An accountant at one of our partner firms spent two months on a client cleanup where the AR aging and the general ledger would not tie. Two weeks ago she got access to Claude for the first time, handed it the same data, and had the analysis back in forty minutes.
Her story is one of four that Asa Laws, Baysora’s director of technology, walked through on a closed webinar we run for owners inside the Baysora partner community. Every example came out of a live client engagement, and in each case the person who built it was an accountant. The firms producing this work are the ones where a preparer or a partner decided to try something on a Tuesday afternoon.
None of the four replaces the work that actually requires a CPA. Multistate exposure, passive versus active determinations, returns with a hundred K-1s: AI does not resolve those. What it resolved was the volume work surrounding them.
A tax organizer that runs in thirty minutes
Clients send messy documents. They arrive unlabeled, scanned badly, three forms deep in a single PDF, and no document management system fixes that on its own. One partner firm decided to treat the mess as a solvable problem.
They connected Claude to their existing file system and SharePoint, then built a reusable skill that sorts incoming client documents into the structure their preparers already use, renames them, and produces a tax organizer from what it finds. The organizer pulls the relevant figures, flags what is missing, and marks the judgment calls. In the example Asa showed, it caught income on the P&L with no corresponding 1099, identified 1099s issued against a personal SSN instead of the business EIN, and surfaced a Section 179 decision and a meals deduction question for a human to resolve.
The firm’s baseline for this work was roughly four hours per client. The skill runs in about thirty minutes, unattended, and one partner reported keeping four to seven of them going at once while he worked on something else. The run Asa demonstrated processed 41 files. Usage cost, on the most capable model available, came to about three dollars.
The preparer still verifies the figures, resolves the flags, and talks to the client.
Two months of reconciliation work, reproduced in forty minutes
The AR cleanup had consumed two months of that accountant’s time. She had gone line by line through the ledger and email by email with the client, with no clear view of where the aging and the GL diverged.
She got access to Claude two weeks before the session and had done one training with Asa. She pulled the GL and AR extracts, 24,000 rows in total, explained what she was trying to accomplish, and let it run. Forty minutes later she had an Excel report of the findings, an auditable record of how it reached them, and a consolidated list of questions for the client.
She used it to check work she had already finished, which is the right way for anyone to start. She told Asa she wished she had had it eight weeks earlier. She and her team are now building the process into a skill for the entire bookkeeping group.
Any firm waiting for the technology to settle down before letting anyone touch it is giving up results available this quarter.
A $250,000 IRS assessment reduced to zero
A prospective client came to one of our partners with a payroll tax assessment from the IRS for roughly $250,000. Their previous accountant had already conceded the point and told them the agency was right.
The partner read it and disagreed on instinct. He pulled four years of general ledger data, added the IRS correspondence for context, handed the whole set to Claude, and asked it to work through the discrepancy. It ran about forty minutes in the background while he did other work, and it came back having identified hundreds of miscategorized entries. The categorization errors were the reason the IRS believed tax was owed.
He audited the findings himself and agreed with the analysis. He then had Claude draft the defense memo and build an auditable evidence list in Excel tied back to the source data. He signed it, sent it, and the client ended up owing nothing.
Total time was about an hour, forty minutes of which was unattended. He billed $2,500, roughly one percent of the assessment, and could likely have billed more. The client was glad to pay it.
The instinct that the IRS position was wrong was his. What he got was the ability to test it against four years of ledger data before the end of the afternoon.
Onboarding that finds what is missing before the client does
Onboarding a large new client means assembling prior returns, meeting notes, email threads, and whatever files have already come across. A traditional checklist tells you what a complete file looks like. It cannot tell you what you already have.
One partner firm built a skill that reads everything they have collected, determines what a complete engagement requires for that specific client, and identifies what is absent. In one case it noticed a K-1 from an entity mentioned on a discovery call that had never arrived, along with five other missing sources. It then drafted a single email requesting all of it at once.
The firm reports saving two to ten hours per onboarding depending on client complexity, and the client gets one clear request instead of six rounds of back and forth. It took the firm weeks to go from nothing to a skill deployed across the whole organization, where every partner and every preparer can run it.
How firms get from here to there
These four firms are separate businesses, and Asa’s work with each followed the same sequence.
He starts with leadership. He meets with the partner group before anyone else and gets them logged in and using it themselves. Reading about capability does not move anyone. Watching your own client problem get solved does. Fear turns into enthusiasm at the moment a partner sees their own work come back better than they expected.
Champions come next. Every firm has people willing to dive in, make mistakes, and test workflows that might not survive. Find them, train them deeply, and let them carry it across the business. They are the ones who will notice the fifth use case, and the twelfth.
Then the limits. Trust in a tool requires knowing where it fails. Clients with dozens or hundreds of K-1s, multistate and international exposure, passive versus active determinations, the accumulated judgment that makes a return actually save someone money: AI does not resolve those. Firms that pretend otherwise burn credibility with their own staff on the first bad output. Concentrate the effort where the return is real and be honest about the rest.
This sequence is the kind of work Baysora does alongside our partner firms. Any firm can run it with the right person leading.
What this asks of a firm
Adopting AI changes how a firm decides what work is worth doing by hand, and the answer keeps moving. Software migrations have a completion date. Moving from Lacerte to UltraTax or rolling out a new CRM is a training problem that ends. This one runs continuously, because the models available in six months will be meaningfully more capable than the ones these four examples were built on.
So the firms getting this right are building a culture that looks for use cases, iterates, and lets go of processes that have outlived their reason for existing. The tool matters less than that. Asa’s own recommendation is to start with a general-purpose tool like Claude Cowork and learn how AI works before buying a specialized product, because the options will look different by the time you have formed an opinion.
The decision belongs to the partner group. Capacity is coming back to these firms in real amounts: three and a half hours per organizer, two to ten hours per onboarding, weeks off a single reconciliation. Some of them will put it toward more clients on the same headcount. Some will finally launch the advisory service they have been describing to each other for years. One partner told Asa he wants his senior people home at a reasonable hour in March, because he has watched two of them leave for firms that promised exactly that.
The capacity arrives either way. What does your firm do with it?
