An accountant’s name goes on the work. Not the software’s or the model’s. That single fact shaped every decision we’ve made about letting AI agents near a client file.
We’ve been building agentic AI into Silverfin and running early access with real firms. Before any of it went near client work, we set ourselves three conditions. And if a capability doesn’t meet all three, it doesn’t ship yet.
1. You stay in control, and so does your data
The scariest version of an AI agent is the one with keys to everything. It reaches into systems you can’t see, changes things you didn’t ask it to, and you find out afterwards. That’s not the AI agent we built.
The agent gets the same access rights as the person using it, and nothing more. It can only reach what you could reach yourself, so it works as an extension of you rather than a separate system sitting above you.
Control also means visibility. You see the steps the agent takes and the changes it proposes before anything lands. Nothing is written into your work or removed from it behind your back. At the system level we hold to the transparency expectations of the EU AI Act, then go further and explain in plain terms how it works rather than asking you to trust a black box.
In summary, access is scoped to the individual user and every change arrives as a proposal you can inspect and turn down. If either isn’t true, it isn’t ready.
2. A human stays accountable
The signoff is a person putting their judgement and their reputation behind the numbers, and AI helping with the work doesn’t change who’s answerable for it.
That shaped one of our key design decisions. The easy path was to let the agent write straight into the templates and working papers but we chose the harder one. The agent creates proposals, data to add or change, and none of it enters your work until you accept it. You see how it was prepared, decide whether it’s good enough, and only then bring it in and mark the page as prepared.
The decision stays human. AI does the legwork of proposing, the accountant does the judging and gives the sign-off. That’s what we mean by building AI to support accountants, not replace them. The tool gets faster, but the stamp of approval stays with a person who can stand behind it.
3. It has to actually save you time
A capability can be secure, controlled and accountable and still not be worth having. If checking the agent’s work takes as long as doing it yourself, you’ve gained nothing. You’ve just moved the effort around.
So the third condition is a simple sum. The AI’s time plus your review time has to come in under what the manual job would have cost. If it doesn’t, we hold it back.
We won’t pretend we’ve proven that yet. That’s the honest reason the early access programme exists. We want to see what real usage looks like in someone’s actual working day, not a controlled demo, and we’re watching for the trap of speeding one step up while quietly creating work somewhere else. We have confidence that we’ll be able to reach this bar, but we want to take a robust approach and be able to measure it before wider releases.
Who we build on
Those three are about each capability we ship. There’s a fourth commitment that sits underneath them, about who we build on.
“AI infrastructure is heavily subsidised right now. The big providers aren’t charging the true cost of running these systems. They’re buying market share and embedding themselves into workflows so that switching gets harder over time. Today’s pricing isn’t tomorrow’s, and anyone who builds a deep dependency on a single provider is taking on a risk that isn’t fully visible yet.”
So we build to avoid that lock-in. We connect to a wider ecosystem of models rather than betting everything on one, much as open APIs once let firms build best-of-breed stacks without becoming captive to a single vendor. We use whichever model is genuinely best for a task, including the domain-specific agents starting to outperform general ones, and we keep the freedom to switch as prices and options change. Plenty of vendors are building lock-in on purpose. It might look like a feature today. In a few years it may look like a liability.
The pattern underneath
Look across all four and the thread is the same. On each one we took the harder path on purpose. Scoped access instead of global control. Proposals you approve instead of direct writes. A time-savings bar we have to clear rather than a headline we get to announce. An open approach to providers instead of a convenient lock-in.
We do this because it’s what it takes to put something in front of an accountant that they can trust with a client’s file.
Over the coming weeks we’ll share what our early access customers are telling us after their first month, including the parts that surprised us. For the background on what we’re piloting and why, our previous post walks through it.
What agentic AI actually means for accounting firms
