Everyone’s Connecting AI. Nobody’s Fixing the Data.
Josh Derouin, VP of Marketing
What’s getting the headlines
Look at the AI news in wealth management over the last few months. Almost all of it is about connecting.
Anthropic launched a product for financial advisors. It plugs into custodians, CRMs, portfolio and planning platforms, and ships pre-built skills for meeting prep, follow-up and compliance review. It’s designed to sit on top of the systems you already run.
An AI workforce vendor announced what it called the largest AI workforce partnership in RIA history, on a pitch that firms don’t have to change anything about their data or infrastructure.
And wealthtech vendors are racing to launch their own MCP servers so AI can reach their platforms.
This is good news. The work is real. Nobody can call AI in wealth management unproven anymore.
But read those headlines again. Every one of them is built on top of a firm’s technology stack. Not one of them fixes what’s underneath it.
The question it’s driving
Every one of those headlines lands on a COO’s desk. And right behind it, we hear the same question.
“Why do we need a data layer? We’ll plug Claude into Salesforce, Orion and the custodian with MCP and let the AI sort it out.”
Fair question. Wrong conclusion.
Let’s be clear up front: MCP is great. It’s one of the best things to happen to enterprise AI in years. It gives any model a standard way to reach into your systems, pull data and trigger actions. We use it. You should too.
But MCP is a door, not a brain. It tells an agent how to ask for data. It says nothing about whether the data coming back is true.
Meet the Smiths
Your firm has one client household. Your systems have five.
Salesforce says Robert Smith. The custodian says Bob Smith. The planning software says Robert J. Smith. The portfolio system says Smith Household. The trust document says Bob & Susan Smith Family Trust.
Connect all five with MCP and your agent can now see all five Smiths. Congratulations. You’ve given AI faster access to the same mess your ops team has been untangling by hand for a decade.
Which Smith is real? Which system wins on the address? On beneficiaries? Who made that call, and where is it written down so the next agent doesn’t have to guess again tomorrow?
MCP doesn’t answer any of that. It was never built to do it. The protocol standardizes how systems get accessed and deliberately leaves the actual processing to whatever sits behind it.
So somebody has to do the work. Either you do it once, on purpose, and keep the result. Or your agents redo it on the fly every time someone asks a question, landing on a slightly different answer each time. In wealth management, “slightly different answer” is what an exam finding looks like before it becomes one.
MCP doesn’t solve the problem. It moves it.
Where the work actually happens
Here’s how we handle it at Subatomic. Raw data comes in and gets preserved exactly as it arrived. That’s Bronze. Then it gets cleaned, deduplicated, reconciled and matched to the right entities. That’s Silver. Then it becomes one business-ready record your whole firm agrees on. That’s Gold. One Smith Household. One set of facts. A full trail showing how every source got there.
Our AI Chief of Staff and AI Co-Workers work from Gold. Not from five competing versions of the truth. Not from guesswork.
And MCP fits right into that picture. It’s a great way to get data in and to expose what the platform knows to the models you already use.
Systems of record → MCP, APIs and connectors → Bronze (raw) → Silver (reconciled) → Gold (one truth) → Subatomic IQ → Chief of Staff + AI Co-Workers
MCP makes our architecture more useful, not less.
The plumbing era is ending
For twenty years, software vendors sold you a place to store things. Then integration vendors sold you pipes between those places. Then middleware vendors sold you something to manage the pipes. The average RIA is now running 10+ systems and paying a stack of vendors just to keep them talking to each other.
AI is about to make most of that pipe business worthless.
Connections are becoming free, open and standard. MCP is the proof. When every model can reach every system through an open protocol, “we integrate with X” stops being a feature. It’s the floor. Give it ten years and nobody pays a premium for plumbing.
Here’s what we’ll say that most vendors won’t: data cleansing is getting cheaper too. Can someone build MCP servers that dedupe records and resolve entities on demand? Sure. Some will. The moment they do, they’ve built a data layer. They just built it in pieces, with no persistent record, no lineage, and a maintenance bill they’ll own forever.
So “we clean your data” isn’t a moat. Anyone betting their company on it is in for a rough decade.
What doesn’t get commoditized
Your firm’s judgment.
Which system is the authority on which field. How you define a household. How your team handles a trust, a beneficiary change, a messy transfer. The way your best ops person reconciles an account in four minutes because she’s done it a thousand times. None of that lives in an API. No protocol ships it.
That’s what Subatomic IQ captures: your fragmented data, your institutional knowledge and your firm’s procedures, turned into one persistent intelligence layer that every AI Co-Worker operates from. It gets sharper every time the work gets done. It doesn’t reset when the chat window closes.
That’s the difference between giving AI access and giving AI understanding.
Our line in the sand
Connecting is solved. It’s becoming free and it gets better every quarter without you lifting a finger.
So stop spending a year proving it. If your AI plan is “plug it into everything and see what happens,” you’ll run pilot after pilot and get faster versions of the same mess. Every product you trial reads the same five Smiths and reaches its own conclusion about which one is real.
Use the connectors. Use the models. They’re great. Just don’t mistake access for a strategy.
Fix the foundation first. Make your data mean one thing. Then put AI to work on it.
Three questions before you buy anything
- Where does the agent get its data?
- When five systems disagree, who decides, and where is that decision recorded?
- What do you hand a regulator who asks how the AI reached its answer?
If the answer to all three is “we connect to your systems,” you’re buying plumbing.
Connecting APIs is easy. Making the data mean one thing is hard, and represents the meaningful outcome.