherent structure is most of the work, and there is no shortcut that skips it. What it buys in return is real: bounded, predictable cost instead of a token budget that scales with data room size, and a model that can be swapped for a better one the moment it lands, because the value sits in the structured layer underneath rather than in whichever model happens to be answering today.

How do you tell a connector from a foundation?

Two questions do most of the work. Does the model have a single coherent view of your firm’s information, or is it stitching together five separate views every time you ask it something? And if you want to swap the underlying model in six months, can the platform do that without rebuilding everything?

Both should be a clear yes. If either answer is “sort of,” what is in front of you is a connector, not a foundation, whatever the label on the pitch deck says.

A connector might still be the right tool for the next twelve months, particularly if a firm genuinely does not yet know where it wants to land. The mistake is not using one. The mistake is confusing it with the destination. If you are trying to work out which one you are actually being sold, talk to us and we will walk through it plainly.