Last week I was given the opportunity to sit down with a true accounting tech early-adopter, Dave Sellick, who runs Sidgrove, a UK-based boutique CFO/accountancy firm, and chat about agentic AI.
Dave was sceptical. Well, maybe he still is... but not hostile or dismissive of the technology. Really just sceptical. As for me, I love this type of conversation: it helps me identify and work through my own blind spots. Dave's position is very reasonable: "what does any of this mean for someone running a firm?"
That's a valid question. And I think people in my industry (ie. software vendors) are doing a bit of an underwhelming job of explaining it.
Although our chat covered a lot of ground, we kept coming back to a couple of themes:
Dave was sceptical. Well, maybe he still is... but not hostile or dismissive of the technology. Really just sceptical. As for me, I love this type of conversation: it helps me identify and work through my own blind spots. Dave's position is very reasonable: "what does any of this mean for someone running a firm?"
That's a valid question. And I think people in my industry (ie. software vendors) are doing a bit of an underwhelming job of explaining it.
Although our chat covered a lot of ground, we kept coming back to a couple of themes:
- how agents will impact the work you do
- how it impacts the business you own
How agents will impact the work you do
Steve Jobs had this line I love coming back to:
you've got to start with the customer experience and work back to the technology
What I see is some vendors doing more than a bit of the opposite and that's not right. They're leading with the technical capability and hope the accountant (customer) works backwards to value.
Agents are the most recent stage of generative AI. There's very little magic to it, it's all the same underlying tech.
Stage one: we got Q&A prompting. ChatGPT and similar tools, from around 2022-2023. We asked a question, the model gave us an answer. Useful for research, drafting, general knowledge. A better Google.
Stage two: copilots. Then we figured "if we organise all of a company's proprietary knowledge in this way, we can make the LLM query it and use it to generate answers". The chats got embedded inside apps and we called it a Copilot. It inherits an app's permissions and data, so it can draft emails in your tone of voice, summarise calls, or surface insights without leaving the host software. Many practice management tools have added these as you have probably noticed. Autonomy is very limited, in fact you can say it's quite transactional: give it something and it'll give you something back.
Stage three: agents. This is the shift. This is the thing that got us all excited at Pixie back in the day when GPT3.5 came out. An agent can design its own workflow based on a goal you give it, the boundaries and the quality thresholds you define, follow your policies and adhere to your review process. The agent works out the steps, executes them, and adapts as it goes.
Succinctly: while a copilot helps you do one task faster, an agent should handle a workflow end-to-end helping you do eg. 50 jobs in parallel.
What it can look like in practice
I'll try to go through something concrete: delivering monthly management accounts (MMA) for clients.
An agent given this goal does not just speed up one step. It works backwards from the deliverable and figures out what needs to happen:
- reconcile the transactions, matching bank feeds to the ledger
- categorising unmatched items and flagging anomalies against the client's historical patterns
- (if you let it) asking the client to clarify expenses or confirm those anomalies
- prepare monthly workpapers: prepayments and accruals schedules, fixed asset depreciation, tax reconciliation if/as needed
- apply any month-end adjustments: journals for accruals, depreciation, provisions, based on the firm's methodology and the client's recurring patterns.
- review the trial balance against prior periods and budget. Then compile the management accounts themselves: P&L, balance sheet, cash flow, variance analysis, and draft the narrative commentary.
- at each of these steps, a good agent will have its own review loop, sending things back if the result is not up to standard
Each of these steps can be executed by an agent skill, or a specialised "micro-agent" that's configured to do one thing really, really well. The MMA Agent's job is to figure the right workflow for the outcome (ie. what's needed to produce the monthly management accounts), which skills to chain and how to execute the workflow passing the right context to each step.
And a well-built agent will execute under a solid control model that also includes full audit trail at every step. What it did, why it did it, what data was used, who authorised certain actions (like changing source records or sending client-facing outputs), and what rules were applied. If there's no trail to show a regulator, the agent is not really ready for production. Make it a threshold requirement.
If you want to dig into the detail, drop me a line. I'm always up for this conversation.
So although there's a bit more depth to it, but that's the gist of it.
This is the point that matters. An agent that reconciles transactions faster is useful but limited. To me the real value is in handling the workflow. From reconciliation to the finished accounts.
And I'm not a proponent of an agent running autonomously in the background, monitoring everything like some kind of digital overseer.
In many scenarios I think it's preferable for the user to invoke the agent manually, then review outputs at key checkpoints to approve permanent changes to the clients' records, and also approve the final deliverable.
You set the boundaries. The agent operates within them.
In many scenarios I think it's preferable for the user to invoke the agent manually, then review outputs at key checkpoints to approve permanent changes to the clients' records, and also approve the final deliverable.
You set the boundaries. The agent operates within them.
From a software vendor's point of view, we should look again at the Steve Jobs quote and start with the customer experience. That whole workflow can potentially be condensed into a single "Generate Monthly Management Accounts" action in your app's UI that, when clicked, triggers all the activity in the background. This is in contrast to having to use multiple features across different modules/apps to achieve the same outcome.
Same outcome to the user with orders of magnitude less work.
For firm partners and owners, the right strategic question then isn't so much "should I adopt agents?" but rather "which parts of my software stack will create bottlenecks for agents?"
This is what you'll want to avoid and should be part of every vendor renewal conversation.
Same outcome to the user with orders of magnitude less work.
For firm partners and owners, the right strategic question then isn't so much "should I adopt agents?" but rather "which parts of my software stack will create bottlenecks for agents?"
This is what you'll want to avoid and should be part of every vendor renewal conversation.
How it impacts the business you own
The debate making the rounds is whether AI will replace accountants[1]. We're trying to draw parallels from what's happening in software building (engineering, design, product management, etc.) and projecting it to other professions.
I think we need to be careful with this leap.
In the software industry, we're probably in the preceding stages of something that will look more like what followed the dot-com bust or the 2008 crash: more startups, each doing one thing well with a product targeting a specific niche.[2] We just do not have the hindsight yet.
I spent years building practice management software for accountancy firms. I've seen more than a few operational/workflow differences between a bookkeeping firm, tax preparation shop, and a full-service accountancy firm. Even then, a firm that services trucking companies works differently from one that supports creative agencies, and different again from another one focused on ecommerce.
I genuinely no longer think it'll be a story with winners and losers.
The partner running a 200-person compliance-heavy firm may end up running a 40-person advisory firm with far better margins and better lifestyle. The senior manager who felt stuck may find themselves leading a five-person specialist practice that would not have been viable a decade ago. The vessel (ie. the firm) changes shape but the work does not disappear.
Same GDP, different structure. The profession's output does not shrink. It redistributes. And this is not new. Industries fragment and consolidate in cycles. Give it a couple of decades and some of those specialist firms will likely roll up again. Ebb and flow.
I think we need to be careful with this leap.
In the software industry, we're probably in the preceding stages of something that will look more like what followed the dot-com bust or the 2008 crash: more startups, each doing one thing well with a product targeting a specific niche.[2] We just do not have the hindsight yet.
I spent years building practice management software for accountancy firms. I've seen more than a few operational/workflow differences between a bookkeeping firm, tax preparation shop, and a full-service accountancy firm. Even then, a firm that services trucking companies works differently from one that supports creative agencies, and different again from another one focused on ecommerce.
I genuinely no longer think it'll be a story with winners and losers.
The partner running a 200-person compliance-heavy firm may end up running a 40-person advisory firm with far better margins and better lifestyle. The senior manager who felt stuck may find themselves leading a five-person specialist practice that would not have been viable a decade ago. The vessel (ie. the firm) changes shape but the work does not disappear.
Same GDP, different structure. The profession's output does not shrink. It redistributes. And this is not new. Industries fragment and consolidate in cycles. Give it a couple of decades and some of those specialist firms will likely roll up again. Ebb and flow.
So the question is not so much whether your firm survives this. It is what your firm will look like on the other side.
---
1 - there's a contradiction in this vendor narrative that I'm trying to understand. We're seeing firms being told "adopt or die." But if firms were truly going to disappear, who exactly are these vendors building for?
The only way accountancy firms can genuinely lose is if direct-to-client software becomes good enough, easy enough, and its vendor is willing enough to assume the liability for poor advice and non-compliance. If that is the bet, selling agentic software to accounting firms is a strange business model.
2 - not even joking about this but on LinkedIn I just read about a CRM specifically for recruitment firms servicing the fitness industry. Talk about a niche of a niche... If I'm running such an agency, why would I buy Hubspot rather than this product that "just works" for my business?
1 - there's a contradiction in this vendor narrative that I'm trying to understand. We're seeing firms being told "adopt or die." But if firms were truly going to disappear, who exactly are these vendors building for?
The only way accountancy firms can genuinely lose is if direct-to-client software becomes good enough, easy enough, and its vendor is willing enough to assume the liability for poor advice and non-compliance. If that is the bet, selling agentic software to accounting firms is a strange business model.
2 - not even joking about this but on LinkedIn I just read about a CRM specifically for recruitment firms servicing the fitness industry. Talk about a niche of a niche... If I'm running such an agency, why would I buy Hubspot rather than this product that "just works" for my business?