By Sagar Shankaran, Founder of CallSphere
Consolidated 1099s, SSNs and IRS transcripts cannot leave a tax firm. What local AI hardware changed in 2026, with a season-cost comparison and honest limits.
Key takeaways
Count it once and the number stops being abstract. A practice with 1,100 individual returns and 180 business returns runs somewhere north of 35,000 pages of source documents through the scanner in ten weeks — W-2s, 1099-NECs, consolidated brokerage statements that run 60 pages for a single client, closing disclosures, mortgage interest statements, K-1s with two-page footnotes, a shoebox of receipts from the contractor who has been a client since 1998.
Every one of those pages contains a Social Security number, an account number, or both. Every one of them is return information inside the meaning of Internal Revenue Code §7216, which makes it a criminal matter to disclose or use it without the client's consent in the form the regulations require. And every one of them, in most firms, currently gets uploaded somewhere.
That last sentence is the thing that changed this year.
Because the alternative was hiring. Scanning-and-organizing services — some domestic, plenty offshore — took the shoebox off the firm's hands for a flat fee per return, and firms told themselves the consent problem was handled by a paragraph in the engagement letter. Often it wasn't. Disclosure of return information to a preparer outside the United States has its own consent requirements and its own limits on what identifying detail may be sent, and a generic engagement-letter clause does not satisfy the regulation's format rules.
Meanwhile the cloud AI tools that arrived in 2024 and 2025 had the same shape: to read the document, the document had to go somewhere else. For a lot of firms that was survivable for a summer trust return and unthinkable for the client who is a defence subcontractor with contract language forbidding it, or the client whose attorney sends everything under privilege, or the audit clients whose workpapers sit under peer review scrutiny.
So most practices did the split: automate the easy 60%, and let the partner's assistant hand-key the sensitive 40% at eleven at night in April.
Hear it before you finish reading
Talk to a live CallSphere AI voice agent for financial services in your browser — 60 seconds, no signup.
The hardware caught up with the model sizes. Qualcomm's Dragonwing-class processors put genuinely capable document reading on a machine that fits in a server closet and does not need a data centre behind it. At the enterprise end, Cisco has been rolling a personal AI agent out to roughly 90,000 employees with an explicit on-premises emphasis — meaning the work happens on their own equipment, for control and data protection, not because it is cheaper.
And it is also cheaper. Running this kind of high-volume reading on your own hardware costs roughly 90% less than sending the same volume to a cloud service, which matters when the volume is 35,000 pages in ten weeks rather than forty pages in a quiet August.
Running AI in your own office means the client's tax documents are read by a machine that sits on your own network, behind your own door, and no copy of the Social Security number ever leaves the building. For a tax practice that is not a technical preference. It is the difference between needing a §7216 consent and not needing one, because nothing was disclosed to anybody.
flowchart TD
A["Front desk scans the client's source documents"] --> B["Local machine reads every page inside the office"]
B --> C{"Missing cost basis or an unmatched wash sale?"}
C -->|Yes| D["Request a corrected 1099-B from the client's broker"]
D --> A
C -->|No| E["Numbers written to the current-year workpaper"]
E --> F["Preparer reviews on the same network, nothing uploaded"]
F --> G["Partner signs, return goes to e-file"]
Not all of it, and pretending otherwise makes the conversation useless. Here is the honest tiering that most firms land on.
Layer on top of that the FTC Safeguards Rule, which is the reason your firm has a Written Information Security Plan at all. The Rule wants a named qualified individual, an inventory of where customer information lives, encryption, multi-factor authentication, and oversight of service providers. Every document that stays inside the building is one fewer service provider to oversee and one fewer entry on the inventory.
Illustrative, and every number here is yours to replace. Suppose a 1,100-return practice sends 620 of those returns to an outside scanning and organizing service at $18 each, and the firm administrator spends nine minutes per client chasing and filing the §7216 consents that arrangement requires.
| Line item | Assumption | Season cost |
|---|---|---|
| Outside scan-and-organize service | 620 returns at $18 | $11,160 |
| Consent chasing and filing | 1,100 clients, 9 min, $28/hr admin | $4,620 |
| Cloud document reading for the rest | Illustrative | $1,900 |
| Current total | $17,680 | |
| Local machine, purchased | $6,400, written off over 3 seasons | $2,133 |
| Setup, power and support | Illustrative | $2,400 |
| New total | $4,533 |
The gap is about $13,000 a season, which is real but is not the argument. The argument is the line that does not appear in either column: the cost of explaining to 1,100 clients that their Social Security numbers were sitting with a vendor that had a breach. Firms that have lived through that will tell you the notification letters, the credit monitoring, the state attorney general filings and the clients who quietly moved to the firm across town cost more than three years of anything in that table.
It does not make your building secure. A machine in a closet that anybody can walk to, with the same password since 2019, is worse than a well-run cloud service, not better. The Safeguards Rule's requirements — access controls, multi-factor authentication, encryption at rest, a documented plan — apply to your own equipment exactly as they apply to a vendor's.
Still reading? Stop comparing — try CallSphere live.
See the financial services AI agent handle a real call — complete, industry-specific, and live in your browser. No signup.
It does not eliminate consent. If the return information goes out to a lender, an attorney, or a client's financial advisor, §7216 still governs that, whether the reading happened in your closet or in Virginia. Local reading removes one disclosure from the list; it does not remove the list.
It does not do everything the biggest cloud models do. For reading, sorting and extracting from tax documents, a well-chosen local machine is more than adequate. For the hardest research questions — a novel state nexus issue, a complicated §1031 fact pattern — you will still want a frontier model, and the sane arrangement is to send it the question with the client facts stripped out.
And it does not replace the reviewer. The machine is reading, not concluding. Whether the noncovered lots on that 1099-B were reported correctly by the broker is still a human judgment, and it is a judgment the IRS expects the signing practitioner to have made.
No, but you need one person who owns it. In practice that is the same person who is already your qualified individual under the Safeguards Rule — often the firm administrator, sometimes a partner, sometimes an outside managed services provider on a monthly contract. The machine itself is an appliance. The discipline around who can touch it is the actual work.
Increasingly, yes. Carriers have been adding questions about AI use and about where client data is processed at renewal. Being able to answer "our document reading happens on firm-owned equipment on our own network, and here is the entry in our written plan" is a materially easier conversation than describing a vendor chain.
Ask what they actually mean, because most of them mean "don't send my file to a company I've never heard of." Local processing answers that directly. If a client still objects, keep a manual path for their file and note it in the engagement letter — a handful of exceptions is manageable, and it is a better answer than a blanket policy that costs you the other 1,090 clients' worth of savings.
Absolutely not. Buy in July, test in August and September on last season's filed returns, run it in parallel through the 15 October extended deadline, and make the go/no-go call in November. Nobody should be installing a machine in the file room in March.
Before you price hardware, do the inventory. One page: every place client tax data currently sits outside your four walls, who put it there, and what consent covers it. Most firms find two or three surprises on that page — a partner's personal cloud drive, a scanning vendor nobody re-papered after 2022, a portal from a software product the firm stopped using. Fix those and you have improved your position before spending a dollar.
The one thing that will never live inside the building is the phone call, and February brings a lot of them. CallSphere builds AI voice and chat agents that answer the firm line and website chat, book the review meeting and take the caller's details so the front desk isn't the bottleneck while everyone else is in returns. Keep the documents in your closet; let something else handle the ringing.

Written by
Sagar Shankaran· Founder, CallSphere
LinkedInSagar Shankaran is the founder of CallSphere, where he builds production AI voice and chat agents deployed across healthcare, hospitality, real estate, and home services. He writes about agentic AI, LLM engineering, and shipping voice agents that handle real calls in production.
See how AI voice agents work for your industry. Live demo available -- no signup required.
Supplement brands get adverse events reported by ticket, not by form. On-premises AI reads all 3,400 a month without that text ever leaving the building.
Client PDFs are attacker-supplied documents. How a CPA firm scopes AI agent permissions, and the irreversible tax actions that always need a named human.
Concealed-damage claims cost a warehouse manager half a Friday. On-site AI searches your own footage and scan trail without any data leaving the property.
Casino surveillance footage cannot leave the property. On-premises AI in 2026 cuts a 90-minute disputed handpay review to nine, with no video sent to a vendor.
Custody orders, health plans, subsidy files and classroom video can stay in the building. What on-premises AI unlocks for a child care center in 2026.
A cloned controller can reroute a payroll run or a Bill.com vendor payment. The verification step, dual control and engagement-letter passphrase that stop it.
© 2026 CallSphere Inc. All rights reserved.
Made within San Francisco
Watch how CallSphere handles real customer calls, schedules appointments, and processes payments — live.
Try Live DemoBook a DemoCalculate Your ROI