Skip to main content
Guide9 min read·Updated July 24, 2026
🧩

Best AI Agent Skills for Tax Preparation (2026)

B

A. Frans

Published July 24, 2026

Tax PreparationAccountingAI Agent SkillsBookkeepingFinance

The math of a small tax practice looks fine until February. Then a two-partner firm in Ohio finds itself staring at 340 clients, each dropping off a folder: W-2s, a stack of 1099-NECs, brokerage 1099-B pages that run twelve deep, and the classic shoebox of receipts held together by a rubber band. Somebody has to key every figure into the workpapers. Somebody has to tie the 1099 totals back to the client's own spreadsheet. By mid-March the whole office is running 60-hour weeks, and the bottleneck is never judgment. It's data entry.

That is the part AI agent skills are good at right now. Not filing. Not deciding whether a home office qualifies. The tedious middle: reading documents, pulling numbers, and lining them up so a human can check them fast.

A quick distinction, because it matters. An AI tool is a finished product you log into, like a tax SaaS or a chatbot. An AI agent skill is different. It's an installable capability package, a folder of instructions plus scripts, that an agent such as Claude or Claude Code loads on demand to run a repeatable job over your files. You point it at a folder of PDFs on your own machine and it does the job you defined. Nobody hosts your data for you, and nothing gets filed with the IRS.

Here are the ones worth wiring up before busy season, and what each honestly is.

Skill / ToolkitWhat it does for tax workSource / installWatch-out
pdf-toolsReads 1099s, W-2s, and scanned receipts; pulls text and figures out of PDFs, splits and merges filesAnthropic skill (pdf in the anthropic-skills set)Real skill. OCR on bad scans still misreads digits; verify every number
xlsx-spreadsheetBuilds and edits the workpapers, reconciles totals, formats the summary tabsAnthropic skill (xlsx)Real skill. It writes formulas you must sanity-check, not just values
stripe-agent-toolkitPulls transaction and payout data for clients who run on StripeStripe's agent toolkit (MCP / SDK, not a "skill")Toolkit, not a skill. API keys involved; use restricted read-only keys
financial-services-pluginsConnects an agent to banking / payments data sources for clients on those railsPlugin / MCP collectionServer toolkit, varies by provider; audit scope and permissions
skill-creatorLets you build your own firm-specific skill (your workpaper format, your naming rules)Anthropic skill (skill-creator)Real skill. What you build is only as careful as the instructions you write
Two of those, pdf-tools and xlsx-spreadsheet, are genuine skills in the sense the directory means: folders an agent loads. The Stripe and financial-services entries are MCP toolkits, servers your agent talks to, and I'm calling that out because people conflate them constantly. Both categories help. They help at different layers.

Document intake and extraction with pdf-tools

This is where the hours go, so it's where you get them back. A 1099-NEC has maybe eight fields that matter. A brokerage 1099 consolidated statement has hundreds of transaction rows across pages. Typing those by hand is how tired people transpose 5,400 into 4,500 at 9 p.m.

The pdf skill reads the document and pulls the figures into a structured form you can drop into a workpaper. On a clean, machine-generated 1099 the extraction is reliable. On a phone photo of a crumpled receipt, it's a coin toss on the cents, sometimes the dollars. So the workflow I settled on: batch the whole client folder, let the skill extract everything into a draft table, then a preparer reads down the "source figure" column against the actual PDF thumbnails. Reviewing extracted numbers goes maybe four times faster than keying them from scratch, and the reviewer stays alert because they're checking, not transcribing.

One habit that saved us: keep the page reference next to every extracted number. When the reviewer sees "Box 1: 18,240 (1099-NEC, Acme LLC, p.1)" they can jump straight to the source. An extracted number with no traceable origin is worse than no number at all.

Organizing and reconciling in spreadsheets with xlsx-spreadsheet

Once the figures are out of the PDFs, they need a home and they need to tie. The xlsx skill builds the workpaper, drops the extracted values into the right tabs, and runs the reconciliation you describe: total the client's reported 1099 income, compare it to the sum of the actual 1099s you extracted, flag the delta. When a client swears they made 92,000 and the 1099s add to 96,800, the skill surfaces that 4,800 gap in seconds instead of at the end of a long night.

It also handles the grunt formatting. Standardize column headers across 300 client files, split a merged "name, amount" column, recompute a schedule after one number changes. I trust it to write the formulas. I do not trust it blindly, and neither should you, because an agent can happily write =SUM(B2:B19) when the data runs to B20 and quietly drop a row. Check the ranges. Spot-check the totals against a calculator on a couple of files per batch.

Pulling transaction data for clients on Stripe and similar rails

Plenty of small-business clients run everything through Stripe or a similar processor. Instead of asking them to export a CSV they'll format wrong, the stripe-agent-toolkit lets an agent pull payout and transaction data directly. For a client on that rail it turns "email me your 2025 numbers" into a query.

Say it plainly: this is an MCP toolkit, not a folder-of-instructions skill, and it touches live financial data through API keys. Use restricted, read-only keys scoped to the one account. Never the client's full secret key. The financial-services-plugins collection covers the same idea for banking and payments connections, and the same caution applies harder: know exactly what each connector can see before you turn it on.

Building a firm-specific skill with skill-creator

Off-the-shelf skills don't know your workpaper template, your file-naming convention, or that you always split brokerage 1099s by account. The skill-creator skill lets you build one that does. You describe your process once, it packages the instructions and helper scripts into a reusable skill, and next season your agent already knows your firm's way of doing things.

We built a small one that takes a client folder, runs pdf extraction, drops results into our standard workpaper via xlsx, and produces a "review me" sheet listing every number with its source page and every reconciliation gap. It's not clever. It's just our checklist, encoded, so a first-year staffer and the agent follow the same steps. The skill is only as good as the instructions you feed it, so write them the way you'd train a careful new hire.

For anyone wiring this up in Claude Code, installing a skill from a repo looks roughly like this:

# Add a skills marketplace, then install the pdf skill
claude plugin marketplace add anthropics/skills
claude plugin install pdf@anthropic-skills

# Skills live as folders under your skills directory:
#   ~/.claude/skills/<skill-name>/SKILL.md   (instructions the agent loads)
#   ~/.claude/skills/<skill-name>/scripts/    (code it can execute)

Before you run any of that: read the scripts first. A skill is not passive text. It can execute code on your machine, and anything touching client financial records deserves paranoia. Open the SKILL.md, open every script in the folder, confirm it only does what it claims, and check where it sends data (it should be nowhere off your machine). Treat a skill from an unknown author the way you'd treat a macro-enabled workbook from a stranger. The convenience is not worth handing a script your clients' tax files without reading it.

Accuracy and compliance: the part you cannot skip

An AI agent will occasionally hand you a number that is simply wrong. It might misread a 3 as an 8 on a fuzzy scan, or, worse, "fill in" a plausible figure that never existed on the document. That second failure is the dangerous one, because a hallucinated number looks exactly as confident as a correct one.

So the rule in our office is absolute: every AI-extracted figure gets verified against the source document by a human before it goes anywhere near a return. The agent proposes; the preparer confirms. The productivity win is real, but it comes from faster review, not from skipping review. This is also why I keep the source page reference on every number, so verification takes seconds per figure.

And the hard line: these skills do not file returns. They don't sign anything, they don't transmit to the IRS, they don't practice. A licensed human makes the calls and puts their PTIN on the 1040. If a vendor tells you their AI "files your taxes," read that as marketing, not architecture. What these skills do is prep the data that a professional then reviews, decides on, and files through their actual tax software.

Client data handling is non-negotiable

Do not paste a client's SSN, or a full W-2, into a consumer chatbot. That's the whole rule, and it's the one that gets firms in trouble. Consumer AI chat products may retain and train on what you type. A Social Security number in a chat window is a data-handling problem you created for yourself.

Skills that run locally, or inside an environment you control, are the point here. When pdf-tools reads a 1099 on your own machine, the file doesn't leave. Keep it that way: prefer local execution, understand the data policy of any API-based toolkit before it touches client records, and scope every key and connector to the minimum it needs. Your engagement letters and your state board both assume you control where taxpayer data goes. An agent skill doesn't change that obligation. It just gives you one more surface to get it right or wrong.

If you're evaluating the broader set of options, we keep running lists of AI tools for accountants and AI tools for finance professionals that pair well with the skills above.

FAQ

Can AI do my clients' taxes? No. AI agent skills extract and organize data, reconcile figures, and build workpapers. They don't determine tax treatment, exercise professional judgment, or file returns. A licensed preparer reviews the prepped data and files it through their own tax software. Treat the AI as a very fast, occasionally wrong, data-entry assistant.

Is it safe to put client tax data into AI? It depends entirely on where the AI runs. Skills that process files locally on your machine keep the data with you. Consumer chatbots that may retain or train on your input are not safe for SSNs, full W-2s, or anything identifying. Read the data policy, prefer local execution, and scope every API key to read-only and to one account.

Will AI replace accountants? Not the accountant. It replaces some of the keystroke work, the transcribing and reconciling and reformatting that eats February. The judgment, the client relationship, the signature, and the liability stay with the human. Firms that hand the tedium to skills spend their saved hours on review and advisory, which is where the fee actually is.

Do these skills integrate with tax software like Drake or Lacerte? Not directly. There's no native connector from pdf-tools or xlsx-spreadsheet into Drake, Lacerte, or UltraTax. These skills work one layer earlier, turning raw documents into clean, verified data you then enter into your professional software the way you already do. Think of them as improving what goes into the return, not filing the return itself.

The honest read after a season of running these: AI skills earn their keep at the data-prep layer and nowhere else yet. Extraction, reconciliation, organizing hundreds of messy folders into clean workpapers, that's real time back, and it's the time you were losing anyway. Review stays human because the tools still hand you wrong numbers now and then. Filing stays human because that's the law and your license. And client data handling isn't a setting you tune later, it's the first decision you make. Get those three boundaries right and a skill-equipped agent is the best February hire your firm never had to pay overtime.

Share this article

📬

Get More AI Tool Guides

New comparisons and guides every week. Join thousands of professionals staying ahead of the AI curve.