Every accountant has heard it, usually with a sigh. “I told you that already.” The truck was sold in March, the new lease started in July, and the owner lent the company $40,000 and mentioned it on a call in April. The client isn’t wrong. They did tell you, and when the preparer needed it, nobody could find it, so the client was asked again.
On October 1, Mercor, the company that recruits experts to build tests for the AI labs, published a study of AI against accountants that turned that sentence into a controlled experiment. Twelve licensed CPAs, averaging five and a half years of experience, each worked four month-end-close tasks on simulated companies. Every answer was in the files Mercor shipped, each CPA got a map of what the files held, and each task had a three-hour window. Working alone, the accountants met about 37% of the grading criteria. Claude Opus 5, working alone as an agent (software that opens the files, works through the steps, and submits the answer with nobody driving), met 100% on all 20 attempts.
Then the timing. The CPAs mostly took 30 minutes to the full three hours, and Claude took under 10 minutes a task. Even on the items the humans tended to get right, Mercor measured the model at 34 times faster.
So this wasn’t a test of accountants. It was a test of whether information a client has already given you is somewhere your process actually looks, and most firms would fail it the same way the CPAs did.
The fine print is real, and I covered it on Monday
Monday’s roundup worked through the caveats. These were experienced CPAs, 10 of the 12 senior or manager by title, not the juniors in Mercor’s headline. The tasks came from a benchmark (a fixed set of tasks used to score AI models against each other) that was built to make models fail, and the scoring cascaded, so one missed trap could sink the score. Nobody was allowed to ask a question, and Mercor’s own benchmark paper concedes that knowing when to ask one is “a skill that real staff accountants exercise constantly.” None of it changes what follows.
The humans had everything they needed, and they missed it anyway
Strip away the arguments about the test and one fact is left. Every answer was in the folder, and Mercor rechecked that with the accountants who wrote the tasks after the results came in. Twelve experienced people, paid for the full three hours and paid a bonus for every item they got right, routinely didn’t find it.
Mercor’s paper explains the mechanism: “mistakes cascade,” and “missing one of these hard-to-spot errors could result in a very low score.” One task author observed that participants “tended to settle early on an approach that looked reasonable and carry it through without checking it back against the instructions.” Both are what happens when a preparer doesn’t go back to the source for every item, and going back to the source, every time, is the step a machine does without noticing the cost. The paper puts “reading 100 Excel files in under a minute” on its short list of tasks that only AI can do.
Settling early isn’t carelessness. It’s how a preparer works under time pressure in a real firm, because a real firm has a shortcut: you ask. You ask the client, you ask the colleague who had the file last year, you ask the manager. Mercor removed the shortcut, and what it exposed is that exhaustive search through a pile of documents is something people are bad at and machines are not.
The client who says “I told you that already” is describing the same failure from the other side. The fact was given once, in an email in February or on a call in April, and the preparer who needed it in September didn’t look there, or couldn’t. So the firm asked twice, and the client trusted us a little less.
“Don’t you guys talk to each other?” is the wrong question
There’s a second version of the sentence, and it’s harsher. The client told one member of your team something, and months later a different member asks for it. “Don’t you guys talk to each other?” The honest answer is that the first person wasn’t careless. The fact wasn’t relevant to them at the time, so it went nowhere, and there was no place to put it where the person who’d need it later would look.
That’s a storage problem wearing a communication costume. Telling the team to talk more won’t fix it, because nobody knows in February which facts will matter in September. What fixes it is a place where everything a client tells the firm lands, and a search of that place that doesn’t depend on the right person remembering.
Clients tell you things in six places, and your process looks in one
Think about where a client’s information actually arrives. Email, to whoever they happened to write to. The chat thread in your client portal or practice-management system. The document submission system where they dropped the bank statements and a note nobody read.
The transcript of the planning meeting, a text to the partner’s phone, and last year’s file, where the prior preparer left the answer in a cell comment. The preparer who sits down to a year-end close or a tax return looks in one of those places, maybe two.
An agent doesn’t have to choose among the places you’ve connected, which is why the inventory of those places is the first decision below. Treat the year-end close or the tax return as one process, and its first step is a search across every connected place that client has ever contributed, before a single question goes out. The nuance that sat in a message to someone else is now in reach, and the client is asked once. The questions that do go out are the ones nobody has answered, which a client can’t reasonably resent.
Mercor’s folder came certified complete and with a map. Your archive has duplicates, superseded facts, and no map, and closing that gap is the real work; the search result is the reward for doing it, not a substitute.
The vendors can see where the data is. Gusto’s rebuilt Gusto Pro, announced October 7 and in early access, adds connectors (the links that let an AI read another system) to Gmail and Outlook “for retrieving information like hours worked, reimbursable expenses sent by clients via email.” Financial Cents’ new MCP (the Model Context Protocol, an open standard that lets an AI read from and act in another piece of software) reads client chat on its Scale plan. In the demo I saw in July, Combinely’s Coco lived in the firm’s Gmail or Outlook and drafted replies from what the firm already knew.
Connect the email. Limit what the agent can do, not what it can read
Here’s the objection. Email is where the firm’s risk lives: on Monday I covered Barracuda’s finding of phishing written for the AI assistant that summarizes your inbox, and told you to ask every vendor whether outside content is treated as data only. Most accountants I talk to won’t connect their email to an AI at all, and the fear is nearly always the same one: it’ll send something.
Monday’s warning was about what an agent can be made to do with your inbox. It was never an argument against letting one read, and here is where the two fit. The gain is the entire archive of what your clients have told you, and the risk is a permission you can withhold. Claude’s Gmail connector, for instance, can search and read, draft, and label, but it sends, replies, and forwards only with your explicit approval by default, and on Team and Enterprise plans the owner decides whether anyone can switch that approval off. Gusto describes its connectors as retrieving information, with every change coming back to the accountant to approve.
Read-only, or read plus draft plus tag with no send, is the safe design, and it costs you almost none of the gain. A hidden instruction in an invoice email is dangerous in proportion to what the agent is allowed to do, and to what you do with its output unchecked. Read-first means the agent can’t act on a planted instruction. It can still repeat one to you as though the client said it, so what it reads goes through the same review as anything else it produces, and nothing lands in the client’s permanent file until a person has decided it belongs there.
Two conditions before the connection, both yours rather than the vendor’s. Connecting the inbox is a disclosure to a service provider under your professional code, so the vendor’s confidentiality terms and a line in the engagement letter come first. And the agent searches one client’s contributions for one client’s work, never the whole firm’s mail at once. (The default Gmail connector reads the message and only the attachment’s metadata; a more capable MCP server, or one your firm builds, reads the attachment too, so treat that as a configuration choice rather than a ceiling.)
Where the machine stops
None of this makes the machine an accountant. On Mercor’s full APEX-Accounting benchmark, the unsimplified one, the best model scores 56.4%, and the failures are overwhelmingly judgment failures rather than search failures: in one case the model treated a program director’s report as donor authorization and released a restricted $15,000 balance. Digits’ Jeff Seibert put it plainly a week later: “no amount of thinking will tell a model how a client books things when it has never seen that client’s books.” Give the machine a simplified task and a complete file and it wins. Give it the full task and it fails on judgment, and judgment about this client is context only your firm can supply.
The question your firm has to answer
How much of what your clients have told your firm sits somewhere an agent can search, and how much is in someone’s head or in a call nobody wrote up? Three decisions follow. Inventory the places clients contribute. Connect them read-first, and decide who in the firm approves each connection. Then give facts one place to land once a person has decided they matter: the connections are how facts arrive, and the one place is where they stay, which is the context Seibert says no amount of thinking can produce.
The free version of that place is the client context file I wrote up in May. The firm-wide version is one of the nine modules in the Practice Transformation Program, and yes, that’s the one I sell. The next cohort starts Wednesday, November 11: four live three-hour sessions, and your champion leaves with a transformation plan your firm actually executes. US$2,997, and enrollment closes Sunday, November 1, at theaiaccountant.ai/transformation.