Intuit surveyed 725 US accounting and bookkeeping professionals in May 2026. Three findings from it, read together, explain more about agentic AI in accounting than most of what has been written about agentic AI in accounting.
Situational Use Is Not a Small Version of Embedded Use
Eighty-eight percent use AI for at least one client service. Eighty-six percent use it for at least one firm operation. And only 30% have AI fully embedded as their default, while 54% use it situationally.
It is tempting to read those numbers as a maturity curve, with the 54% on their way to becoming the 30%. They are not on that road, because situational use and embedded use are structurally different things.
Situational means a person decides, task by task, to open a tool. Somebody is stuck on a classification, so they ask. Somebody needs a first draft of a client email, so they generate one. The judgment about when to use AI sits in a human head, freshly, every time.
That is genuinely valuable. Seventy-five percent of AI users in the same survey said the tools delivered more value than expected, and situational use is where most of that came from. It is also invisible to the firm. It leaves no artefact, no defined step, and nothing anyone else can inherit.
What Agentic Actually Requires
An agent is not a better chatbot. An agent is a process that runs without being asked, and running without being asked has four preconditions.
It needs a trigger, so something has to define when the work starts. It needs inputs it can reach on its own, because an agent that waits for someone to paste in a document is a chatbot with extra steps. It needs a defined output, so there is a thing to produce rather than a conversation to have. And it needs a named reviewer, because in this profession somebody signs.
Situational use supplies none of those, by definition. The trigger is a person's judgment, the inputs arrive by hand, the output is whatever the conversation produced, and the reviewer is the same person who asked.
So a firm at 54% situational use has not built a foundation for agents. It has built a set of individual habits, which is a different and more fragile asset.
The Precondition Nobody Costs
Of those four requirements, three are process decisions a firm can make in a meeting. Inputs are the one that costs money.
Look at what the same survey found firms using AI for on the client side: data entry and processing at 54%, financial forecasting at 51%, real-time insights at 49%. Every one of those is input-hungry. None of them can run on a prompt alone. They need the client's statements, invoices, contracts and prior filings, and they need them reachable at the moment the work runs rather than assembled by a person first.
This is where agentic projects quietly die in accounting firms. The pilot works, because during the pilot somebody prepares the inputs by hand. Then it is supposed to run on Tuesdays without that person, and the inputs are not there in a form anything can reach, so the agent either stops or produces something confident and wrong.
An agent is only as autonomous as its access to documents.
The Honest Sequence
Which suggests an order of operations that is less exciting than the vendor version and considerably more likely to work.
Make the inputs retrievable first. Then define one step precisely enough that a stranger could perform it. Then automate that step, with a named reviewer. One step, working, beats a platform that assumes a process the firm does not have.
MetaWurks is built for the first of those. It ingests a client's statements, invoices, contracts, returns and correspondence, and makes the whole set queryable in plain English, so the material an automated step needs is reachable rather than assembled. Role based access controls decide who can open which client's records, audit logs record who opened what and when, and documents ingested into the platform are not used to train models or exposed to other users, which matters when something other than a person is doing the reading.
It is not an agent and it does not replace judgment. What it removes is the assumption underneath every agentic pitch: that the documents an automated process needs are already somewhere a process can get at them.
Eighty-eight percent of the profession has proved AI is useful. The 30% figure is the one that decides whether any of it can run without being asked.
Join the Conversation
Pick the AI use in your firm that saves the most time today. Could it run on Tuesday without the person who currently starts it, and if not, what is missing?