AI Agents for Business Keep Stalling at Your Books
A Brussels company called Chift closed a 10.5 million euro Series A on Monday. BlackFin Capital Partners wrote the lead cheque, joined by earlier backers Entourage, Shapers, Seeder Fund and Wallonie Entreprendre. What Chift sells is a single doorway into the systems small companies actually run on, from the bookkeeping ledger and the billing tool to the card reader, the online storefront and whatever software a landlord uses.
That description sounds like plumbing because it is plumbing. It also explains something frustrating that many owners have run into this year, which is that the AI assistant you bought works beautifully right up until the moment it needs to touch your actual books.
The Gap Nobody Warned You About
Running a small company means living inside a handful of disconnected tools. Your accountant uses one system, your online store uses another, your card terminal reports somewhere else, and none of them were designed to introduce themselves.
An AI assistant runs straight into that wall. It can draft your invoice reminders, but it cannot see which invoices were actually paid unless something gives it permission and a route into the accounting system.
This is why so many AI tools for small business feel impressive in a demo and thin in daily use. The intelligence was never the bottleneck. The access was.
What Chift Is Building Underneath
The company puts its count of hooked-up financial platforms above 120. Roughly 150 software vendors now build on top of that layer, and through them the reach extends past 50,000 smaller firms spread across a dozen or so European markets. Revolut, Qonto, Pennylane and Mollie all sit on the customer roster.
Back in 2024 the company took a seed round worth 2.3 million euros, and the top line has multiplied more than ten times over since. Total disclosed funding now sits around 12.8 million euros. Those figures come from the company, so read them as reported rather than audited.
The piece that matters most for owners is newer. There is now a Model Context Protocol server sitting in the stack, which lets AI agents read and change financial records while permissions and sign-in get handled properly underneath.
| Measure | Figure |
|---|---|
| Series A raised | 10.5 million euros |
| Financial systems connected | More than 120 |
| Software companies using it | More than 150 |
| Small businesses reached | More than 50,000 |
| European countries covered | More than 10 |
Permissions Are the Real Product
Here is the part that deserves your attention. An agent that can read your bank feed and edit your invoices holds far more power than a chatbot that answers questions.
So the interesting question stopped being what the agent can do. It became who decided what the agent is allowed to reach, and whether anyone can review that decision afterward.
Owners have already learned this lesson the hard way through suppliers. The habits behind good vendor risk management apply exactly as well to software agents, and the stakes are similar when the connected system holds your customer payment records.
What to Do Before You Connect Anything
Start by listing every tool that currently holds financial data about your business. Most owners are surprised by the length of that list, and you cannot govern access you have not written down.
Then check what each AI tool is actually asking for. Read permissions are very different from write permissions, and any assistant requesting the ability to change records should earn that access rather than receive it by default.
Finally, keep a human in the approval loop for anything involving money leaving the business. That single rule prevents most of the failures described in reporting on weak AI guardrails, and it costs you almost nothing to maintain.
- Inventory every tool holding your financial data.
- Separate read access from write access deliberately.
- Require human approval on outbound payments.
Why Investors Are Funding the Boring Layer
Try selling financial software across Europe and you inherit a mess. Every country runs on different systems, plenty of them decades old, so vendors routinely burn years wiring up each connection and then babysitting it. Infrastructure that erases that work has obvious value.
AI makes those connections more valuable rather than less. An agent that cannot get securely into the ledger, the billing tool and the payment rail simply has nowhere to do the job, which suggests the next wave of durable companies may sit a layer or two below the assistants everyone is talking about.
Standards bodies are catching up as well. The Model Context Protocol specification is now the common way tools expose data to agents, which should make future switching easier.
Signals to Watch From Here
Watch whether US accounting and payment platforms adopt similar agent-ready connections. European fragmentation created the need first, but American owners juggle nearly as many disconnected systems.
Watch pricing too. Integration infrastructure tends to get cheaper as it commoditizes, and cheaper plumbing eventually shows up as cheaper software on your monthly bill.
Owner Questions About AI Access
Should I let an AI agent into my accounting system?
Read-only access for reporting is a reasonable place to start. Grant write access only after you have tested the tool and confirmed you can review what it changed.
What is a Model Context Protocol server?
It is a standard way for a software system to expose data and actions to AI agents, with authentication and permissions handled explicitly rather than improvised.
How do I audit what an AI tool touched?
Ask the vendor for an activity log before you sign. If they cannot produce one, treat that as a reason to keep the tool away from your financial systems.
