On October 5, 2026, Wolters Kluwer announced an agentic layer for TaxWise. According to CPA Practice Advisor's October 7 report, it organizes client documents, extracts and maps the tax data, flags possible issues and produces a draft return ready for review. Wolters Kluwer's Rocco Impreveduto called it "one of the most significant" advances in preparation. Nine days later, on October 14, CPA Practice Advisor published its 2026 Innovation Awards. One finalist, Kaseload, does nothing at the preparer's desk. It sits in front of the client, asks for what is missing and keeps asking until the checklist is done.
Those two announcements sit at opposite ends of the same file, and I think most firm owners are paying attention to the wrong one. This essay is for owners of CPA and EA practices who want to know how a tax practice can have an AI agent get client organizers completed, following up on the unanswered questions until the file is ready for a preparer. My position is simple. The return was never what slowed your practice down. The client going quiet did. An agent that drafts a return in four minutes is worth very little if the file it needs shows up in April with three blank answers and a 1099 from the wrong year.
None of what follows is tax advice. The preparer decides everything. That sentence is a design requirement, and I will come back to it.
What shipped while you were filing extensions
Look at where the money and the engineering went this year. TaxGPT released an autonomous tax prep agent in March that works through browser automation inside the software firms already run, per CPA Practice Advisor. Byron launched on May 27, 2026 with agents for 1065, 1120 and 1120-S returns and a $6.5 million seed round. Karbon put Kai, which it calls an AI coworker, into early access on June 3. And now Wolters Kluwer, one of the two or three companies whose software already sits under a large share of American returns, has made the draft return an agent's job.
The adoption numbers say firms are ready for this. A Blue J and CPA.com survey of more than 1,000 US tax professionals, released in June, found 60% use AI for tax research at least weekly, up from 33% in 2025. When asked where they would put the time AI saves, 50% said "improving client response and delivery timelines." That answer is interesting because research tools do not touch delivery timelines much. Delivery timelines are set by the slowest input, and the slowest input in a tax practice is a person who got your organizer in January and opened it in March.
Almost every product named above starts working once the documents are in the building. They are preparer-side tools. They make the fast part of the job faster.
The preparer was never the constraint
Every firm owner knows the shape of the season. The organizers go out. Some come back fast and complete, from the clients who were always going to be easy. Most come back partial. A few come back never, and turn into extensions. Meanwhile your staff spend the weeks from February to April writing the same email in forty slightly different tones: we still need your K-1, did you sell the rental, is the brokerage statement the final one or the preliminary one.
That work has a strange property. It is low-skill in the sense that the question is already known. It is high-cost because the person who knows which question to ask is usually a senior preparer or a manager, since a junior staffer cannot look at a file and see that something is absent. The firm spends its most expensive judgment on noticing a gap, then its scarcest attention on chasing the gap closed.
Kaseload's own nomination text claims that a firm that "previously spent forty to sixty minutes per return on collection and follow-up" gets that time back. That is a vendor's number, offered to an awards program, and I would not underwrite a budget with it. The shape of the claim still matters. It says the chase is a per-return cost that runs into the same order of magnitude as preparation itself on a simple 1040, and that it scales linearly with your client count, which is the one thing a practice owner cannot afford in a labor market that will not give you more staff.
Here is the structural point. When you speed up the preparer and leave the chase alone, you do not get a faster firm. You get preparers idling in February and drowning in April, because the files arrive on the client's schedule. The capacity you bought in October sits unused until the inputs show up. Throughput in a tax practice equals the rate at which files become ready, and a draft-return agent has no effect on that rate.
How a tax practice can have an AI agent get client organizers completed
Say you want to build this properly instead of buying a reminder tool. The architecture has three parts, and each one exists to keep a line clear: the agent handles completeness and the preparer handles correctness.
The question bank belongs to the preparer
The agent should never invent a question. Every question it can ask comes from a bank your preparers write and own, keyed to facts in the prior-year return and the current organizer. If last year's return had a Schedule E, the agent asks about the property. If last year's return had a 1099-B from a particular broker, it asks for this year's. If the client answered "yes" to a new dependent, the agent asks for the name, date of birth and Social Security number, because the preparer wrote a rule saying those three items close that question.
Kaseload describes a version of this in its award entry: it builds a document checklist from the client's prior-year return or the firm's intake form, and checks each upload against the item it was meant to satisfy, so a wrong-year statement gets caught and the client is told what to send instead. Unclear items go to the firm. That last clause is the whole design. When the agent cannot tell, it escalates instead of guessing.
The prior-year return is the richest source of questions you own, and most firms treat it as a reference document rather than a set of triggers. Turn it into triggers. Each line that existed last year raises a question for this year: still true, changed, or gone.
The agent may ask, never advise
This is the boundary your engagement letter, your errors and omissions carrier and your conscience all care about. A client will reply to a follow-up with a question of their own. Some of those questions are logistical: where do I upload this, which statement do you mean, can I send a photo. The agent should answer those directly and well. Some are tax questions in disguise: does my Uber driving count as a business, can I deduct my car, do I qualify for the overtime thing my coworker mentioned.
The agent's job on the second kind is to collect the facts the preparer will need and route the question, with no hint of an answer. "Your preparer will decide how this is treated. To help them, can you tell me how many months you drove and whether you have mileage records?" That reply is useful to the client and safe for the firm, and it moves the file forward. A reply that says "yes, you can probably deduct that" is a liability with a friendly tone.
Write the refusal as a rule, then test it. Before the season, have someone on your team spend an afternoon trying to get the agent to give advice. Ask it leading questions. Ask it in Spanish. Ask it as a frustrated client who says the last accountant always just told them. Every failure you find in December is one you do not find in a client's inbox in March.
Ready is a definition, written down
The agent cannot know when to stop unless you tell it what "ready" means. Most firms have never written this down. Ready lives in a manager's head as a feeling: I could start this one. The agent needs it as a list. Every triggered question has an answer or a documented "not applicable." Every expected document has arrived, matches its year and has passed the upload check. Every client-raised tax question sits in the preparer queue with its supporting facts attached.
Once ready is written down, the agent has a stopping condition and your workflow has a gate. Files that pass the gate go to a preparer. Files that do not stay with the agent, which keeps following up on a schedule you set, in a voice you approved, until they pass or until a human decides to call the client. The preparer's day stops starting with a chase and starts with a file that is actually workable.
The phrase "until the file is ready for a preparer" carries the entire product. Organizer completion tools have existed for years. The new thing is persistence tied to a definition, with someone who never gets tired or embarrassed doing the asking.
The new deductions made the organizer longer
There is a timely reason to build this now rather than next year. The One, Big, Beautiful Bill created a set of deductions that run on paperwork clients do not usually keep, and those questions now belong in every individual organizer through tax year 2028.
The IRS spelled out the records in fact sheet FS-2026-05, published in March 2026. For the tips deduction, clients need "Tip records for 2025, such as daily or weekly tip logs, employer reports, point-of-sale summaries, invoices," with separate records for each employer or client. For overtime, they need "Pay stubs or payroll summaries showing overtime pay." For car loan interest, the IRS wants the lender statement plus "Vehicle identification number and purchase details." The enhanced senior deduction needs date of birth and income documents that bear on eligibility.
Look at what that does to the chase. These items do not arrive on a single form the client receives in the mail in late January. They live in a restaurant's point-of-sale system, in a payroll portal the client has forgotten the password to, in a dealership folder in a kitchen drawer. Each new deduction adds a question whose answer requires the client to go and find something, which is exactly the kind of question that gets skipped and then forgotten.
These are also the questions most likely to bring the "do I qualify" reply. A client who reads a headline about no tax on tips will ask your agent whether their job counts. The agent should gather the facts (the occupation, the employer, whether tips were reported) and hand the eligibility call to the preparer. If you are going to test the advice boundary anywhere, test it here.
Where the design breaks
I have watched firms try versions of this with plain email automation, and the failures repeat.
The first failure is tone. A reminder that reads like a collections notice gets ignored or, worse, gets a client to call the partner and complain. A reminder that reads like a person who remembers what the client said last time gets answered. The agent should know that the client already said the rental was sold in June and should never ask about rental income again. Generic reminder tools do not remember. Agents can, if you give them the thread.
The second failure is the channel. Some clients answer email. Some only answer texts. Some will only talk to a voice on the phone. Kaseload pitches a branded portal that needs no account, and that detail matters more than it seems, because every login wall you put in front of a client loses a share of them. Pick the channel the client already uses and let the agent meet them there.
The third failure is escalation that never happens. An agent that follows up forever is as broken as one that never follows up. Set a ceiling. After a set number of unanswered touches, or once a deadline is close, the file goes to a named human who picks up the phone. The agent's work makes that call shorter, because the human sees exactly which three items are open and what the client said about each.
The fourth failure is data handling. Your client's Social Security number, their W-2 and their brokerage statements are moving through this system. Kaseload's entry lists a full audit trail, tenant isolation and an on-premises option, and those are the right questions to ask any vendor. You need to know where the data sits, who at the vendor can see it, and whether any of it trains a model you do not control. Your Written Information Security Plan should name the tool and describe the flow. If it does not, you have a gap your insurer will find before you do.
The arithmetic of the chase
Suppose a practice files 1,200 individual returns, and suppose each one costs 40 minutes of staff time in collection and follow-up, the low end of Kaseload's claim. That is 800 hours a season, most of it in a ten-week window, much of it done by people whose time you would rather spend reviewing returns. Suppose an agent removes three-quarters of it. You get 600 hours back, which is roughly the season of one additional full-time preparer, without a hire you probably cannot find.
That illustration is not a forecast for your firm. Your client mix, your current process and your staff all change the numbers. The direction holds, though. The chase cost scales with clients, and the agent's cost per client is small and flat, which is why Kaseload prices per client with no platform subscription.
There is a second gain that does not show up in hours. When files reach "ready" earlier and more evenly, the preparers' work spreads across February instead of piling up in the first two weeks of April. Fewer files become extensions because the client simply never finished. Remitian, which won a 2026 Innovation Award for tax payments, reported that one firm ran 194 payments totaling $2.65 million through it and reclaimed an estimated 275 senior manager hours. The lesson I take from that figure is that the hours firms get back from AI tend to come from the administrative edges of the return, the payments and the documents and the chase, not from the technical center.
The preparer-side agent needs this to exist
Return to Wolters Kluwer for a moment. Its agentic TaxWise ingests client documents and produces a draft return, and the CPA Practice Advisor report says "preparers remain in control throughout." Every preparer-side agent shares one dependency. It can only draft from what it was given. Feed it an incomplete file and it will either stop and flag, which pushes the work back to your staff, or it will draft around the gap, which pushes the risk onto your reviewer.
So the follow-up agent is the missing half of the preparer-side agent, and the order of investment matters. A firm that buys the drafting agent first has made its preparers faster at waiting. A firm that builds the completion layer first has made every downstream tool, human or software, work on better files. When the drafting agents get cheaper next year, and they will, given how many companies are now building them, the firm with clean inputs gets the full benefit on day one.
Karbon's 2026 report, cited in its awards entry, says "81%" of clients feel their accountant becomes more valuable when AI handles routine work. I would read that cautiously, since it comes from a vendor. The instinct behind it still rings true to me. Clients do not experience your preparation speed. They experience your emails. The chase is the part of your firm they see.
What to build before the organizers go out
Organizers for the 2026 tax year go out in the next eight to ten weeks. That is enough time to do this properly if you start now, and not enough if you wait until January.
Start by writing the question bank from last year's returns, with your best preparer deciding which prior-year facts trigger which questions, including the new Schedule 1-A items. Then write the definition of ready, item by item, for your three or four most common return types. Then write the advice boundary as explicit rules, along with the handoff language the agent uses when a client asks a tax question. Then pick the tool, or build on the platform you already run, and test it against your own hardest clients before a single real reminder goes out. Choose the escalation ceiling and the human who owns each stalled file.
Notice what most of that work is. It is writing down knowledge that currently lives in your senior people's heads. The software is the easy part. The question bank, the stopping rule and the boundary are your firm's judgment turned into a form a machine can follow, and that is the part a vendor cannot sell you.
Architect the chase, or keep paying for it
Buying a reminder tool will not fix this. A tool without your question bank asks generic questions. A tool without your definition of ready never knows when to stop. A tool without a written advice boundary will, sooner or later, tell a client something your preparer would not have said, in writing, under your firm's name. Each of those failures comes from a missing piece of architecture, and none of them is a missing feature.
The firms that come out of the 2027 season ahead will be the ones that treated client silence as a systems problem and designed against it: the questions owned by preparers, the stopping condition written down, the boundary tested, the escalation assigned to a person. That work crosses your workflow, your engagement terms, your security plan and your staff's daily habits, and it has to be done before the first organizer leaves the building.
Agor AI Advisory does this work with tax practices: mapping the chase, writing the rules, choosing or building the agent, and testing it until the advice line holds. The preparer keeps every decision. The agent keeps asking.
Sources
- CPA Practice Advisor, "Wolters Kluwer Adds Agentic Tax Preparation to TaxWise," October 7, 2026
- CPA Practice Advisor, "2026 Innovation Awards Announced: AI Tech Drives Advancements for Accounting Firms," October 14, 2026
- CPA.com, "Blue J and CPA.com Survey Finds AI Adoption Among Tax Firms Has Nearly Doubled in One Year," June 2026
- IRS, "What you will need to file your taxes under the One, Big, Beautiful Bill," FS-2026-05, March 2026
- CPA Practice Advisor, "Byron Launches with AI Agents for Business Tax Preparation Workflows," May 27, 2026
- CPA Practice Advisor, "TaxGPT Releases Autonomous Tax Workflow Agent," March 24, 2026
- CPA Practice Advisor, "Karbon Introduces AI Coworker Kai," June 3, 2026