5 things you can do with private AI without uploading anything to the cloud

You do not need to upload a client's file to the cloud to get leverage from AI. Research IRS guidance on a local model, extract fields from PDFs on-device, draft client letters, translate between English and Spanish, and classify documents into the right folder — the same five jobs firms currently send to public chat tools, done without the trip. These are capabilities, not a shopping list for a device that is not on sale this week.

These five jobs match what staff already paste into Level 1 tools during tax season. They belong on Level 3 when a client file is involved — meaning the processing happens on hardware in the office and the file never crosses a third-party API. Framework: three levels of AI security. The knowledge-vs-documents routing: questions vs. documents.

1. Research IRS guidance locally

Load IRS publications and instructions — documents you have already retrieved through legitimate public channels — onto a machine in the office. Ask "where does Publication 17 discuss gambling winnings?" without attaching the client's casino statements or W-2G. The model works from the publications, not from the client's file.

Knowledge research can be local or cloud — the knowledge-vs-documents split allows either, as long as no client PII travels with the question. This item is specifically for offices that want even their research to stay offline — perhaps because they work in environments where any internet-connected AI use requires additional approvals.

Practical advantage: local research does not depend on internet speed, cloud vendor uptime, or API rate limits during the January–April crunch when everyone is querying the same cloud endpoints.

2. Extract data from PDFs on-device

W-2 boxes, 1099 payer names, employer EINs, dates, dollar amounts. OCR and field mapping do not require a Silicon Valley endpoint. They require a model or rules engine that runs on the machine next to the file. That is the core of "the data never left."

The journey you are specifically avoiding: the W-2 upload path — where the file crosses the internet, lands on an external server, gets processed by software you do not operate, and is "deleted" on a schedule you cannot independently verify.

On-device extraction does not require perfection. Bad scans and handwritten notes will still sometimes need a human eye. The goal is not to eliminate human review — it is to eliminate the outbound trip for the file. If the OCR misreads box 1, you fix it locally. If the OCR runs in the cloud, you already made a disclosure before you even checked the accuracy.

3. Draft client letters

Organizer reminders, missing-document checklists, "please sign Form 8879" in plain language, year-end thank-you notes, engagement-letter boilerplate. The local tool takes a structured list of missing items from your CRM — not a dump of the entire organized PDF into a public model.

XyloDocs already holds the client conversation in SMS, WhatsApp, and portal, so the letter has a delivery channel that is not Gmail. The draft job belongs on local infrastructure when it references specific client facts; the delivery job uses the communication platform the client already responds to.

This is one of the fastest wins for a firm exploring local AI: letter templates with fill-in-the-blanks are straightforward for even modest local models, and they eliminate the temptation to paste a client's entire file into a public chat "just to draft a quick reminder."

4. Translate English ↔ Spanish

Bilingual practices should not have to choose between "paste into a public translator" and "hire a professional translator for every thread." A local translation model keeps the facts of the letter in the office — the client's name, their specific dollar amounts, their case details — while producing the Spanish (or English) version on-device.

The communication platform can then deliver the message through WhatsApp or SMS — channels the client already uses for daily communication. Translation is the job that stays local. Delivery is the job the communication platform handles.

Colorado's tax community serves a significant Spanish-speaking population. Bilingual communication is not a nice-to-have — it is how some firms get half their book of business. But the bilingual workflow is especially sensitive because it often involves client-specific facts (dollar amounts, filing statuses, document names) embedded in the translated text. Those facts are return information, and they belong on local infrastructure.

5. Classify documents

W-2 vs. 1099-NEC vs. 1099-INT vs. driver's license vs. prior-year return — sorting uploaded documents into the right categories is how piles become organized files. Doing it in a public assistant teaches a third party your client's entire document mix. Doing it on-device fills the same folder your preparer already opens in the tax software.

Related reading: tax document automation for the broader workflow problem. Classification is the first step in that workflow — and it is the step most commonly outsourced to a public chat tool because "it's just sorting." The sorting still sees every SSN and employer name in the pile.

A local classifier can also flag incomplete submissions — "this pile has a W-2 and a 1099, but no ID and no signature form" — which turns the classification step into a completeness check that feeds your missing-document workflow.

What these five are not

• Not a claim that XyloDocs ships a local LLM appliance today. The 2027 tax software is designed with document-local architecture. The communication platform is available now. These are different products on different timelines.
• Not a claim that local models never make mistakes. They hallucinate, misread, and miss edge cases just like cloud models. You still review. Privacy is about where the file lives during processing, not about whether the answer is perfect.
• Not permission to skip §7216 when a human vendor still receives files through another channel. Local AI for documents does not exempt the rest of your vendor relationships from disclosure analysis.

CRM you can use now: $49/month platform base, additional seats $9/month, metered usage from prepaid balance. Client accounts are unlimited and always free.

Frequently asked questions

Is on-device OCR as good as cloud OCR?

It depends on the engine, the model size, and the scan quality. Privacy is the reason to try local first for documents with SSNs — not a promise of perfect extraction on every crumpled receipt. Bad scans still need a human review regardless of where the model runs.

Can I do these five in a spreadsheet macro instead of AI?

Sometimes, for structured data. AI helps with unstructured inputs — photos, handwritten forms, inconsistent PDFs. The test remains: did the file leave the office? If a macro keeps it home and gets the job done, the macro is private enough.

Why not just use the cloud for all five with a BAA-style contract?

You can choose Level 2 with eyes open — that is a legitimate decision if documented in the WISP and supported by proper consents. This article exists because most offices are accidentally on Level 1 — using consumer chat tools with personal accounts — for these same five jobs. The minimum improvement is moving from undocumented Level 1 to documented Level 2 or Level 3.

What if my firm does not have the IT capacity for local AI?

Then enforce the knowledge-vs-documents split by policy: public AI for research questions only, no client files in any external AI tool. Use the firm's existing software for document tasks, and evaluate 2027 product options that include local processing as a built-in feature.

The bottom line

Research, extract, draft, translate, classify — none of those jobs inherently requires a public upload. Stop sending the file to do work the file could stay home for. The five jobs are the same ones offices already do in public chat windows; the difference is routing them to local infrastructure where the file does not leave.