The AI Tax Automation Market Is Moving Fast — and That Should Make You Cautious
The pitch sounds compelling every time: AI reads your clients' tax documents, extracts the data, and populates your tax software automatically — no manual keying, no data-entry errors, no staff burnout in January. CPA firms are right to be interested. The manual processing of W-2s, 1099-DIVs, 1099-INTs, and the rest of the 1099 family is genuinely expensive, genuinely error-prone, and genuinely miserable for the people doing it.
But the AI vendor landscape right now is a minefield. Every category attracts opportunists during a hype cycle, and AI is no different. Some vendors are selling polished demos of software that doesn't work in production. Others are making claims about accuracy that fall apart the moment a real document hits the system. A few are quietly using your clients' sensitive financial data to train their models — something buried in a terms-of-service clause that nobody reads until something goes wrong.
Before your firm signs a contract, writes a check, or hands over a single client document, there are specific questions you should be asking — and specific answers that should give you confidence or send you running. Here is the list that matters.
1. Does It Actually Work With the Tax Software My Team Already Uses?
This is the question most firms forget to ask with enough specificity. "Does your product integrate with ProSeries?" is not the same question as "Does your product type data directly into ProSeries fields, in the correct locations, the way a trained staff member would?" The difference matters enormously in practice.
Many AI tools in this space stop at extraction — they read the document and give you a structured data file or a formatted spreadsheet. That is useful, but it still leaves your staff doing the transfer work, just from a different source. The data-entry burden isn't eliminated; it's reorganized. True automation means the system completes the entry end-to-end, in the software your team is already using, without requiring anyone to rekey values.
Ask vendors to be precise: Does the data go into the tax software, or does it go into a middle layer that your staff then works from? If it goes into the software, how? Which fields? Which forms? What happens with edge cases — a W-2 with a Box 12 code your staff rarely sees, or a 1099-DIV with foreign tax paid?
Also ask about the software roadmap. If a vendor currently supports only one tax platform, what is their timeline for expanding? This matters for firms evaluating switching costs down the road. For context, Drake support is an example of functionality that may be in active development at a given vendor but not yet available — and you want to know exactly where a vendor draws the line between "coming soon" and "works today."
2. What Happens When the System Isn't Sure?
This is the most important question on this list, and it is the one that separates responsible AI design from reckless AI design.
Every AI system encounters documents it cannot read with full confidence. A W-2 scanned at low resolution. A 1099-INT where a handwritten correction was made. A box that's partially obscured by a staple mark or a fold in the paper. What does the system do in that moment?
There are two philosophies. The first: make a best guess and move forward, keeping the workflow clean and the throughput numbers high. The second: stop, flag the field, and require a human to review it before anything gets entered. The first philosophy optimizes for speed. The second optimizes for accuracy. In tax preparation — where a wrong number on a client's return can mean penalties, amended filings, and a damaged relationship — there is only one acceptable philosophy.
Ask every vendor directly: When your system encounters a value it cannot confidently read, what does it do? Push past marketing language. Ask for a specific, technical answer. A system built correctly will have a clear, well-defined flagging and review workflow. A system built to hit demo metrics will hedge, or tell you the confidence threshold is "very high" without defining what happens at the edges.
Verification matters on the output side too. After data is entered, does the system check what it typed against the source document? Or does it assume the entry was correct and move on? An end-to-end verification step — where the system reviews its own work before flagging anything for staff — is a meaningful signal of engineering maturity.
3. Where Does My Clients' Data Go — and What Is It Used For?
Your clients trust your firm with some of the most sensitive financial information in their lives: Social Security numbers, employer identification numbers, dividend income, interest income, wages. When an AI tool processes that data, it has to go somewhere — typically to a cloud API or an AI model inference endpoint. The question is what happens to it after that.
Some AI vendors — particularly those built on top of large general-purpose AI models — have terms of service that allow submitted data to be used to improve or train their models. That is a direct conflict with your obligations to your clients and, depending on your engagement letters and state-specific privacy rules, potentially a legal problem.
Ask vendors three specific questions. First: Do you have a data-processing agreement available, and will you sign it? Second: Are client documents used in any way to train, fine-tune, or improve your AI models? Third: Where, physically or logically, does the data reside during processing, and for how long is it retained?
A vendor who hesitates on any of these questions, or who cannot produce a signed DPA without a lengthy procurement process, is a vendor whose data practices are not built around your clients' interests. That is disqualifying.
4. Is This Running in Production Anywhere — Right Now?
The AI software market is saturated with products that perform impressively in controlled demos and struggle badly under real-world conditions. A polished walkthrough using a clean, perfectly formatted W-2 tells you very little about how a system handles the actual documents your clients send — crumpled, photographed at an angle, generated by payroll platforms with unusual formatting, or covering a tax year with a mid-year employer change.
Ask vendors directly: Do you have paying customers running this software in production today? How many? How long have they been live? What volume of documents have been processed?
Be appropriately skeptical of "pilot programs," "beta partnerships," and "design partners" — these arrangements often mean the software is still being shaped around the customer's feedback, not that it has been validated in real-world conditions. There is a meaningful difference between a firm that agreed to test software for free and a firm that is paying for it because it works.
Production use is the only real proof of function. A vendor who can point to a paying customer actively running the platform is in a categorically different position from one who cannot — regardless of how impressive the demo looks.
5. How Does the Pricing Model Scale — and What Are the Real Costs?
AI tax automation tools are typically priced on volume — per document, per return, per user, or some combination. The headline number often looks attractive. The fully loaded cost, once you account for implementation, training, exception handling, and the staff time required to manage the tool's output, is frequently higher than the vendor's initial figure.
Ask vendors to walk you through cost at three different volume levels: your current volume, 1.5x your current volume, and 2x. A pricing model that scales fairly is a sign of a vendor who wants a long-term relationship. A model with sharp inflection points — where your cost doubles when you cross a threshold — is a vendor optimizing for their margin at your expense.
Also ask what is not included. Is onboarding and training billed separately? Is there a support tier required for firms above a certain size? Are there per-form costs that only apply to certain document types? The 1099 family is large — 1099-DIV, 1099-INT, 1099-R, 1099-B, 1099-NEC, 1099-MISC, and more — and some tools support only a subset of it. If your clients generate document types the tool doesn't handle, you need to know before you commit.
6. What Does the Firm's Workflow Actually Look Like After Deployment?
The final question is the most practical, and it's often the most revealing. Ask the vendor to walk you through a day in the life of a staff member at a firm using their product during the height of tax season. Not a demo — a realistic description of what the workflow looks like, what the staff member's responsibilities are, what the exception queue looks like on a busy day, and how the firm knows everything that needed review was actually reviewed.
This question surfaces assumptions that vendors make about your operations. Some tools assume you have a document management system already in place. Others assume staff will monitor a separate dashboard outside your normal workflow. Some require meaningful change management to implement correctly. All of these are reasonable — but none of them are free, and you need to know about them before you sign.
A vendor who can answer this question concretely, with specificity about the staff experience, has thought seriously about how their product fits into a real firm's operations. A vendor who pivots back to features and slides is telling you something important about how much they understand the work your team actually does.
The Standard Has Been Set — Hold Every Vendor to It
The AI tax automation category is no longer entirely theoretical. As of late 2026, platforms like Kairos are running in production with paying CPA firm customers — processing W-2s and 1099-DIV and 1099-INT forms, entering data directly into Intuit ProSeries, flagging anything the system can't verify with confidence, and operating on the firm's own computer under a data-processing agreement that explicitly prohibits using client documents to train AI models. That is the benchmark.
When a vendor can't tell you whether their software is in production, won't commit to a DPA, or gets vague when you ask what happens when the system isn't sure — measure that against what is already possible. The bar has been set. Your clients' data and your firm's reputation deserve vendors who can clear it.
Kairos, built by Selah Systems, is an AI-powered W2 and 1099 tax automation platform designed specifically for CPA firms. It eliminates the manual processing burden, reduces errors, and scales with your practice — so your team can focus on work that actually moves the firm forward. If you're ready to see what that looks like in practice, download Kairos free and try it on your own returns.