Every investment management vendor in the market right now is racing to ship a chatbot.

We’re not.

We built InvestNext MCP as a connector, not a chatbot. Instead of building our own AI experience inside the platform, we connected InvestNext to the AI environment our customers already use. The data stays inside InvestNext, the reasoning runs inside Claude, and the connector is what makes the conversation possible.

Here’s why we made a different call, and why we believe the architectural choice is going to look obvious in five years.

What the chatbot pattern actually is

When I say “the chatbot pattern,” I mean a purpose-built AI experience shipped as a feature inside a SaaS product. Most often, this is a  box in the corner of the screen with a guided prompt experience. It’s a fine-tuned model with retrieval pointed at the platform’s data, and a vendor that owns the conversation end to end.

This is not a lazy or poorly planned investment. It takes engineering. It takes fine-tuning. It takes UI design. The vendors who ship this are betting that owning the AI experience inside their product is the right long-term move. It’s a single pane of glass with absolute brand consistency. 

I take the case seriously, and I don’t think the people building this are wrong about everything. But I do think they’re wrong about one specific thing, and that one thing matters more than the rest.

The chatbot pattern has three structural cracks, and each one gets worse on a five-year horizon. Together, they’re the reason we built something different.

Crack #1: Always one step behind

The first structural problem with purpose-built AI inside a SaaS product is the release cycle.

Industry leading LLMs ship improvements weekly. 

New models. 

Better reasoning. 

Longer context. 

Better tool use. 

The pace of capability gain over the last 24 months has been the fastest in software history, and it shows no sign of slowing.

SaaS products ship on a different cadence: quarterly at best for major features, monthly for incremental ones. By the time a platform’s built-in AI catches up to what was state of the art when development started, the frontier has already moved. The client is using a capability the platform engineers were targeting six months ago, with no good way to upgrade except wait for the next release.

I don’t think this is a vendor competence problem. The people building purpose-built AI inside SaaS products are talented and well-resourced. The problem is structural. You cannot ship an institutional-grade AI experience on a quarterly release cadence. The math doesn’t work.

A connected product solves this by default. When Claude ships an improvement, every InvestNext MCP client gets it instantly. We don’t have to update anything. LLMs ship improvements and our client capabilities grow with it.

A connected product gets better every release. A purpose-built chatbot has to wait for the next update.

Crack #2: Prescription disguised as simplicity

The second problem is the interface.

Purpose-built AI tools, by design, prescribe how you should use AI. They limit what you can ask, they suggest prompts, and they route you through templates. They give you a “send” button and a friendly hint at the top of the box. 

The product makes a thousand small decisions on your behalf about what good AI use looks like.

For a buyer who has never used AI before, that scaffolding is genuinely useful. It gives the client a softer on-ramp. And I’m not against scaffolding when scaffolding helps.

But the sophisticated GP teams I talk to every week aren’t at the bottom of the ramp. They’ve already climbed it. They have working sessions in Claude that run for hours. They have favorite prompts, skills, custom instructions, multi-step workflows, and the whole machinery of someone who actually knows how to use this stuff. For them, a purpose-built tool with a guided experience is a downgrade. The scaffolding is a ceiling.

This is the audience question every vendor in our category needs to answer honestly. Are your clients AI novices who need to be taught, or AI fluent users who’ll resent being limited? For us, the answer is increasingly the second. The GPs we serve are the second group, and the gap is widening.

A purpose-built AI tool that tells you how to prompt is a ceiling, not a feature.

Crack #3: The Right Architecture to Build From

The third problem is the one I’ve been thinking about the longest, because it’s the one that matters most for the industry we serve.

LPs trust GPs with capital. GPs trust their platform of record with the data behind that capital. The chain is long and the stakes are real. When you introduce AI into that chain, the question every thoughtful client is going to ask is the same one I’ve been asking myself: what does this AI actually do with our LP data, and how does it grow into more without breaking the trust that got it through the door?

I believe the answer is easier to give honestly when the architecture is designed to scale in a direction the client can audit and maintain control over. 

We are heading toward an agentic future. I am as convinced of that as I am of anything we are working on. Agents will do meaningful work for GPs on the common tasks and the sensitive ones, with humans in the loop where judgment is required, and the firms that learn how to work that way will pull ahead. We are building toward that. The way you get there with an industry like ours is to start with an architecture where the client can see exactly what’s happening, then extend capability in steps the client’s compliance team can actually defend.

That is what the connector gives us. A foundational framework that can grow with the client – ready for agentic workflows that we will build with our clients.

What MCP actually does

So if not a chatbot, what?

We built InvestNext MCP on the Model Context Protocol. MCP is an open standard for connecting an AI environment to a data source. It is the connective tissue between the place reasoning happens and the place data lives.

When a client asks Claude a question about their portfolio, Claude reaches into InvestNext through the MCP connector. It pulls the relevant records, synthesizes an answer grounded in the actual data, and returns it inside the AI environment the client already trusts. The reasoning happens inside Claude. The data stays inside InvestNext.

That is the move. The data never leaves the platform it is supposed to live in. The reasoning runs in the environment the client already chose. The connector is what makes the conversation possible without collapsing the two into a single product.

The architecture is built for what it does today and designed for what we know comes next. Today, V1 gives a sponsor a real answer drawn from real data in seconds, inside the permission rules they already trust. This is the architecture that will enable use to build agentic capabilities with our clients. 

What this looks like in a sponsor’s actual workflow

The architecture is not theoretical.

A CEO has a 9 a.m. call with one of her top LPs. Twenty minutes out, she asks Claude for talking points. The answer comes back grounded in the LP’s actual history, recent activity, and capital position. Sourced from InvestNext and synthesized into a brief she can actually use. She walks into the call ready.

A finance lead has a question about an active raise on a Tuesday morning. He doesn’t open the platform. He asks Claude. Eight seconds later he has a fund-level answer with the figure he needs, drawn from current data.

An IR lead sees an investor across the floor at a conference. Phone out. Question typed. Briefing on screen. Walks over already informed.

Three different moments. Three different people. One architecture making every answer possible.

The longer arc

This is the bet I’m making with InvestNext.

Over a five-year horizon, and it likely won’t take that long, the connected product grows with every LLM enhancement. The client experience gets better automatically. The roadmap is freed from the impossible job of competing with Claude, ChatGPT, Gemini or any other LLM, which means our engineering capacity gets pointed at the work only we can do: deepening the InvestNext source of truth and extending the connector into the agentic capabilities the industry is moving toward.

That is the second half of this bet, and I think it is the part that gets undersold when people read the announcement. Agentic capability is where this industry is going. Agents will do real work for sponsors, on the common tasks and the sensitive ones, with humans in the loop where the stakes earn it. The firms that get there first, with their clients’ trust intact, will compound their advantage in ways the rest of the category will struggle to close.

A connected architecture is what makes that journey possible without breaking the trust that got us into the room. The connector earns the right to do more by doing the small thing reliably first, in front of the client, in an audit they can read. Then it extends, and extends again. Every step is one the client can grant, the auditor can defend, and the LP would recognize as the way a serious business introduces new capability.

The purpose-built product grows with the vendor’s own engineering capacity. Quarter by quarter, always reaching for a moving target. Always behind on capability. Always asking the client to use the platform’s idea of AI instead of the one they already chose, and reliant on the platform to build the agents they need instead of leveraging the platform’s agents in concert with their own. 

Closing

The right AI for your firm is the one you’ve already chosen. Our job was to make sure your InvestNext data could meet you there.

We’re running a small Early Access cohort to build V1 alongside the sponsors who’ll use it most. If you want a seat at that table, apply at investnext.com/mcp-early-access. I read every application.

Frequently Asked Questions