What happens when the accounting system stops being somewhere you go?

The accounting system becomes infrastructure and the assistant becomes the workspace. AI prepares the transaction and puts it in front of the accountant, who reviews it and posts it. What may change is how often a person has to go to the system at all.

For 30 years, using business software has meant the same thing. Open a browser, log in, click through a few screens, do the work. Accounting systems, CRMs, banking platforms: different software, same model.

I recently watched a presentation on MCP Apps that put forward a different idea. What if the assistant becomes the place you work, and the applications sit underneath it? Instead of going to five systems, your assistant goes to them for you. That made me think about accounting.

What happened when I tried it?

The functionality rolled out overnight, so I got up early on Sunday morning to test it. Being an accountant makes me a bit of a pessimist, so I went in expecting to find holes. But AI has improved so quickly over the last few years that another part of me was thinking: give this the right data, context and guardrails and it is probably going to work pretty well.

My first prompt was: "Can you help me draft a journal entry?"

It asked what the transaction was. I explained it. Straight away it started referring to accounts from our actual chart of accounts. No padded explanation of what a journal entry is. It identified the relevant accounts, picked up that direct posting was blocked and that the transaction needed to go through the relevant subledger, and shortly afterwards the journal was sitting in our accounting system, ready for me to review.

The clever part was not that a language model knew the debit and the credit. We have been able to ask AI accounting questions for years. This time it was doing the work inside our environment.

Why won't the AI be posting the journal?

There is a line here that is clear for us. AI can prepare the accounting. A human posts it. I do not see Flagship handing that responsibility over.

One journal entry looks insignificant on its own, but it does not stay one journal entry. It flows into a quarterly NAV, into investor statements, into audited financial statements and into the data our clients query. There is a lot sitting downstream of one debit and one credit, so someone needs to look at it, approve it and be accountable for it. That is not an AI question for us. It is part of the service. Clients pay for peace of mind, and there has to be a person who can stand behind the books.

That is how I separate the two uses of AI in our business. Helping a client find an answer without knowing which report holds it is something we already do and want to do more of. Doing the accounting is a different thing, and it stays with a person.

What work doesn't need an accountant?

Accounting contains a lot of boring work. Invoice number, invoice date, payment date, amount, vendor, description, account. All of it needs to be right, and very little of it needs a qualified accountant copying information from a PDF into another application. There isn't much glory in eliminating that work, but there is a lot of value in it. If the AI can prepare the transaction in seconds, the accountant can review it while they are still thinking about it, instead of picking it up hours later when the context has gone. The admin work goes away and the judgment stays with the professional. The machine does not get bored on invoice 75, get tired at 5pm or transpose two digits.

The journal entry is probably the boring use case. What interests me more is the assistant questioning the accounting as well as preparing it. "You treated a similar invoice differently last quarter. Is that intentional?" "This invoice relates to last quarter and there is an accrual still sitting on the balance sheet. Do you want to reverse it?" "This vendor has historically been coded somewhere else. Is there a reason this one is different?"

The accountant still makes the decision. But the assistant has the supporting document, the historic transactions and the accounting records in front of it at once, and can point out something worth a look. Strip out the work that does not need judgment, and give the professional more information at the moment judgment is required.

Then there is the data

This could matter more than the efficiency. There has always been a cost to capturing data, so accounting systems naturally tend to capture what is needed to do the accounting. AI changes that. If a document is already being read, why only extract the invoice number, date and amount? Why not payment terms, counterparties, relevant periods, investment attributes, contractual dates and relationships to other documents? The incremental effort starts getting very small. So we get faster accounting, and we also get better data. Better data is useful well beyond the finance team: investor relations, liquidity management, investment monitoring, reporting.

Fund administration has always produced a huge amount of valuable data, most of it buried in accounting systems, spreadsheets and PDFs. If AI makes it economical to capture and easy to get at, the administrator can become a much more useful data partner to the investment manager.

We may have got one thing slightly wrong

Earlier this year we launched Clarity, which lets clients have a conversation with their fund data. I still think the idea is right. A client should be able to ask a question about their fund and get the answer, without knowing which report to open.

What I am less convinced about is whether they should have to come to our application to do it. We may have been right about the destination and wrong about the interface. Every software company has been building its own AI, and I am increasingly wondering whether that is backwards. Maybe the client has their AI, and our job is to make our data and services available to it. The client shouldn't necessarily need a Flagship AI, a bank AI, a legal AI and ten other assistants. They may just want their assistant to be able to ask Flagship, securely, for the information it needs.

In that world, the most important thing we built through Clarity isn't necessarily the chat window. It is the work underneath it: structuring the data, giving the model the right context, keeping clients separated, controlling permissions, working out how an AI should interpret fund accounting information. Those things matter regardless of which assistant sits on top. Clarity page

The systems aren't going anywhere

None of this means the accounting system disappears. It remains the system of record: it holds the data, applies the controls and keeps the audit trail. My Sunday test still ended with me opening the accounting application and reviewing the journal before anything happened, and I expect we will be doing that for a while. But it is now easy to imagine doing most of the work somewhere else. A few years ago I assumed I would always Google questions. Now I ask ChatGPT without thinking about it. I don't think it is crazy to suggest the same thing will happen to a lot of the software we currently spend our days clicking around in.

I am not interested in AI replacing accountants. I am interested in taking away the work that does not need one, in giving accountants better information at the point they need to exercise judgment, and in capturing more of the data that already passes through our hands so our clients can actually use it elsewhere in their business. We intend to start bringing your fund data directly into your own assistant later this year. Flagship MCP page