Automate Document Intake
It is the second week of March and your inbox holds a W-2 photographed on a kitchen counter, a bank statement named "scan0042.pdf," and a message that says "see attached" with nothing attached. Somewhere a client believes they sent everything. Somewhere a staff member is about to start a return or a filing with half the file missing, and nobody has written down what is still outstanding.
Document intake is the unglamorous front door of a law or accounting practice. When it works, the professionals see complete, labeled, safely stored files and spend their hours on judgment. When it does not, billable time disappears into hunting, renaming and chasing.
This article is for a small firm, perhaps a handful of professionals and one or two administrators, that wants to automate intake without buying an enterprise platform or compromising client confidentiality.
Start by naming what "intake" means
Intake is not one thing. It is a short pipeline with distinct stages, and each stage can fail on its own.
- Request. You tell the client what you need and by when.
- Receipt. A document arrives through some channel.
- Identification. You work out who it belongs to and what matter or engagement it supports.
- Classification. You decide what kind of document it is.
- Extraction. Where useful, you read data out of it.
- Storage. It lands in the right place with the right permissions.
- Tracking. The request list updates, and anyone can see what is still missing.
- Retention. It is kept, or destroyed, according to a schedule.
Most firms automate stage six by buying a document management system, and stay manual on the other seven. The better return is usually at the front, in request and tracking, because that is where clients and staff lose the most time.
Intake channels, and why fewer is better
Clients will use whatever is easiest for them. If you allow every channel, you inherit every channel. A reasonable target is one preferred channel, one fallback and one exception path.
| Channel | Strengths | Weaknesses |
|---|---|---|
| Secure client portal upload | Authenticated, logged, tied to a request | Clients must create an account and remember it |
| Unique upload link per request | No account, easy on a phone, can expire | Link forwarding is a risk, so keep links short lived |
| Email attachments | Familiar to everyone | Weak confidentiality, size limits, easy to misfile |
| Text message photos | Fast for clients | Lowest quality, no structure, personal phones |
| Mail or hand delivery | Needed for originals and signed paper | Requires scanning, labeling and a log |
| Shared cloud folder | Handy for bulk | Permissions drift, and files get renamed or deleted |
For many firms, the best arrangement is a secure upload link generated per client and per request, with email acceptance as a fallback that an automation moves into the same pipeline. The point is not the specific product. The point is that every path ends in the same queue.
Attachments sent by email deserve a sentence of caution. Standard email is not a confidential channel by default, and your professional rules may have specific expectations about client communications. Check your obligations with your licensing body or counsel, and treat this article as general information rather than legal advice.
Make the request the center of the system
The most useful object in intake is not the document. It is the request item. A request item says "Client A, 2025 tax engagement, needs Form 1099 from brokerage, due March 20, status outstanding."
When a document arrives, it is attached to a request item, and the item changes state. This one design choice creates the client-facing tracking that people ask for, because the checklist and the files are the same data.
A simple state model works well.
| State | Meaning |
|---|---|
| Requested | Sent to the client, nothing back yet |
| Reminded | One or more reminders sent |
| Received | A file arrived and is awaiting review |
| Accepted | Staff confirmed it is complete and legible |
| Rejected | Staff asked for a replacement and gave a reason |
| Not applicable | The client says it does not exist, and staff agree |
A "Received" file is not the same as an "Accepted" one. Clients routinely send the wrong year, a single page of a multi-page statement, or a blurry photo. If the system marks the item done at upload, the checklist lies.
Give the client a simple status page or a weekly summary that shows what is accepted and what is still outstanding. Clients respond well to seeing a short list rather than receiving a general reminder to "send the rest."
Identification before classification
Before you can sort a document, you need to know whose it is. Identification is the step firms underestimate.
If the document arrives through a per-request link, identification is free. The link itself says who and what. That is the strongest argument for that channel.
If it arrives by email, you rely on the sender address, which fails when a spouse, an assistant or a bookkeeper sends it. A sensible rule is to match known sender addresses to a client record, and send anything unmatched to a human triage queue instead of guessing. A wrong guess here is worse than a delay, because a document in the wrong client's file is a confidentiality incident.
Never use the file name as the sole identifier. File names are chosen by scanners and phones, and they carry no meaning.
Classification, and the difference between OCR and structured data
Classification means deciding that a file is a bank statement, a deed, an engagement letter or a pay stub. There are three approaches, and firms often use a mix.
Rules. If the request item the file was uploaded against says "brokerage 1099," the file is classified as that. This is the cheapest and most reliable method, and it works whenever upload is tied to a request.
Text matching. Optical character recognition (OCR) turns an image of a page into text. Once you have text, simple searches for phrases such as "Form W-2" or "Statement Period" can classify many documents. OCR quality depends heavily on the image, so a skewed phone photo reads worse than a clean scan.
Machine learning or language models. These can classify messy documents that rules miss. They return a confidence level, and you decide a threshold below which a person must review. Treat the output as a suggestion, and keep a human in the loop for anything consequential.
The distinction between OCR and structured data matters. OCR gives you text, which is unstructured, and then you must interpret it. Some documents arrive as structured data already, such as electronic statements in a defined format, exports from accounting software, or e-filed forms. If a structured version exists, request it. It is exact where OCR is approximate, and it removes a whole class of transcription errors.
A worked scenario (hypothetical)
A three-person CPA practice receives about 40 documents on a busy day. Each takes roughly two minutes of staff time to identify, rename and file, which is about 80 minutes. If upload is tied to request items, identification and classification are automatic for most files. Staff review only the files that fail a basic check, say 8 out of 40, at two minutes each, about 16 minutes. The difference of roughly an hour a day is illustrative and will vary, but the shape is general. The savings come from removing decisions, not from reading faster.
Automatic checks that catch problems at the door
A handful of cheap checks, run when a file arrives, prevent a lot of later rework.
- Readability. Is the page count plausible, is the resolution adequate, is the file unencrypted or password protected in a way you cannot open?
- Completeness. If a statement says "page 1 of 4" and there are two pages, flag it.
- Period and name. Does the tax year or statement period match the request, and does the name on the document match the client?
- Duplicates. A file hash (a short fingerprint computed from the contents) identifies exact duplicates. Near duplicates need a similarity check, but exact duplicates alone remove much noise.
- Malware scanning. Any file from outside your organization should be scanned before it is opened by staff.
- File type. Reject executable files and unusual formats at the door.
Tell the client at once if a file fails, with a plain reason such as "this statement is missing page 3." A fast, specific message beats a generic reminder weeks later.
Confidentiality is a design requirement
Law and accounting firms hold sensitive information, such as identification numbers, financial records, medical details and privileged communications. Automation must not weaken the protection of any of that.
Practical principles worth applying.
- Least privilege. Staff see only the matters they work on. Automation accounts get only the access they need, nothing more.
- Encryption. Data should be encrypted in transit and at rest. Ask vendors to state how they do this in plain terms.
- Authentication. Use multi-factor authentication for staff and for any client portal.
- Expiring links. Upload and download links should expire, and be single purpose.
- Logging. Record who uploaded, viewed, downloaded and deleted what, and when. Keep logs free of the document contents themselves.
- Vendor terms. Ask where data is stored, who can access it, whether it is used to train models, and what happens when you leave.
- Language model caution. If you send documents to an external AI service for classification, confirm in writing how that provider handles and retains your data. Some matters may not be suitable for that at all.
Your own professional obligations, and local law, shape what is acceptable. This is general information and not legal advice, so have counsel or your compliance contact review the design.
Retention is part of intake
Every document that comes in will one day need to be kept for a defined period or destroyed. Decide that at the beginning, not after a decade of accumulation.
Attach a retention category to each document type or each matter, with a trigger such as "closed date" or "filing date." The system can then compute a review date, and flag documents for deletion or for a legal hold review. Requirements vary by profession, jurisdiction and document type, so confirm your schedule with the relevant authority or counsel. Also remember that deletion should apply to backups and to copies in email, and that a legal hold must suspend deletion.
Exceptions and failure handling
An intake process is judged by how it behaves when things go wrong.
- Unmatched sender. Send to triage, never auto-file.
- Unreadable file. Notify the client, keep the original, and log the reason.
- Wrong document. Move the request item to rejected, not accepted, and explain.
- Client silence. Escalate by reminders on a schedule, then to a staff call after a defined number of attempts.
- System outage. Have a fallback channel and a written procedure. For firms in Puerto Rico, plan for storms. Cloud-based intake that stays reachable when your office has no power or connectivity is a real advantage, and so is a documented way to receive documents by phone call and secure link when the usual path is down.
- Retries. If a job that files documents is interrupted and rerun, it should not create duplicates. Use idempotent steps, meaning that doing the same step twice leaves the same result as doing it once, for example by storing a file hash and skipping files already filed.
What staff still own
Automation should handle the repeatable parts. People own the following.
- Deciding whether a document is sufficient for the professional purpose.
- Reviewing anything flagged as low confidence or unusual.
- Talking to clients who are confused, anxious or slow.
- Making judgment calls on privilege, relevance and exceptions.
- Approving changes to the rules and the retention schedule.
A good intake system gives them a short, ordered queue and good context, rather than an inbox.
Bilingual clients
Many firms in Puerto Rico serve clients in both Spanish and English. Request messages, reminders and status pages should be available in the client's preferred language, stored as a field on the client record. Documents themselves arrive in either language, so test your OCR and classification on Spanish documents before relying on them. Accented characters and local form layouts can trip up tools tuned mainly for English.
Questions for a vendor
- Can uploads be tied to a specific request item or matter, without the client creating an account?
- Where is data stored, how is it encrypted, and who at the vendor can access it?
- Does the product log every access, and can you export logs?
- Is there malware scanning on upload?
- Can classification confidence be set and reviewed, with a human review queue?
- Does it support retention schedules and legal holds?
- Can I export all documents and metadata if I leave?
- Is the client experience available in Spanish?
A Monday morning plan
- Choose one document type that causes pain, such as year-end tax documents or engagement onboarding.
- Write the checklist for it and put every item into a spreadsheet with states from the table above.
- Pick the one channel you will promote, and generate per-client links or portal folders for it.
- Write the three automatic checks you will run on arrival, and the messages sent when each fails.
- Test with two friendly clients. Time how long it takes staff to handle each file before and after.
- Review the exceptions after two weeks and adjust before widening.
If you would like to talk through the design for your firm, I am glad to. Ozcorp InfraOps builds this kind of workflow for small teams, and InfraOps Base starts at $500 per month with vendor and license charges billed separately. Either way, start with the request list. Once clients can see what is missing, most of the chasing takes care of itself.
Discuss your infrastructure requirements
Review the current service offer or schedule a conversation about the scope you need.
Stay in touch
Follow Ozcorp
Leave your details for occasional notes about new articles and Ozcorp updates.