On October 22, 2026, Thomson Reuters released the latest version of CoCounsel Tax, Audit & Accounting. It writes and runs code for depreciation, amortization, required minimum distributions and qualified business income. It builds multi-tab Excel workbooks with live formulas, connects to UltraTax, and grounds its work in Checkpoint, which the company says holds more than 12 million primary source and proprietary analytical documents. Joel Hron, the company's chief technology officer, told CPA Practice Advisor that code execution is "table stakes now; every company in this space has it." He's right. That is also why it doesn't help you much this winter.
The back office of a tax firm now has an engine. The front door still runs on a phone, a shared inbox and a staff member named, in most firms I talk to, the person who knows where everything is. Every January the shoebox shows up. It holds a W-2, half a brokerage statement, a photo of a 1099 taken at an angle on a kitchen counter, and a note saying "same as last year." The CoCounsel launch makes the gap between those two ends of the firm wider, and the only way to close it before filing season is to start building an AI intake agent for your CPA or tax firm by mid-November. This essay covers what to set up during that pre-tax-season build window, what to leave to the preparers, and why the intake agent is the system where your firm's real AI policy ends up living. None of it is tax advice. Most of it is about paperwork, consent and calendars, and that's the reason it matters.
The engine is waiting on the clerk
Every serious tax AI product I've watched launch this year makes the same assumption. It assumes the documents are already there. CoCounsel's pitch, in Thomson Reuters' own words, is to take work "from source documents and client data to a finished product" inside one workflow. The words "source documents" are carrying a lot of weight in that sentence. Someone has to get them. Someone has to know which ones are missing, ask for them, ask again, notice that the K-1 in the upload is from 2024, and route the client who mentions a new rental property to a human before anyone pretends the return is simple.
That someone is your intake process, and in most firms it's the weakest system in the building. It isn't weak because people are careless. It's weak because nobody ever designed it. It grew out of habit: an organizer PDF, a portal invite, a reminder email written in 2019 and copied forward each year since. It works at September volume and falls apart in February.
A preparation agent fed by a broken intake process does its broken work faster. It will produce a confident workpaper from an incomplete file, with citations, and the citations will be correct for the facts it was given. That's the trap. The more capable the back-office tool, the more expensive a missing document becomes, because the error now arrives looking finished.
So the order of operations is the reverse of what the vendor demos suggest. The clerk comes before the engine. The firms that will get real use out of this season's preparation agents are the ones that fix the front door first, and the front door has a deadline.
Why mid-November is the line
The tax calendar doesn't move for your project plan. According to Greenback Tax Services, updated October 23, 2026, the IRS e-file shutdown for the year-end transition "begins at 11:59 a.m. Eastern Time on Saturday, December 26, 2026," and e-file "will not reopen until January 2027." The same page notes that "The IRS has not announced the reopening date." The season opens when the IRS says it opens, and the shoeboxes start arriving the day clients get their first W-2.
Work backward from that. Suppose you want the intake agent running with real clients in the first week of January. Suppose you want three weeks of December to test it against a few dozen friendly clients, the ones who will forgive a clumsy reminder and tell you about it. Suppose you need two weeks before that to write the consent language with counsel, set up the data controls and decide what the agent may say. That puts the start of real configuration around November 15. Anything later and your testing happens on live clients during the busiest weeks of the year, which is the same as not testing.
December is also when your people are least available to think. Year-end planning calls, estimated payment questions, the client who discovers in the last week of the year that they sold a business. Partners have attention in November. In December they have none. An intake agent needs decisions from people with authority, and those decisions have to be made while those people still have time.
There's a softer reason too. Clients form their impression of how your firm works this season during the first request for documents. If that request arrives from a new system in late January, half-tested, the client blames the firm for the system's mistakes. If it arrives in early January, clear and patient and correct, the client starts the season trusting the process. You get one first message per client per year.
What the intake agent does, and only that
The intake agent is a clerk on a short leash. Its job is logistics. The more boring its scope, the more useful it gets. Here is roughly what belongs inside the leash.
It sends the organizer and the document request list for each client, built from what that client sent last year. If the prior-year file had a 1099-DIV from a specific brokerage, the agent asks for it by name. It chases missing documents on a schedule you set, in the client's preferred channel, and stops chasing when the document arrives. It checks uploads for the obvious things: wrong tax year, unreadable image, a page count that doesn't match, a form type that wasn't expected. It answers logistics questions about deadlines for documents, portal access and appointment times, and when to expect a draft. It books appointments. It routes the engagement letter for signature. It captures consent, which gets its own section below because it's the hardest part.
It also flags. When a client writes "we bought a house in Arizona" or "my mother moved in with us" or "I started selling on Etsy," the agent doesn't interpret any of it. It tags the message, sends it to a named human, and tells the client a preparer will follow up. Spotting a life change is clerical work. Deciding what it means for the return belongs to a professional.
The test I'd use for every capability is simple. If a smart, careful receptionist in her first week could do it with a checklist, the agent can do it. If it needs the judgment of someone who has signed a return, it goes to the preparers.
What to leave to the preparers
Some things stay with licensed humans, and the agent should be built so it can't wander into them. That means any statement about what a client owes or will get back, even a range. It means any interpretation of what a document means for the return, such as whether an expense is deductible, whether a form changes filing status, or whether a client qualifies for anything. It means anything that sounds like planning advice, including "you should" sentences about retirement contributions, entity elections or estimated payments. It means the conversation with the client who is upset, confused or in trouble with the IRS.
You enforce this in the configuration, through refusal patterns, routing rules and a short list of phrases the agent may use to hand off. Don't rely on the model's general good manners. Write the handoff sentence yourself, word for word, and have a partner sign off on it. Something like "That's a question for your preparer, and I've sent it to them with your note." The client should never be unsure whether they're talking to the clerk or the CPA.
This limit also protects the preparers' time. A good intake agent sends them fewer interruptions about missing pages and more well-formed questions about real facts. The preparer's day changes from hunting to judging, and judging is the work you're licensed and paid for.
Consent is the first form
Most firms get this part wrong because they treat it as legal boilerplate. It's actually architecture.
On October 12, 2026, Steven de la Fe, a CPA and co-founder of Precision AI OS, laid out the problem in CPA Practice Advisor. He wrote that "Section 7216 requires a taxpayer's written consent before a preparer uses or discloses return information," and he pointed to the unanswered question underneath every AI workflow in the profession: "whether sending return information to an AI tool counts as a disclosure to a third party." The IRS hasn't answered it. His recommendation is "a standalone, signed AI consent, worded with your counsel," and he adds two points that change how an intake system has to be designed. First, "a consent must be knowing and voluntary." Second, "conditioning services on it generally makes it involuntary."
Read those two sentences as an engineer would. They mean your intake system needs two paths from the first message. One path is for clients who consent to AI-assisted handling. The other is for clients who decline, and on that path their documents never touch the model. That's a routing requirement, and it has to exist before the first organizer goes out, because the organizer is where the client's data first arrives.
This is where the mid-November deadline turns from a preference into a hard limit. The consent has to be collected before the agent handles return information. In practice, that means the consent request is one of the first things the intake system does, in the same flow as the engagement letter, and it has to be written, reviewed by counsel and tested before January. A consent clause stuffed into the engagement letter in a hurry is how a firm ends up with a pile of AI-processed files and an argument it would rather not have.
The decline path is also a design gift. Say 15 percent of your clients decline. Those clients get the old process, run by humans, and you learn exactly what that process costs per client compared with the agent path. Most firms have never had a clean comparison like that. The consent requirement gives you one for free.
The alert names your partner
On June 24, 2026, the IRS Office of Professional Responsibility issued Issue Number 2026-19, titled "Introductory Guidelines for Responsible AI Use in Federal Tax Practice." It's short and plain, and every managing partner should read it before configuring anything.
Three lines in it shape an intake agent directly. "Practitioners cannot rely solely on AI; human scrutiny and editing are essential." That's the reason for the leash. "Practitioners must strictly handle all client data using only secure, enterprise-approved AI." That rules out the staff member who pastes a client's brokerage statement into a consumer chatbot to work out which numbers matter, and it means the intake agent must run on a platform with a data agreement you've actually read. The alert also warns of risk "especially when data is uploaded to unsecured or public systems."
The fourth line is the one people skip. Under Circular 230 Section 10.36, individuals with principal authority over a firm's practice "must take reasonable steps to ensure that the firm has adequate procedures in effect." That points at a named person, your managing partner. Lawyers at Kilpatrick Townsend, writing on September 29, 2026, framed the alert as something practitioners need to absorb "as tax professionals prepare for the upcoming filing season," and their advice is direct: "Practitioners should treat AI-generated material as a preliminary draft."
An intake agent is a procedure. It's a written, configured, testable statement of how your firm handles client data at the moment it arrives. Build it carefully and it becomes the firm's evidence that adequate procedures exist. Build it carelessly, or let staff assemble their own versions out of whatever tools they like, and you have the opposite: proof that nobody was in charge.
The fee line nobody wants to read
The alert has one more sentence that will matter more each year. Under Section 10.27, "Cost savings should be passed on openly, with billing practices that reflect the efficiencies gained from the use of GAI."
That sentence is about pricing, and pricing is strategy. De la Fe reports that in his work with agent workflows at a high-volume tax practice, "human time came down to less than 5 minutes per return." Whatever your own numbers turn out to be, the regulator has told you that hourly billing for work a machine did will draw questions. Intake is the safest place to absorb that pressure, because the savings there come from removing chasing and re-asking, work that clients never wanted to pay for anyway. Firms that automate intake first can tell a clean story: we spent the savings on faster turnaround and more preparer attention on your actual situation. Firms that start with preparation automation will have a harder conversation about what the fee covers.
The pre-tax-season build window, week by week
Here's the setup I'd run if I were starting an AI intake agent for a CPA or tax firm today. Treat the timings as illustration. Your calendar will differ, but the order shouldn't.
For illustration, the first two weeks of November go to decisions, not software. The managing partner names an owner for the intake system, one person with authority to say no. That person writes a one-page scope listing what the agent may do, what it must hand off, and the exact handoff sentences. Counsel drafts the standalone AI consent and the language for the decline path. Someone reads the data terms of the platform you plan to use, including retention, training use and where the data sits, and writes down the answers. That last task takes an afternoon, and in my experience it's the one most often skipped.
Suppose the next two weeks, up to roughly the start of December, go to configuration. You load last year's client list and the document history for each client, so the agent knows what to ask for. You build the two paths, consent and decline, and confirm that a decline really keeps a client's documents out of the model. You write the reminder cadence. You connect the agent to your portal and your calendar, and to nothing else yet. The urge to connect it to the tax software in week three is strong. Resist it. Intake that writes into preparation systems is a second project, for a quieter month.
Suppose December goes to rehearsal with real clients. Pick a few dozen who are organized, friendly and honest. Send them the real organizer through the real agent. Read every conversation. Count three things: how many documents arrived complete on the first request, how many messages the agent sent to a human, and how many of those handoffs were unnecessary. The second and third numbers tell you whether your leash is the right length. If the agent hands off everything, it's an expensive autoresponder. If it hands off almost nothing, look harder, because it's probably answering questions it shouldn't.
January is for going live and watching closely. Keep a human reading a sample of agent conversations every day through the first month of the season. That person will be tired and will want to stop. Don't let them.
What the intake agent should never be asked to do
Most firms will be tempted to make the intake agent "smart." They'll want it to pre-classify documents into return lines, estimate complexity, suggest the right preparer, or flag which clients look like planning opportunities. Some of that is fine later. None of it belongs in the first season.
The reason is accountability, not capability. The models can attempt all of it. The question is who owns the outcome when the classification is wrong. In the first season you want every judgment in the chain to belong to a person whose name is on a return. An intake agent that stays clerical keeps that chain clean. An intake agent that starts interpreting makes a quiet decision on the firm's behalf, and under Section 10.36 the firm owns that decision whether anyone saw it or not.
Firms that keep the agent boring in season one earn the right to make it clever in season two, because by then they'll have a year of logged conversations that show exactly where the leash could safely be let out. That log is worth more than any feature a vendor can sell you this fall.
Why the front door decides the season
Back to the October 22 launch. Thomson Reuters built a system that stays anchored to authoritative sources and shows its assumptions and calculations for review. That's real progress, and the firms that adopt it will prepare complex returns faster than they could a year ago. Hron told the trade press that code execution is table stakes. The same logic applies one layer up. Once every vendor can produce a fine workpaper, the workpaper stops being where firms compete. They compete on what goes into it.
Intake is where a firm's quality gets decided upstream of any model. It sets whether the file is complete, whether consent exists, whether the life change got flagged, and whether the client trusted the process enough to send the awkward document instead of hiding it. None of those things can be bought from a tax research vendor, because none of them live in the tax research. They live in your client list, your partners' judgment about what a clerk may say, and your counsel's wording of a consent form. That makes them yours, which is the only kind of advantage that survives a market where everyone licenses the same engine.
There's also a staffing argument the profession is slow to admit. The person who knows where everything is will leave someday. When she does, most of the firm's intake knowledge leaves with her, because it was never written down. Writing it down as agent configuration (document lists per client, reminder schedules, handoff rules, consent paths) is the first time many firms will have their intake process in a form that outlasts a resignation letter. The agent is the clerk, and the written configuration is also the firm's memory of how it treats people at the door.
Build the door before the crowd arrives
The pressure this fall is to buy. Vendors launched aggressively in October, and the demos are impressive. Buying a preparation agent before you've fixed intake is like hiring a brilliant surgeon for a hospital with no admissions desk. The surgeon is fine. The patients show up in the wrong rooms with half their charts.
What a CPA firm needs before January is a designed front door: a written scope, a consent architecture with a real decline path, a named owner under Section 10.36, a test with real clients in December, and a log that turns this season into next season's evidence. Those are decisions about how your firm handles people, data and liability. A tool can carry them out, but it can't make them for you. The firms that make them in November will spend the season judging returns. The firms that wait will spend it chasing PDFs while their new preparation agent turns incomplete files into finished-looking work.
Agor AI Advisory designs these systems with firms before the season, not during it. We map your intake process as it actually runs, write the scope and handoff rules with your partners, work with your counsel on the consent and decline paths, configure the agent on a platform whose data terms you understand, and run the December rehearsal with you so January holds no surprises. The window is about three weeks wide, and it's closing.
Sources
- Thomson Reuters press release, CoCounsel Tax, Audit & Accounting, October 22, 2026
- CPA Practice Advisor, Thomson Reuters Releases Next Version of CoCounsel, October 22, 2026
- CPA Practice Advisor, What to Settle Before Client Tax Data Touches AI, October 12, 2026
- IRS Office of Professional Responsibility, Issue Number 2026-19, June 24, 2026
- Kilpatrick Townsend, Using AI in Tax Practice, September 29, 2026
- Greenback Tax Services, IRS E-File Shutdown, October 23, 2026