Sumary's $25M Raise vs. Building Your Own AI Accounting Agent
What Sumary’s Raise Actually Signals
As of mid-2025, Danish startup Sumary has been reported to have raised a $25M seed round to expand its AI-powered accounting platform from Scandinavia into the UK and Benelux markets. (Specific coverage dates and links referenced in earlier versions of this article cited future publication dates in error; readers should check Sumary’s own press page and outlets like Tech.eu or ArcticStartup for the most current, verified reporting.) Sumary positions itself as an AI-native ledger — not an AI feature bolted onto an existing general ledger, but a system designed from the ground up around automated transaction processing and categorization.
This fits a broader pattern. Venture capital has been flooding into AI-first accounting tools globally, and new AI agents are launching across accounting, construction, healthcare, and other sectors. These platforms will reach North American and broader global markets within a few years.
But here’s the question that matters for your firm right now: should you wait for one of these platforms, build your own AI agent over the tools you already run, or do both?
What AI-Native Platforms Like Sumary Do Well
AI-native accounting platforms promise a fundamentally different data model. Instead of you categorizing transactions and the software recording them, the software categorizes transactions itself, learns your patterns, and surfaces only the exceptions for human review.
The appeal is real:
- Lower switching cost for new firms. A firm starting from scratch (or onboarding a new client with no history) gets AI categorization from day one without training a custom model.
- Vendor-managed AI. The platform handles model updates, prompt engineering, and accuracy improvements. You don’t staff it.
- Integrated workflow. Categorization, reconciliation, and reporting live in one system — no glue code.
But there are clear limits, especially for established firms:
- Migration cost. Moving clients off QuickBooks Online, Xero, or another GL is a massive lift. Most firms won’t do it for incremental AI gains.
- Client-data jurisdiction. For a European platform expanding internationally, where data is processed and stored matters — particularly under IRS Publication 4557 guidance on safeguarding taxpayer data and under various state board of accountancy rules. Verify with the vendor and your compliance advisor.
- Lock-in. An AI-native ledger that owns your data is the deepest possible vendor dependency.
What Building Your Own AI Agent Gets You
The alternative to adopting a new platform is connecting an AI assistant — Claude, GPT-4, or another model — to the systems you already run. The cleanest way to do this today is via MCP (Model Context Protocol), an open standard that gives an AI secure, governed, tool-by-tool access to your data.
A custom AI agent built over your existing GL and practice-management system can:
- Categorize transactions by reading your chart of accounts and historical patterns directly from QuickBooks Online or Xero, without moving data to a third-party platform.
- Prep bank reconciliations by pulling bank-feed data and matching it against GL entries, flagging discrepancies for human review.
- Draft workpapers and variance analyses from a trial balance, using skills (reusable instruction sets) that enforce your firm’s standards.
The key advantage: you keep your existing GL, your existing client relationships with that GL, and your existing integrations. The AI layer sits on top.
The honest downside: you’re responsible for building, maintaining, and improving the agent. That means either in-house technical talent or an outside partner. And as underlying AI models update — which they do frequently — you’ll need ongoing prompt tuning and regression testing to keep accuracy stable. It’s not zero-effort.
Comparing the Two Paths Side by Side
- Vendor manages the AI model and updates
- Requires GL migration for full benefit
- Integrated workflow out of the box
- Data lives on vendor’s infrastructure
- Limited customization to your firm’s specific processes
- Availability depends on vendor’s market expansion timeline
- You (or a partner) manage the AI layer
- Works with your current GL — no migration
- Requires integration work (MCP server, skills)
- Data stays in your existing systems
- Fully customizable to firm-specific workflows
- Available now, but requires ongoing prompt tuning and testing as AI models update
How to Think About the Economics
Don’t trust anyone who gives you a single dollar figure for ROI here — the math depends entirely on your firm’s size, client mix, and current tooling. But you can model it yourself:
The formula: (hours saved per client × clients × blended rate) = recovered capacity per month. That recovered capacity either reduces overtime or — more valuably — gets reallocated to advisory work billed at a higher rate. The 40–60% range above is a rough starting assumption based on the general pattern that AI categorization handles routine entries well but still struggles with unusual or context-heavy transactions. Treat it as a hypothesis to test, not a promise. Plug in your own numbers; don’t trust anyone else’s.
For a broader framework on when to buy automation software versus building custom, see Accounting Automation Companies: Buy, Build, or Wait.
When to Wait, When to Act
-
If you're on QuickBooks Online or Xero and happy with the GL
A custom AI agent via MCP is likely your faster path to value. You skip migration risk entirely and get AI capabilities layered on top of a system your team and clients already know.
-
If you're starting a new firm or a new CAS practice line
Evaluate AI-native platforms seriously — Sumary, or competitors as they emerge in your market. The total cost of adoption is lowest when you aren’t migrating existing clients.
-
If you want to test AI categorization without commitment
Start with a read-only MCP connection to your existing GL. Let an AI agent categorize transactions in a shadow environment, compare its output to your staff’s work, and measure accuracy before giving it any write access. This is a low-risk pilot regardless of which long-term path you choose.
-
If client-data confidentiality is your top concern
Neither path is automatically safer. An AI-native platform processes data on its infrastructure; a custom agent sends data to an AI model’s API. In both cases, review the vendor’s data-processing agreements, confirm whether data is used for model training (and opt out if so), and ensure compliance with IRS Publication 4557 and your state board’s requirements. Verify specifics with your compliance advisor.
The Honest Bottom Line
Sumary’s raise is a signal, not a call to action. AI-native accounting platforms are coming, and some will be very good. But most firms reading this already have a GL, a practice-management system, and clients who aren’t switching anytime soon. For those firms, the most practical move right now is to build AI capabilities on top of what you have — starting with read-only access, running pilots, and expanding scope only after you’ve measured accuracy and ROI with your own data.
The firms that will benefit most from this wave aren’t the ones that jump to the newest platform or build the most complex agent. They’re the ones that start small, measure honestly, and keep a human in the loop where it counts — which, in accounting, is most of the places that matter.
Not sure where to start?
Get a free automation audit: we map your bookkeeping, month-end close, client onboarding, document collection, and AP/AR — and show you what's worth automating before you spend a dollar.
Get a free automation audit