A few months after connecting SumIt to their AI assistants, our customers have settled into their new routine. Here's what they use the connection for.
Picture this — a controller at a single-family office with about 20 entities closes the books, opens Claude, and runs one saved prompt. Within an hour, her principal has the month's report, consolidated across all entities and compared against last year.
This isn’t a made-up scenario; however, it’s reality for one of SumIt’s clients. "All of this would have taken me days before," she told us.
How is this possible? Simple — she's using the SumIt MCP server we launched in May. MCP (Model Context Protocol) is an open standard that lets AI assistants query other software directly. SumIt’s MCP server gives them read-only access to the general ledger. In our case, it connects the general ledger to the assistants our customers already use, like Claude, ChatGPT, and Microsoft Copilot, and applies the same permissions each person has in SumIt.
Since then, we've spent a lot of time with the people using the SumIt MCP server every day to see what it looks like in real-world practice. Our team joins calls to walk through what our clients have built, and we’ve noticed that nearly all of it is the usual recurring work a family office handles each month, like the reporting package for the principal.
AI Reporting for Family Offices: Monthly Packages and Principal Summaries
Reporting is the first job our customers hand to their preferred AI assistant. The people at the end of the reporting chain want a short, clear picture, and as a CPA who runs a multi-family office put it, they "just don't know it until they see it."
Build it once, run it every month. The controller we mentioned earlier typically closes her books in about two days. Then, her saved prompt has Claude pull an income statement and a balance sheet for every entity from SumIt, along with consolidated financial statements for the full structure. The income statements show the month, year-to-date, and prior year-to-date, and the balance sheets compare month-end with the prior year-end. She checks it in Excel, then turns it into a snapshot for the principal, with income, outflows, and recurring versus one-time spending broken out by entity and category.
Let the AI build the deliverable. That CPA built a family's monthly package during a two-hour working session, with the family in the room, and made changes live as they reacted. Now, they have an interactive file with a tab for each view and a dropdown that switches between each entity and the consolidated picture, styled in the family's own branding. Each month after that, a new run takes only a few minutes. "Every tab had information that the family wanted to see," he told our team. "Like there was no extraneous information."
Family Office Cash Flow Forecasting with AI
A different CFO we work with had already sketched the cash forecast he wanted in Excel. He is particular about wanting his own layout and entity groupings, so he wants to maintain this structure even when using AI to help his processes. He pointed Claude at every account tagged as cash in SumIt and handed it the spreadsheet as the spec.
In the forecast, credit card payments are spread across the expenses they cover, based on the prior month's transactions, so the forecast shows meals and travel rather than a single card payment. Intercompany transfers net out, so nothing is counted twice. Both depend on seeing individual entries, entity by entity.
The first version took 2–3 days, with almost all of that time spent shaping and tweaking the output. "All the data was there and available," he told us. Now he refreshes it each month with a short prompt.
AI-Assisted Bank and Intercompany Reconciliation
Bank reconciliations still happen at close. An AI assistant will still follow its typical workflow of adding checks around them, and our customers follow the same steps — they tie out an account's opening and closing balances, pull the period's activity, check which way each entry posted, look for duplicates against the custodian or bank statement, and, when something is missing, ask whether the payment was ever booked in any entity.
The controller applies that check to intercompany balances. Her report pulls every due-to and due-from balance across her entities and shows, entity by entity, where they're out of balance. In one run, Claude flagged an entity as missing transactions. The CFO then ties out his cash and credit card accounts against the statements.
In both cases, the AI assistant finds the problem and a person fixes it. Claude tells the CFO the entity, the account, and the amount, and he then opens SumIt, goes to the register, filters, and recodes the entry himself. "The piece that took the longest," he told us, "was having to find and identify the specific transaction." Every answer already carries a link back to the matching page in SumIt, and making that link impossible to miss is on us.
Where We're Taking It
A different multi-family office tried pulling years of trial balances through the connector for a client reporting tool, and went back to downloading them straight from SumIt. For that much data, an export is simpler.
Where the connector does beat the old workflow, customers want more from it, specifically help preparing:
Recurring accruals, including interest
Depreciation and amortization schedules
Journal entries, including the intercompany entries that keep entities in balance
The CFO already has Claude draft his interest accruals, then exports the entries and uploads them to SumIt by hand.
When Claude flagged the entity with missing transactions, the controller pushed back. She was wrong, and Claude backed down anyway. "I'd rather approve it before it happens than review it and catch it after it happens," she said. The CFO agreed: "I don't know if I'd be like, ‘Alright, Claude, go ahead and fix that journal entry."
The CPA weighs changes differently, though — he would let AI fix the tags and account codes on its own, as long as there's an audit trail, but would hold journal entries for a person to approve. He already works that way with SumIt's bank feeds, where every suggestion waits for review: "…you aren't afraid to make a mistake because you can, you know, undo it and redo it."
This is what we're building toward. The assistant proposes, a person reviews it in SumIt, and no entry posts to the ledger until it's approved.
Our Bet
In May, I wrote that the chat layer belongs to frontier models, and that our place is the clean, well-shaped foundation of family-office accounting beneath it. Every workflow here started with one job and a result someone could check against SumIt — the controller started with her reporting package and has since added intercompany reconciliation. That's the recurring work of the month, and it's where AI on the books is already paying off.
— Fernando
Fernando Ramirez is the co-founder and CTO of SumIt Software, where he has led engineering for the platform's general ledger, multi-entity accounting, and integration infrastructure since 2022. He brings more than a decade of software engineering experience, with the beginning of his career focused on flight dynamics at NASA's Goddard Space Flight Center. Before co-founding SumIt, Fernando worked in staff engineering and engineering management roles at companies like Rhino, Blue Apron, and Capsule.
FAQs
Is SumIt’s MCP server read-only?
Yes. SumIt’s MCP server gives AI assistants read-only access to your family office’s general ledger. The AI can look up balances, transactions, and entities, but it cannot create, edit, post, or delete entries. Every tool call is logged with the tool used, who asked, when, and what data was returned, so your team has a complete audit trail of AI activity.
Which AI assistants work with SumIt?
SumIt’s MCP server works with Claude, ChatGPT, and any other AI assistant that supports the Model Context Protocol (MCP). Claude users can connect through Custom Connectors on Pro and Team plans, and ChatGPT users can connect on Business, Enterprise, and Edu plans. Other MCP clients connect using OAuth 2.1 with PKCE, and setup takes about three minutes.
Can the AI see data a user can’t?
No. Every query an AI assistant runs through SumIt’s MCP server uses the permissions of the person who authorized the connection. If that user can’t see an entity in SumIt, the AI can’t see it either. Access is also limited to the organizations you authorize, and you can disconnect the AI in one click, which immediately invalidates all access tokens.
Try SumIt Today
