GPT-6 Astra raises a practical question for accounting firms: what changes when producing the first version of the work becomes much easier? Someone would still need to decide whether the records were complete, whether the conclusions made sense, and what the client should do with the information. But the number of people involved, the skills they needed and the way the firm charged could all change.
At Fintant, we think that is the conversation accountants need to have about increasingly capable AI. There is a real opportunity to make useful financial work more accessible, alongside a real possibility that some of today's work will command less money. We should be willing to examine both.
For accountants, the distinction between today's AI capabilities and the prospect of artificial general intelligence matters. The changes firms can make now and the decisions they may face later deserve separate attention.
Why Astra changes the discussion
OpenAI introduced GPT-6 Astra in September 2026, describing improvements in computer use, multistep workflows and the creation of spreadsheets, documents and presentations. These are OpenAI's capability claims, rather than results from accounting engagements.
What interests us is the potential to connect activities that usually sit between different files and applications. Suppose a firm needs to compare supporting records, investigate discrepancies and assemble a workpaper. A system that can carry context between those steps could reduce the effort involved in getting the work to a reviewer.
We would still want evidence from that firm's actual workflow before putting a number on the benefit. A polished spreadsheet tells us very little about whether an unusual transaction was understood or a missing document was noticed. The possibility of handling a longer assignment is significant, but the work has to survive checking.
We do not need an AGI verdict to take this seriously
AGI means artificial general intelligence. In OpenAI's Charter, it refers to highly autonomous systems that outperform humans across most economically valuable work. That is a much broader proposition than being very good at preparing a spreadsheet.
Astra has produced striking benchmark results, but the detail matters. ARC Prize reports a best observed score of 62.7% on ARC-AGI-3 Semi-Private with its Standard setup, compared with 99.9% using a different setup that preserves reasoning state between requests. The surrounding software and context management are part of the result.
ARC Prize also explicitly says it is not claiming Astra is AGI. Its environments have bounded rules and goals, while real work is more open-ended.
Our view is that accountants can take the progress seriously without treating a benchmark as a professional qualification. The more useful test is whether a system can complete a defined assignment with dependable evidence, appropriate limits and an acceptable total cost. That test remains useful whichever label eventually wins the debate.
Better accuracy does not guarantee faster work
In their 2026 paper, Human + AI in Accounting, Jung Ho Choi and Chloe Xie examine how AI assistance affects accounting work. It combines a survey of 277 accountants, platform records covering 79 US client businesses, and an experiment involving 99 accountants.
In that experiment, AI assistance improved transaction-classification accuracy by roughly 18 percentage points, although overall completion time did not significantly improve. The authors also identify the risk of accountants following problematic AI recommendations.
This is evidence from a particular platform and task, predating Astra. It does not establish what Astra would deliver or account for all adoption and operating costs.
For a firm adopting AI, the practical question is which part of the service actually improves. Better accuracy, faster preparation and lower delivery cost are different outcomes. We would want to know which one improved before claiming that an AI workflow had made a firm more productive.
The value of accounting will need a clearer explanation
If AI makes preparation cheaper, we think firms will face a harder question about what clients are paying for. That does not mean every fee should fall, but it does mean that time spent may become a weaker explanation of value.
Suppose a firm can prepare the same monthly reporting package with fewer hours of work. Under an hourly arrangement, that could reduce revenue unless the scope or demand changes. Under a fixed fee, it could improve the margin, provided review, software and correction costs do not absorb the saving. If competing firms pass on similar savings, the pricing advantage may also narrow.
These are possible commercial outcomes, not a prediction that one pricing model will suit every practice. An efficiency gain also creates a business decision. The firm can use the capacity to serve more clients, improve the service, reduce workload or lower costs, but each choice has different consequences.
We also think the familiar advice to “move into advisory” needs more thought. An owner may value earlier notice of a cash problem, a clear explanation of a margin change or help weighing a business decision. That demand has to be understood and served. Calling an additional report advisory does not establish that anyone wants it, and more capable AI may make parts of analytical work cheaper too.
Professional judgment has to remain something we practise
We are cautious about building the profession's future around the claim that AI will never be able to exercise judgment. That is a prediction about future technology, and it gives accountants little guidance about what to improve today.
We would put more emphasis on the ability to establish what a conclusion depends on, challenge it with evidence and explain why it is appropriate in the circumstances. A reviewer who accepts a convincing answer without checking its basis has added little assurance simply by being human.
ACCA's analysis of changing accounting work anticipates less routine processing and changing responsibilities around judgment, controls and technology. It also recognises that data, costs and operational constraints affect how quickly automation can spread.
For us, the practical implication is that review needs to be designed into the work. The accountant should be able to see what information was used, which conclusions remain uncertain and what requires a decision. Access to software should follow the authority granted for the engagement. Preparation should not silently become approval to post, file, make a payment or communicate with an accounting firm's client.
We should be honest about the effect on careers
We do not think it is credible to promise that every accounting role will remain unchanged. If fewer hours are needed for a service and demand does not expand enough to absorb the capacity, staffing could be affected. The existence of a human sign-off does not guarantee the same number of jobs behind it.
The ILO's 2025 research treats job transformation as the most likely overall effect of generative AI, while identifying clerical work as particularly exposed. It measures potential exposure, not actual layoffs, and it predates Astra.
We would also take the training consequences seriously. If juniors see mostly completed outputs, firms need another way to help them learn what good work looks like and why errors happen. Asking someone to form an independent view, compare it with an AI suggestion and explain the difference could be part of that process. The aim should be to build understanding alongside tool proficiency.
This matters to experienced accountants as well. Knowing a process thoroughly can help us evaluate automation, but experience should also make us willing to change a process when the evidence supports it.
Preparing an accounting practice for more capable AI
If AGI eventually becomes capable of performing most economically valuable work, accounting's exposure could extend well beyond data entry. Analysis, review and advisory preparation would also need to be reconsidered. We cannot establish an arrival date from the evidence available, and we would not assume that today's division between human and machine work is permanent.
Our recommendation is to build the ability to adapt through specific, measured changes. Start with one recurring workflow and define what acceptable work looks like. Compare the complete effort required to reach that standard, including checking and correction. Use approved data and access, keep consequential decisions with the authorised people, and examine how the change affects both delivery and learning.
For an individual accountant, we would pair AI fluency with stronger accounting fundamentals and clearer communication. It should be possible to explain a conclusion to a colleague or client without relying on the fact that a tool produced it. For a firm owner, we would connect the same exercise to pricing and client needs, so that faster work leads to an intentional business choice.
As AI becomes more capable, we believe firms need to reconsider how accounting work is organised and valued, while testing whether changes improve the service clients receive. A useful starting question for the next team discussion is: if preparation became much easier, what would our clients still need us to understand, decide and explain?