Most teams today are connecting Claude, or another AI tool, directly to a handful of systems: a CRM, a document store, a market data provider, one connection at a time. This is one of the practical problems we set out in our whitepaper, How to scale AI in private markets, which covers the architecture question in full. This piece goes deeper on one part of it: what it actually takes to operationalise Claude, rather than just connect it.
That approach runs into three problems fast. First, every connection has to be rebuilt for every tool. Set it up for your Claude instance, and the moment someone wants to use ChatGPT instead, the work starts again from zero. Second, the insight generated in any one conversation stays trapped on that person's machine, in that one chat. It never makes it back to the rest of the team. Third, it gets expensive quickly, because each of those direct connections is querying the firm's systems separately, and that adds up in token cost with nothing to show for it beyond a single conversation.
A single layer instead of many connections
The fix is not a bigger AI model. It is a layer that sits between any AI tool and the firm's intelligence warehouse: one connection that reaches everything, instead of many separate ones that each reach a part of it.
Once that layer is in place, every tool call, whether from Claude, ChatGPT, or whatever comes next, goes through the same route into the firm's data. That cuts token cost significantly, because the same information is not being pulled and reprocessed by every separate connection. It also means the firm is not locked into today's model. When a better one comes along, there is nothing to rebuild.
What this looks like day to day
For a market-facing professional, it starts before any meeting. Instead of piecing together old notes, past conversations, and where things were left, a single prompt pulls together everything already sitting across the CRM, the document repository, and market data: the last conversation, any outstanding actions, the latest news and financials. No dashboards, no switching between systems.
The same logic scales up to something like conference preparation. Rather than someone manually cross-referencing a five hundred name attendee list against a target universe, a single prompt does that matching automatically: flagging who has already been engaged, who is new, and who is worth the most time, ranked by fit. Work that used to take hours becomes one request.
Behind both of these, the mechanism is the same. Anything run on a schedule, a daily market scan or a weekly CRM update, runs centrally rather than on somebody's laptop, so it is reliable and visible: the team can see exactly what is running and what it costs. And any task that does not need a premium model, simple screening or formatting, can be handled by a cheaper one, with only the harder reasoning passed to a frontier model. That means the firm is not paying premium prices for routine work.
Why this matters
None of this requires a bigger or better model. It requires the infrastructure that lets whichever model a firm uses reach everything it needs, in one place, without rebuilding that access every time the tooling changes. This is the architecture question we set out in full in our whitepaper, How to scale AI in private markets.
Firms operationalizing Claude or other LLMs this way are seeing meaningful reductions in manual research time and in the token cost of running AI at scale, because the same information is no longer being fetched and reprocessed separately by every tool and every person who uses it.
If you want to see what operationalizing an LLM looks like for your own tech stack, book a demo or get in touch with the team.
