A couple of weeks ago Basis, a US-based tech startup building AI agents for accounting firms, announced its $100m funding round. A serious amount of money. This is both a strong signal of the opportunity for Agentic Accounting and a clear indication of the economic benefit flowing up to the model builders (eg. Anthropic, OpenAI).
However my mind couldn't help wonder "what's the moat" supporting that valuation? For founders, VCs and tech analysts alike, "system of record" vs. "system of action" is still a hot topic. Client data isn't it.
Every owner/partner I've met through the years believes their client relationships are core to the firm's moat. There's truth in that, but I think it sits on something deeper: the accumulated judgement behind every decision.
The partner who knows which treatment to apply before the client finishes the sentence. The handling of the exception that isn't in any SOP because it never needed to be. That's the real asset.
One thought led to another, and in the next moment I'm wondering: when you deploy an AI agent in your accountancy firm, who owns what it learns? (side note: this excellent article popped into my radar via Gary Turner, both are really great thoughtpieces)
If we’ve spoken at any point over the past couple of years, you’ll have endured my ramblings about agents and accounting, so it’s exciting to see these strong signals materialise and reassuring to know it’s not just a Quixotic vision.
However my mind couldn't help wonder "what's the moat" supporting that valuation? For founders, VCs and tech analysts alike, "system of record" vs. "system of action" is still a hot topic. Client data isn't it.
Every owner/partner I've met through the years believes their client relationships are core to the firm's moat. There's truth in that, but I think it sits on something deeper: the accumulated judgement behind every decision.
The partner who knows which treatment to apply before the client finishes the sentence. The handling of the exception that isn't in any SOP because it never needed to be. That's the real asset.
One thought led to another, and in the next moment I'm wondering: when you deploy an AI agent in your accountancy firm, who owns what it learns? (side note: this excellent article popped into my radar via Gary Turner, both are really great thoughtpieces)
If we’ve spoken at any point over the past couple of years, you’ll have endured my ramblings about agents and accounting, so it’s exciting to see these strong signals materialise and reassuring to know it’s not just a Quixotic vision.
What's an agent anyway?
Most of the AI tools in use by accountancy firms are assistants. You ask something, they answer. Useful, but passive.
An agent is a bit different. It still sits on top of an LLM (the reasoning engine like Claude Opus, GPT, etc.) but it takes action more or less autonomously. You ask it to do something, it tries to discern what the goals are, defines a plan, identifies the tools it needs, executes the plan, loops back until the results are good enough, stores future memory.
The shift from assistant to agent is effectively the shift from a (still powerful) Q&A tool you use, to a system that really does the work. This is an important distinction because what the system learns is valuable.
The key takeaway is that an agent without context is just expensive guesswork.
Which is something you need to press your agent vendors on: an agent trained on generic accounting knowledge is, at best, a decent starting point.
The thing that makes it useful and valuable and effective is the layer of context that gets built on top of it. Not just your SOPs (although these matter) but the intersections of that particular SOP for that particular client. In fact, what are the rules for this SOP in this firm, for this client, in this industry? This kind of tuning takes real work to build. And it's really never finished because clients change, partners change, the exceptions keep coming.
(One more side note: the good news is that this can now be captured in what effectively are plain-English text files, whereas 3 years ago this would be extremely expensive software source-code)
An agent is a bit different. It still sits on top of an LLM (the reasoning engine like Claude Opus, GPT, etc.) but it takes action more or less autonomously. You ask it to do something, it tries to discern what the goals are, defines a plan, identifies the tools it needs, executes the plan, loops back until the results are good enough, stores future memory.
The shift from assistant to agent is effectively the shift from a (still powerful) Q&A tool you use, to a system that really does the work. This is an important distinction because what the system learns is valuable.
The key takeaway is that an agent without context is just expensive guesswork.
Which is something you need to press your agent vendors on: an agent trained on generic accounting knowledge is, at best, a decent starting point.
The thing that makes it useful and valuable and effective is the layer of context that gets built on top of it. Not just your SOPs (although these matter) but the intersections of that particular SOP for that particular client. In fact, what are the rules for this SOP in this firm, for this client, in this industry? This kind of tuning takes real work to build. And it's really never finished because clients change, partners change, the exceptions keep coming.
(One more side note: the good news is that this can now be captured in what effectively are plain-English text files, whereas 3 years ago this would be extremely expensive software source-code)
The IP of an accounting firm was never the client data
I'm probably preaching to the choir, but it's important to flag this one up. Client data is the client's. Even if the system of record (eg. Xero) wants to lock this down behind tall paywalls, it doesn't really matter much. Thousands of businesses change GL every week, it's a known process regardless of how painful it's made to be.
The real gold is in what already belongs to you. The encoded logic that determines how that data gets processed: by whom in the firm, in what sequence, with what quality checks, and how competently are any exceptions handled. This logic partly lives in documented SOPs that apps like Pixie helped standardise. But it also lives in the heads of your experienced staff. The partner who always reviews a particular kind of work before it goes out because they've seen things and understand its many nuances. The client who needs some expense to go into COGS because of how they operate. This rarely goes into SOPs. It exists because people learned it. It's captured, if not in unstructured client notes then in people's minds.
An agent over time will pick these up too. That's the whole point. Every decision it makes, every exception treatment, goes into the context layer for that workflow, for that client, in that firm.
This is pretty much a compounding asset: the more refined and complete the knowledge, the better the outcome from the agent. In an accounting firm, this is the knowledge that takes years to develop and is likely the reason why clients stay.
So we can assume that the accumulated knowledge is one of the firm's most valuable assets. The "prompts", the decision traces, the exception handling, built over years.
And this is the bit that struck me: whoever owns this context layer effectively runs the firm's operation. Not ideal. But there's yet another angle to it..
The real gold is in what already belongs to you. The encoded logic that determines how that data gets processed: by whom in the firm, in what sequence, with what quality checks, and how competently are any exceptions handled. This logic partly lives in documented SOPs that apps like Pixie helped standardise. But it also lives in the heads of your experienced staff. The partner who always reviews a particular kind of work before it goes out because they've seen things and understand its many nuances. The client who needs some expense to go into COGS because of how they operate. This rarely goes into SOPs. It exists because people learned it. It's captured, if not in unstructured client notes then in people's minds.
An agent over time will pick these up too. That's the whole point. Every decision it makes, every exception treatment, goes into the context layer for that workflow, for that client, in that firm.
This is pretty much a compounding asset: the more refined and complete the knowledge, the better the outcome from the agent. In an accounting firm, this is the knowledge that takes years to develop and is likely the reason why clients stay.
So we can assume that the accumulated knowledge is one of the firm's most valuable assets. The "prompts", the decision traces, the exception handling, built over years.
And this is the bit that struck me: whoever owns this context layer effectively runs the firm's operation. Not ideal. But there's yet another angle to it..
The governance issue
If eg. HMRC (or ATO if you're in Australia), asks how a particular conclusion was reached, you'll very likely be pressed to demonstrate the decision trace in detail. The agent's answer has to be documented, it needs to produce and store auditable decision traces, the entire reasoning behind an output.
If a context layer is a blackbox because it's entirely shaped, held and controlled by the vendor, you may not have much in terms of regulatory defence. The vendor will just point to their terms and conditions and say they have no liability.
Quite the professional risk.
If a context layer is a blackbox because it's entirely shaped, held and controlled by the vendor, you may not have much in terms of regulatory defence. The vendor will just point to their terms and conditions and say they have no liability.
Quite the professional risk.
Back to IP retention
If you let the firm become a passive consumer of the context layer controlled by the vendor, your firm is a bit of a hostage.
You can't really just lift and shift that context.
If the service degrades vs. their competition, if they pivot strategies, etc. you can't just switch without leaving behind everything that made the agent valuable to your firm. This will have an immense negative impact on your operation.
Picture how hard it'd be if all the core people in your firm left at the same time.
And also, the incentive problem: a vendor sitting on top of encoded expertise of hundreds of accounting firms has something way more valuable than the SaaS subscription revenue itself. They have the aggregate knowledge to do the work itself and cut off the middleman.
Onshore is a good example. They spent a few years building and selling R&D tax software to accounting firms, but found tremendous growth just going direct to the corporate end clients and now effectively compete with the firms they once served.
I don't think this will be unusual, in the end those VC dollars have to return tenfold to their LPs.
So the next time an agentic AI vendor comes pitch you, here's what I'd ask:
You can't really just lift and shift that context.
If the service degrades vs. their competition, if they pivot strategies, etc. you can't just switch without leaving behind everything that made the agent valuable to your firm. This will have an immense negative impact on your operation.
Picture how hard it'd be if all the core people in your firm left at the same time.
And also, the incentive problem: a vendor sitting on top of encoded expertise of hundreds of accounting firms has something way more valuable than the SaaS subscription revenue itself. They have the aggregate knowledge to do the work itself and cut off the middleman.
Onshore is a good example. They spent a few years building and selling R&D tax software to accounting firms, but found tremendous growth just going direct to the corporate end clients and now effectively compete with the firms they once served.
I don't think this will be unusual, in the end those VC dollars have to return tenfold to their LPs.
So the next time an agentic AI vendor comes pitch you, here's what I'd ask:
- How does your firm retain full ownership of the context layer?
- How is this context layer ring-fenced, contractually, from the vendor's own model tunings/product development?
- How much control will you have of what and how new tunings or system prompts get rolled out into your workspace, to avoid a decision today contradicting the same decision yesterday?
- What open format can this context layer be exported to/imported from? If the answer here is anything but text files, do not accept it.
- Will the vendor record and give you full access to the complete decision traces of its agents, at any time?
If the vendor turns hand-wavy or looks uncomfortable, that'll be the clearest signal that they haven't thought much about this. At least not yet.