On Nov. 4, 2026, TaxDome shipped something that sounds small. A firm's clients can now connect their TaxDome account to Claude, ChatGPT or any assistant that speaks the Model Context Protocol and ask, in plain language, what is outstanding and whether their return is ready. The press release calls it an industry first. The answer the assistant gives comes from the firm's own records. In practice that means it comes from a card on a board that a preparer dragged from one column to the next, or forgot to drag.
This essay covers how a tax practice can have an AI agent send clients status updates (received, in review, ready to sign, filed) so the phones stay quiet in filing season. The agent reports the status staff set, and it never advises. A build started this week can be live before January. Most firms will treat it as a messaging project. I think it is a truth project, and the hardest part of it is the card nobody moved.
The phone call was always four words long
Strip the pleasantries off almost every inbound call a tax practice takes between February and April and you get the same four words. Where is my return?
Clients do not call because they love talking to their accountant. They call because the queue is invisible to them. They handed over a shoebox of W-2s and 1099s, and from their side of the glass the work went into a dark room. The phone is the only window they have, so they use it, and every call pulls a preparer off a return to read a status off a screen.
HubSync put fresh numbers on this on November 13, 2026, when it released its first State of Client Experience in Accounting report. According to Inside Public Accounting, 53% of respondents ranked improving client experience among their top business priorities, while only 12% rated their firm's current experience as excellent. Half of respondents named chasing documents as the process that eats the most staff time. Only 9% described their firm's AI adoption as leading.
Read those figures together and a shape appears. Firms know the client experience is weak. They know the weakness lives in the back and forth over documents and status. Very few of them have used the tools now available to fix it. The gap between 53 and 12 is mostly made of phone calls about where things stand.
The client brought a robot
The client MCP was the second half of a two-step move. On October 28, 2026, TaxDome, which says it serves more than 15,000 accounting firms, released a public API and an MCP server to every firm on the platform. Firms can connect Claude or ChatGPT to their practice data, ask which jobs are overdue, and let a finished meeting move a job to its next stage. More than 200 firms took part in the beta. Access is set by the firm owner, area by area, read only or read and write, and a connection can never do more than the person who set it up.
Then, a week later, the same door opened from the other side. The client connection is scoped to what the client already sees in the portal. When something needs a signature or a payment, the assistant hands back the exact link. TaxDome made it free to firms and clients alike, and put it live in the Claude directory and the ChatGPT plugin store on launch day.
Ilya Radzinsky, TaxDome's co-founder, was candid about the timing in the release: "Today this is for early adopters." Neil Mukhi, founder of Six10 Ventures, gave the line that I expect every managing partner to hear in some form within two years: "I like their portal, but never going into it is even better."
That sentence is the structural shift. A portal is a place a human chooses to visit. An assistant is a proxy that visits on the human's behalf, as often as the human wonders, at any hour, and turns whatever it finds into a fluent, confident sentence. The firm no longer controls when its status gets read, how often, or in what tone it gets repeated back to the client.
The status field became speech
Every practice management board in the country was built for internal triage. The columns were a shared memory for staff. Everyone knew that "In review" sometimes meant "a reviewer will get to this eventually" and sometimes meant "this has been sitting under a partner's other work for three weeks." The looseness was harmless because the only people reading the board were the people who understood its dialect.
Connect that board to a client's assistant and the dialect disappears. The column name stops being a note to colleagues and becomes a statement of fact made to a customer, by a machine that has no idea the card has not moved since February 3. The assistant will say, warmly and without hesitation, that the return is in review. The client will believe it, because it came from the firm's own system. When the client finally calls, the call will be longer and angrier than the one you were trying to prevent.
Suppose a firm runs 800 individual returns through a season and a fifth of the cards sit in the wrong column at any given moment. Under the old arrangement that error rate cost a little internal confusion. Under the new one it means 160 clients holding a false statement the firm effectively signed.
So the first design decision for a status agent has nothing to do with the model. It is a decision about what each column promises. Clay Welch, EA, of Trailhead Tax, put the goal well in the October release: the point of wiring tools to the data is that "nothing depends on someone remembering." That only works if the data is right. An agent that reports status inherits every lazy habit the board ever had, and broadcasts it.
What the IRS already expects of you
The regulator has been ahead of most firms here. In June 2026 the IRS Office of Professional Responsibility issued introductory guidelines on responsible AI use in federal tax practice. Emily Maxie, chief growth officer at Firm360, wrote in Inside Public Accounting on November 16 that the guidance makes firm leadership responsible, under Circular 230 §10.36, for establishing adequate procedures for AI use, including written policies and documented staff training.
Her piece carries a story every practice owner should read twice. A managing partner told her that an AI tool drafted partner status emails and nearly sent them before anyone reviewed them. His fix was to have the tool save drafts only. Maxie frames this as the difference between "human in the loop" and "human at the helm", a person deciding what gets handed off before the work starts.
I agree with the frame and draw a sharper line from the same story. The near miss came from drafting. A status email composed fresh by a language model, with whatever reassurance or hedging the model chose, is a small piece of advice dressed as an update. Once the model writes the sentences, someone has to read every one of them, and the efficiency evaporates into review. The safer design removes composition from the client-facing path entirely. The agent selects a fixed message keyed to a status a human set, fills in a name and a date, and sends it. Nothing in that message was invented at runtime, so there is nothing to review except the status itself.
Build the agent that only reports
Here is the build I would start this week. It has four states, and each one is a small contract with the client rather than a label on a board.
- Received means every document on the client's checklist has arrived, and someone on staff has confirmed it. It does not mean the first upload landed. If the checklist is incomplete, the client stays in a document request state, and the agent tells them exactly which item is missing, because that list is also a fact staff set.
- In review means a named preparer or reviewer owns the return and has opened it. It carries the owner's first name and the date the state began. If you are unwilling to show the client the date, you have learned something about how your firm uses this column.
- Ready to sign means the return is final, the e-signature request has gone out, and the client's next action is theirs. The agent sends the signing link and nothing else.
- Filed means the return has been transmitted and acceptance has come back. Until acceptance arrives, the agent says transmitted. Filed is the one word clients repeat to their spouses, so it has to be true.
Every message the agent can send for these states gets written once, by a partner, in the firm's own voice, before the season starts. That small library is the whole of the agent's vocabulary toward clients. The model's job is narrow. It watches for state changes, picks the right message, personalises the name and date, and answers the client's inbound questions about status by reading the same fields.
The refusal is part of the product
Clients will ask the status agent other things. Should I amend last year's return? Can I deduct my home office? Did you include the brokerage account? The agent answers none of these. It replies with one fixed sentence saying the question has gone to the client's preparer, logs the question against the client record, and creates a task for a human.
That log turns out to be one of the most valuable things the build produces. Each question a client asks a status agent and gets routed is a piece of advisory demand the firm would otherwise never have seen, because most clients do not phone to ask a question they suspect will be billed. Say a firm sees 300 routed questions in a season and a tenth of them describe a planning need. That is 30 warm conversations about paid work the firm did not have to go looking for.
Staleness is the bug that matters
Each state gets a clock. Suppose "In review" is allowed to sit for 10 business days before the system flags it. The flag goes to the reviewer and the partner, never to the client. The client sees the honest date the state began. The staff see a list of every card that has outstayed its promise. The agent's real gift to the firm is that it makes the board's lies visible internally before they become visible externally.
Most practices that try this discover, in the first week, that their board was a rough sketch of reality. Cards in the wrong column, returns signed but never marked, clients marked received whose W-2 is still missing. Cleaning that up is tedious, and it is the actual work. A firm that skips it and switches the agent on will have automated its own embarrassment.
Six weeks to January
The calendar is not generous. On November 16, 2026, the IRS issued IR-2026-140, the first of its Get Ready reminders, urging taxpayers to begin preparing now for the 2027 filing season. The agency itself is changing shape. ExecutiveGov reported on October 6 that IRS CEO Frank Bisignano told employees an "AI-first IRS" would support staff rather than eliminate jobs, as the agency prepares for a more complex 2027 season. If the IRS is reorganising around software before the season, a 20-person practice has no excuse to walk into February with the same phone tree it had last year.
A workable plan fits between now and the new year. For illustration, here is how I would sequence it across six weeks.
- In the first week, export last season's board and compare every card's final column with what actually happened. Count how many were wrong at the moment they closed. This number is your baseline honesty rate, and you will be shocked by it.
- In the second week, write the four state contracts above in your own words, with an owner and an entry condition for each. Get every preparer to sign off. Disagreements here are cheap now and expensive in March.
- In the third week, write the message library. One message per state, one per missing document type, one refusal sentence. A partner writes them. Nobody lets a model draft them.
- In the fourth week, wire the agent. On TaxDome this now means the public API and MCP server; on other platforms it means whatever read access the vendor exposes. Give the connection read access to status fields and write access only to its own message log and task queue.
- In the fifth week, run it in shadow mode on your own staff and family as clients. Every message it would have sent gets logged and checked against reality.
- In the sixth week, switch it on for a small group of friendly clients who file early. For illustration, start with 25 and widen it in January when the first W-2s arrive.
None of these steps needs a data scientist. Every one of them needs a partner who will spend real hours deciding what the firm's words mean.
Where the quiet goes
The pitch for this build is quieter phones, and that is true as far as it goes. The deeper effect lands on the shape of a preparer's day. Status calls are expensive mostly because of what they interrupt. Say a preparer takes 12 status calls on a heavy day and each one costs four minutes on the phone plus ten minutes finding their place in a complicated return again. That is close to three hours of broken attention, and the four minutes were the cheap part.
When status flows out on its own, two things move. Preparers get long, uninterrupted stretches during the weeks when they matter most. The firm also gets a clean record of every client's position in the queue, with dates, which is the raw material for staffing the next season. Most practices plan capacity on memory and gut. A status agent that runs for one season leaves behind a timestamped history of how long each state really takes, by preparer and by return type.
The quiet also changes what the client pays for. A client who knows exactly where their return stands, without having to ask, experiences the firm as competent before reading a single number on the return. HubSync's 53 and 12 sit in that gap. The firms that close it this season will not need a better brochure.
The partners' objection
Some partners will resist, and their objection deserves a straight answer. Clients pay for a person, they say, and an automated update feels cold.
Clients pay for judgment from a person. Status carries no judgment. "Your return is ready to sign" holds exactly as much warmth whether a partner's assistant types it or a system sends it, and the system sends it the minute it becomes true rather than whenever someone gets around to it. The warmth belongs in the conversations the agent routes to humans, and the firm will have more time for those conversations precisely because it stopped spending hours reading statuses aloud.
The second objection is risk. What if the agent says something wrong? Under the design above, the agent can only say one of a small set of sentences a partner wrote, keyed to a state a human set. The remaining risk sits entirely in whether the state is accurate. That risk existed before the agent. The agent simply stopped letting you ignore it.
The third objection is that only a few clients use AI assistants yet. Radzinsky said as much himself. Clients will catch up faster than firms do, though, and the status agent pays off even for clients who never touch Claude, because it emails and texts them on its own. The client MCP is a second reader of the same truth. Build the truth once and every reader gets it.
Build the truth layer before you buy the voice
The tools are now cheap, and in TaxDome's case the client side is free. That makes the tools the least interesting part of the decision. Any firm can switch on a connector this month. Very few will do the slow work of deciding what each status means, who owns it, how long it may sit, and which sentences the firm is willing to have repeated by a machine to every client who asks. That work is architecture. It decides whether the agent quiets your phones or hands every client a confident, machine-voiced reason to call you angry.
A firm that buys the voice without building the truth beneath it will learn in February that it has industrialised its worst habits. A firm that designs the states, the clocks, the message library and the refusal path now will walk into the 2027 season with a practice that tells clients the truth without being asked, and with preparers who get to finish a thought. The window for that design closes when the first W-2s arrive. Agor AI Advisory builds these systems with practices: we audit the board, write the state contracts with your partners, wire the agent to your platform with the narrowest access that works, and run it in shadow mode until the numbers say it is honest. You keep the judgment. We make sure the status you set is the status your clients hear.
Sources
- TaxDome Releases Client MCP, PR Newswire, Nov. 4, 2026
- TaxDome Releases Public API and MCP Server for AI Agents and Developers, CPA Practice Advisor, Oct. 28, 2026
- HubSync Report Reveals CPA Firm Client Experience Gaps, Inside Public Accounting, Nov. 13, 2026
- What Human in the Loop Gets Wrong About AI in Accounting Firms, Inside Public Accounting, Nov. 16, 2026
- IRS reminder, Take steps now to make tax filing in 2027 easier, IR-2026-140, Nov. 16, 2026
- IRS Expands AI Use as Agency Prepares for 2027 Tax Filing Season, ExecutiveGov, Oct. 6, 2026