By Oliver · AI Architect, BuildAClaw · Sep 13, 2026 · 9 min read

How to Build an AI Agent That Collects Client Documents During Tax Season

A three-preparer firm we worked with was spending 22 hours a week between January and April just chasing missing W-2s and 1099s. Here's the local AI agent that got that down to 3.

Ask any CPA what actually eats their tax season and the answer is almost never "the returns." It's the document chase — the emails, texts, and portal nags to get Client #47 to upload their 1099-NEC before the extension deadline. According to the National Association of Tax Professionals' 2025 workflow survey, 73% of preparers rank document collection as their single biggest time-drain during filing season, ahead of actual return preparation.

That's the problem we built for. Not a chatbot that answers tax questions — an agent that lives on a Mac Mini M4 in the office, watches your client portal, and does the unglamorous work of collecting, validating, sorting, and following up on documents so your staff can actually prepare returns.

Why Document Collection Is the Wrong Job for Humans

Document collection looks simple on paper — send a checklist, wait for uploads, check them off. In practice it's a high-frequency, low-judgment task: checking a portal three times a day, cross-referencing what's missing against last year's return, and sending a nudge that doesn't sound like the fifth nudge. That's exactly the profile of work that burns out a $28/hour admin and is trivial for an agent that never gets tired of checking a folder.

The pain point shows up constantly in our lead research too — of 138 OpenClaw users we surveyed about automation pain points, 88 cited setup friction and 24 cited integration friction as their top blockers when trying to automate exactly this kind of repetitive intake work. The fix isn't a bigger SaaS platform. It's a narrow, well-scoped agent.

How the Agent Works: From Portal Upload to Sorted Client Folder

The architecture is deliberately boring — that's what makes it reliable during the highest-stakes ten weeks of a firm's year:

• Watch step: The agent polls your existing client portal (TaxDome, SmartVault, Canopy, or even a shared SharePoint/Drive folder) every 15 minutes for new uploads.
• Classify step: A local vision-capable model (we run Claude Sonnet 4.6 for this) reads each upload and tags it — W-2, 1099-NEC, 1099-INT, K-1, mortgage interest statement, or "unrecognized."
• Validate step: It cross-checks the tag against a per-client checklist built from last year's return, so it knows Client #47 should have two W-2s and one 1099-DIV, not three.
• File step: Documents get renamed (LastName_2025_W2_Employer.pdf) and dropped into the correct folder in your DMS.
• Chase step: Anything still missing after 5 business days triggers a personalized reminder — by email or SMS — naming the exact document, not a generic checklist.

Step-by-Step: Building It on OpenClaw

This is the same pattern we use for every BuildAClaw client build, adapted for tax intake:

• 1. Hardware: A Mac Mini M4 (16GB unified memory is enough for this workload) running locally in the office — not a cloud VM. Documents containing SSNs never leave the building.
• 2. Connect the portal: Most portals expose a webhook or polling API. If yours doesn't, OpenClaw can drive a headless browser session against the portal UI directly.
• 3. Build the per-client checklist: Feed the agent last year's return PDFs once. It extracts the document types that generated each line item and builds this year's expected checklist automatically.
• 4. Set the classification model: Run document classification and OCR locally where possible; route only the ambiguous 5-10% of documents to a cloud model call for a second opinion.
• 5. Wire the notification channel: Connect email (Microsoft 365 or Gmail) and SMS (Twilio) so the agent's chase messages go out on your firm's existing sending identity, not a no-reply bot.
• 6. Set the escalation rule: After 3 automated chases with no response, the agent flags the client to a human staffer instead of continuing to nag — this is what keeps clients from feeling harassed.

Automating Follow-Ups Without Sounding Like a Robot

The single biggest driver of response rate wasn't frequency — it was specificity. In a pilot with a 3-preparer firm running this setup, we A/B tested two nudge styles over a six-week window:

The agent's reminders name the exact missing item ("we still need your 1099-NEC from Acme Consulting") instead of resending the full checklist, which cuts the mental load on the client and gets a faster response.

Handling SSNs and Financial Data Without Putting It in the Cloud

This is the part firms get nervous about, and rightly so — IRS Publication 4557 and most state board of accountancy rules put the burden of data security squarely on the preparer, not the software vendor. Running the classification and filing steps locally on the Mac Mini instead of piping every W-2 through a third-party SaaS API means:

• No client SSN or income data touches a vendor's cloud storage by default.
• You control retention — documents can be auto-purged from the agent's working directory once filed to your DMS.
• Full access logs live on hardware you own, which is what most cyber-insurance questionnaires actually ask for.

What This Actually Saves a Small Firm

For the 3-preparer firm mentioned at the top: 22 hours/week of manual document chasing dropped to roughly 3 hours/week of exception handling once the agent was live for a full season. At a blended $35/hour cost for the staff time previously spent on this, that's about $665/week, or $10,640 across a 16-week tax season — against a Mac Mini M4 that costs under $700 once and roughly $75/month in model tokens during peak season.

If you're scaling client automation beyond intake, it's worth reading how we approached a similarly repetitive, high-volume workflow in our recruiting pipeline agent build — the pattern of watch → classify → chase → escalate shows up in both. For more build breakdowns, browse the full Clawticles archive.

Frequently Asked Questions

Can an AI agent legally handle SSNs and W-2s?

Yes, as long as the data stays under your control. Running the agent locally on a Mac Mini M4 instead of routing documents through a third-party SaaS cloud keeps you in control of encryption, retention, and access logs — which is what most state CPA data-security rules and IRS Publication 4557 actually require.

How much does it cost to run a document collection agent during tax season?

A single Mac Mini M4 running OpenClaw handles document intake for 150-250 client households on roughly $60-$95/month in model tokens during peak season (Jan-Apr), plus the one-time hardware cost of the Mac Mini itself.

Will clients actually respond to an AI agent instead of a human staffer?

In our pilot with a 3-preparer firm, automated reminders sent via email and SMS got a 61% same-day response rate, compared to 34% for a manual weekly email blast, mostly because the agent nags on a per-document basis instead of sending one generic checklist.

Does this replace my client portal software?

No. The agent sits on top of your existing portal (SmartVault, Canopy, TaxDome, or a shared drive) and watches for uploads, then sorts, names, and validates them — it doesn't replace the portal, it replaces the person who checks it three times a day.