AI went from a distant curiosity to an everyday tool at our family-office customers within six months. Here's how SumIt is showing up for it.
Every SaaS product is shipping a chatbot these days. At this point, it’s expected — when you go to any SaaS or ecommerce website, a sidebar opens, you type a question, you get a written response or a link to a resource from the chatbot, and your data stays inside their walls.
At SumIt, we’ve taken the opposite shape. Today, we're announcing early access to the SumIt MCP server. It connects your books to the AI assistant your team already uses — Claude, ChatGPT, Microsoft Copilot, Gemini, or others — so you can ask questions in natural language about your ledgers from inside those tools. This development is read-only at first, OAuth-secured, and has deep links back into SumIt on every answer.
What our customers are already doing
Family offices have moved fast on AI. Most of the family offices and professionals we work with hold enterprise contracts with Anthropic, OpenAI, or Microsoft. The compliance work is done already, with strict permissions and SSO wired up to keep family office data fully secure and confidential.
The teams inside those organizations use these assistants every day to draft memos, summarize fund documents, parse PDFs, build dashboards, and remove the generally redundant, time-consuming tasks that take away from deep work. They've also started wiring in their CRM and project management tools, for example, so the AI has full context across all these channels.
However, one piece that has been missing from that graph is the general ledger. Getting GL data into the conversation has meant exports, CSV files, copy-paste, lots of Excel sheets, and the loss of provenance. SumIt’s MCP server was built to close that gap completely.
Why this works for family offices
There are two paths a SaaS company can take to bring AI to its customers: enclose them in your product, or connect them to the AI they already use. SumIt is taking the connection path. Why?
Frontier models improve faster than any software company can ship
The chat interfaces in Claude, ChatGPT, and Copilot get better every month. A chatbot we build today would compete with whatever Anthropic and OpenAI release next month, so our time is better spent on the parts of the stack only we can build — the ledger, the close, the audit trail, etc.
We saw this directly at Trove Family Office Week earlier this month, where we demoed SumIt connected to Claude through the MCP server. Roughly 80% of what made the demo feel magical was Claude. We’ve never been able to ask questions in natural language before and get complex breakdowns from our data. Dashboards can be spun up in seconds, and this used to require the work of an analyst to build and visualize.
The other 20% of the magic was us — the SumIt team — supplying clean, structured ledger data through a stable interface. That ratio tells you where to put your engineering hours.
Family-office tooling is fragmented by design
There's a maxim in this industry: Once you've seen one family office, you've seen one family office.
Every shop runs its own mix, including a CRM here, Excel there, a bespoke entity tracker somewhere else, and a stray visualization somewhere in the middle. No two stacks match, and no vendor will ever ship the right turnkey suite for this market — because there is no single shape to build for.
AI is the first piece of infrastructure that bends to that fragmentation. A frontier model reaching across every system the family office has chosen — CRM, document store, project tracker, general ledger — is the right shape for the questions that matter at the principal level: "given our distributions, our commitments, our cash position across these eight entities, and what's in the deal pipeline, what should we do?" No other technology has been able to answer that without a human behind the wheel.
Our data model speaks family office
Partnership allocations, intercompany eliminations, multi-entity consolidations, capital calls, family-office cash sweeps — all the things that are tried and true for family offices. A horizontal accounting connector built on QuickBooks or Xero simply cannot expose these without rebuilding our data model. The MCP server makes them first-class primitives that any frontier model can reason about.
The most prominent lesson we’ve taken away is that primitives only matter if the answers built on them can be trusted.
How we make AI answers trustworthy
The general ledger is the right place to be exact. Numbers in the GL flow into financial statements, tax returns, presentations to family members, and investor letters. Therefore, they have to be right, and they have to be traceable. "The AI said so" is not an answer, though it’s something we hear more often than we’d like to nowadays.
SumIt’s MCP lives in Claude, while the verification lives with us. Here are a few mechanics that make that real:
Self-attesting responses — Every tool call returns metadata alongside the data, such as totals, counts, period boundaries, and debits-equal-credits checks. The math is visible, so you know exactly what is happening behind the scenes to get these numbers.
Deep links back to SumIt — Every response includes a link to the canonical view in our product. Verification happens in SumIt, where the audit trail lives, and hallucinations and composition errors surface on click.
Read-only at first — No tool in this release can create, edit, or delete anything.
Customer permissions, end-to-end — Every query runs as the authorizing user. Entities that a user can't see in SumIt are invisible to the AI.
Full audit trail — Every tool call is logged, meaning who, when, what was asked, and what came back.
Revocable in one click — You have absolute control to disconnect from the AI, and tokens will be invalidated immediately.
What this looks like in practice
Our customers have been pairing SumIt with Claude and ChatGPT for months, through a hybrid of file uploads, exports, and custom scripts our team has built alongside them. The accounting workflows that have emerged cluster into three shapes.
1. Creating journal entries
Some common examples we’ve seen include:
A PDF bank or credit-card statement becomes a clean CSV ready for SumIt upload
A pile of receipts becomes coded journal entries with employee, GL code, and tags
An HR payroll export becomes a fully-formed payroll JE
Raw credit-card transactions become auto-coded expense entries with confidence levels per line
2. Reading and transforming ledger data
Drop a SumIt reconciliation and an AMEX statement into one conversation and let Claude find the gap. You can run tasks to reformat a standard report into a principal-facing presentation, run inter-entity balance checks across the family structure, or answer something as simple as "how much did we spend on legal expenses last year?" in seconds. For a more comprehensive reporting package, including a balance sheet, income statement, budget vs. actual, full register with sub-detail, two minutes is all you’ll need to have it in your hands and ready to go.
3. Visualization
The visualizations might be our favorite component of the AI model's capabilities. Get customized graphics displaying P&L packages with top vendors, recurring vs. non-recurring spend, and category breakdowns drawn straight from the ledger. If you’re working with family members who don’t read a P&L, you can create principal-facing chart packs. It’s the same data, just translated visually.
All of this runs today through scaffolding like uploads, exports, custom Python scripts, and others. The MCP server replaces the scaffolding with a clean, governed connection.
The cross-system questions come once SumIt joins your other connectors. It can match this week’s bank feed against journal entries posted on Tuesday, for example. The MCP server is one node in a graph our customers are already building, but our job is to make it the cleanest, best-shaped node for family-office accounting.
Read first, but write when it's right
Early access is read-only. When you connect SumIt to Claude or ChatGPT, you can ask anything across your full chart of accounts and entity structure, and every answer will ship with a deep link back into SumIt for the audit trail.
Writes come later — posting journal entries, categorizing transactions, running multi-step close workflows, etc. They'll be gated by pilot evidence and the same review-and-approval workflow that a controller already runs on any human-authored entry. We’re keeping a careful eye on when our customers repeatedly ask Claude to post entries through us, as that’s when writes become real.
Anthropic recently released finance agent templates, including a general ledger reconciler and a month-end closer. They signal where the industry is heading, and our place in that stack is upstream of them: the clean data the agents reason over.
The bet
The chat layer belongs to frontier models. Our place is the clean, well-shaped layer of family-office accounting beneath it, and one that you can trust with full security and transparency.
Our chips are on depth, including the domain primitives that we can carefully build, and the trust mechanics that make them safe to expose. If you run a family office and want early access, request it here, and we'll get you connected.
— Fernando
Try SumIt Today
