# Can OCR Vendor Employees See My Documents? How to Check

- Author: Schema Law (https://www.schemalaw.com/)
- Published: 2026-02-07T10:05:03Z
- Updated: 2026-10-11T19:22:20Z
- Category: Confidentiality
- Canonical: https://www.schemalaw.com/blog/can-ocr-vendor-employees-see-my-documents

> Can OCR vendor employees see my documents? Assume some can until the vendor says otherwise. How to find out who, why, and what controls a vendor documents.

Whether OCR vendor employees can see your documents depends on the vendor, and you should assume some can until the vendor says otherwise in writing. An online service has to read your file to convert it, and people run its systems. What differs is who is allowed, for what reasons, and whether each access is approved and recorded.

This post shows where staff access comes from, what documented controls look like at a few large providers that publish theirs, and which questions to send a vendor. It describes how to check. It does not tell you that any particular vendor passes.

## Can an OCR vendor's employees open the documents you upload?

Assume that some employees can, until the vendor tells you otherwise in writing. An online OCR service has to read your file to convert it, and the systems that hold the file are run by people. A statement that a file is encrypted or kept apart from other customers does not answer whether the vendor's own staff can open it.

Three kinds of promise do not answer the staff question.

| What a vendor's page promises | What it covers | What it leaves open |
| --- | --- | --- |
| Encryption in transit and at rest | Interception on the network, and theft of the storage media | People who operate the system that holds the keys |
| Documents stay in your own account | Other customers and the public | The vendor's own engineers and support staff |
| No training on your documents | Use of the file to build a model | A person reading the file for another reason |

Each of those promises is worth having. None of them is a staff access policy. The post on [OCR tools that do not share data between customers](https://www.schemalaw.com/blog/ocr-tool-that-does-not-share-data-between-customers) deals with the second row, which is separation from other firms, a different subject from the vendor's own people.

Vendor security includes which people can open a page and for which purpose. The duty of competence is why you ask before a privileged memo leaves the office. Client confidentiality still covers that memo when a marketing page sounds careful. Put information security questions about approvals and logs in the same written request. Save the reply next to the matter if the file is restricted.

The ethics guidance points at the same question. [ABA Formal Opinion 512](https://www.americanbar.org/content/dam/aba/administrative/professional_responsibility/ethics-opinions/aba-formal-opinion-512.pdf) says lawyers should read the terms and policies of a generative AI tool "to learn who has access to the information that the lawyer inputs into the tool". The opinion is about generative AI tools, and it is advisory, but the habit it describes fits a conversion service as well: find out who has access before the file goes up.

Schema Law is no exception to the method. Its site makes four commitments and says nothing either way about which of its people can open a file. The post on [Schema Law security](https://www.schemalaw.com/blog/schema-law-security) lists what the site states and what it does not, including staff access.

## Why would vendor staff look at an uploaded document?

The two vendor pages quoted in this section give a short set of reasons: fixing a customer's problem, investigating misuse of the service, and complying with the law. Reading the file to improve the product, or for no stated reason, is what your diligence is meant to rule out.

Published wording gives you a vocabulary for this. OpenAI's [enterprise privacy page](https://openai.com/enterprise-privacy/), as archived in October 2026, says of its ChatGPT Business product that its access to stored conversations "is limited to (1) authorized employees that require access for engineering support, investigating potential platform abuse, and legal compliance and (2) specialized third-party contractors who are bound by confidentiality and security obligations, solely to review for abuse and misuse." For its Enterprise product the same page is narrower: authorized employees "will only ever access your conversations for the purposes of resolving incidents, recovering end user conversations with your explicit permission, or where required by applicable law."

That is a chatbot company describing chatbot products, not an OCR service. It is useful here because it names purposes one by one, and because it shows one company stating different access rules for two of its own plans. Check which plan a vendor's statement applies to.

Microsoft's documentation describes the support case in plain terms. Its page on [Customer Lockbox requests](https://learn.microsoft.com/en-us/purview/customer-lockbox-requests) says engineers "fix issues using extensive telemetry and debugging tools", and then: "However, some cases require a Microsoft engineer to access your content to determine the root cause and fix the issue."

Turn those examples into a list of purposes to look for in your own vendor's terms.

- **Support.** You report that a page converted badly, and someone opens the page to see why.
- **Debugging.** A file crashes the converter, and an engineer examines the file.
- **Abuse review.** The vendor checks whether an account is being used for something its terms forbid.
- **Legal compliance.** The vendor is compelled to produce or preserve files. The post on [OCR vendor subpoenas](https://www.schemalaw.com/blog/ocr-vendor-subpoena) covers what legal process can require of a vendor's staff.
- **Quality checks and product improvement.** Staff compare output against source pages to find errors.

For each purpose the vendor names, three follow-ups matter: who decides that access is needed, whether you are asked or told, and whether a record is kept.

## Human review of OCR output and data labeling by vendor staff

Human review means a person looks at a page and its machine-read text and corrects or grades the result. The documentation opened for this post shows that such review is built into machine-reading products as an option. It does not show whether any given OCR vendor has its own staff review customer pages. You have to ask.

Here is what the documentation does establish. Amazon's page on [using Amazon Augmented AI for human review](https://docs.aws.amazon.com/sagemaker/latest/dg/a2i-use-augmented-ai-a2i-human-review-loops.html) describes a service that provides "human review of low-confidence predictions or random prediction samples". It gives a document example: "extracting information from scanned mortgage application forms can require human review due to low-quality scans or poor handwriting." The same page notes that the service is no longer open to new customers.

Two things follow from that page, and one does not.

- A human review step for scanned documents is a documented design. The page's own example ties it to low-quality scans and poor handwriting.
- The page describes review that a customer of the service switches on and configures for its own application. A company building a document product could be such a customer.
- The page says nothing about any particular OCR vendor's practice. It is not evidence that a vendor you are considering uses human reviewers.

Data labeling is the related practice of having people mark up sample pages so a model can be trained or tested on them. If a vendor promises not to train on your files, labeling your files for training would break that promise, so the two questions overlap. The post on [how to verify a no-training claim](https://www.schemalaw.com/blog/ocr-that-does-not-train-on-your-data#how-can-you-verify-a-no-training-claim) covers the training side.

### What to look for in the terms

Search the vendor's terms, privacy policy and help pages for these words: "human review", "manual review", "quality assurance", "annotation", "labeling", "contractor", "workforce" and "improve". Then read the sentences around each hit for three facts.

1. **Is review on by default, or only when you request it?**
2. **Who are the reviewers?** Employees, contractors, or a third-party workforce. OpenAI's wording quoted in this post names contractors as a separate group with a separate purpose, which shows that the distinction is one vendors can state.
3. **Which files are eligible?** All uploads, a random sample, or only files you flag.

If the documents are silent on human review, that is a finding to record, and a question to send. Silence does not mean no review happens.

### A fit for the document, not only the vendor

Suppose a vendor tells you that staff review a small random sample of pages for quality. For a stack of recorded deeds that may be acceptable. For a privileged memo or a sealed record it may not be, and the right response is to keep that document out of the tool or to ask whether sampling can be switched off for your account.

## Support tickets and screen shares that expose a client file

Some staff access is access your own firm hands over. When you attach a client's page to a support ticket, paste text into a chat window, or share your screen with a technician, you have shown the document to a vendor employee through a channel that the product's security commitments may not cover.

This route is easy to miss because it feels like asking for help. Consider a fictional example. Priya, a paralegal at a small employment firm, uploads a scanned personnel file for a client, Ms. Alder. A table of pay dates comes out with the columns shifted. Priya emails the vendor's support address, attaches the PDF, and writes "see page 6". The PDF now sits in a help desk system, in the support agent's mailbox, and in Priya's sent mail. None of those is the workspace where the vendor's deletion control operates. Deleting the document from the product later may not reach the copy attached to the ticket, so Priya has to ask.

A safer routine for the same problem:

1. **Describe before you show.** Say what kind of page it is and what went wrong: "A six-column table on a scanned payroll report loses its third column."
2. **Reproduce on a stand-in.** Build a page with the same layout and invented content, confirm it fails the same way, and send that instead.
3. **If only the real page shows the fault, ask first.** Ask how the vendor's staff would view it, where the copy would be stored, and when it would be deleted. Get the answer in writing.
4. **On a screen share, close everything else.** Share one window, not the desktop. Close the matter list, email and other documents first. Ask whether the session is recorded.
5. **Never grant account access casually.** If support asks to sign in as you or to be added to your workspace, treat that as a decision for the supervising lawyer.
6. **Close the loop.** When the ticket is resolved, ask the vendor to delete attachments and confirm it, and note that in the matter file.

A lawyer's duty to supervise extends to this. Comment [3] to [ABA Model Rule 5.3](https://www.americanbar.org/groups/professional_responsibility/publications/model_rules_of_professional_conduct/rule_5_3_responsibilities_regarding_nonlawyer_assistant/comment_on_rule_5_3/) lists "sending client documents to a third party for printing or scanning, and using an Internet-based service to store client information" among its examples of outside help, and says a lawyer "must make reasonable efforts to ensure that the services are provided in a manner that is compatible with the lawyer's professional obligations." Your state's version of the rule and comment governs.

## What do staff access controls look like in a vendor's documentation?

Documented staff access controls have three parts: access limited to named purposes and roles, an approval step before a person opens customer content, and a log of each access that the customer can read. The examples here come from large cloud and software providers, which are not OCR vendors, because they publish the detail.

### Approval: the customer says yes or no

Microsoft's Customer Lockbox page says the feature "ensures that Microsoft can't access your content to do service operations without your explicit approval." It describes the workflow: a support engineer makes a request that states the service request number and "the estimated amount of time the engineer needs access to the data", a Microsoft manager approves it, and the customer's designated approver is notified. Then: "If the customer rejects the request or doesn't approve the request within 12 hours, the request expires and no access is granted to the Microsoft engineer." Access is "automatically revoked" after the requested duration. The page describes this as a control the customer turns on.

Google Cloud documents a similar step. Its [overview of Access Approval](https://cloud.google.com/assured-workloads/access-approval/docs/overview) says the feature "ensures that Cloud Customer Care and engineering teams require your explicit approval whenever they need to access your Customer Data", and that "Access is granted only after the Access Approval request is approved."

### Logging: the customer can see what happened

Google's [overview of Access Transparency](https://cloud.google.com/assured-workloads/access-transparency/docs/overview) says its logs "record the actions that Google personnel take when accessing Customer Data", and that entries "include details such as the affected resource and action, the time of the action, the reason for the action, and information about the accessor." The Microsoft page says: "All actions performed by a Microsoft engineer are logged in the audit log. You can search for and review these audit records."

### Purpose limits: access only for stated reasons

The OpenAI wording quoted in this post is an example of a purpose limit: access for "engineering support, investigating potential platform abuse, and legal compliance", and nothing else listed.

### Using the examples

These pages describe specific products from large cloud and software providers, and some of the controls are optional features. Do not expect to find the same thing elsewhere, and do not read this post as saying any OCR service, Schema Law included, has them. Use the examples as a yardstick for what a complete answer contains.

| Control | A complete answer says | A thin answer says |
| --- | --- | --- |
| Purpose | The listed reasons staff may open content | "Only when necessary" |
| Roles | Which job functions, and roughly how many people | "Authorized personnel" |
| Approval | Who approves, and whether the customer is asked | Nothing |
| Time limit | Access expires after a stated period | Nothing |
| Record | Each access is logged, and the customer can obtain the log | "We monitor access" |
| Contractors | Whether non-employees are included, and on what terms | Nothing |

A thin answer is not proof of a problem. It tells you the next step is a written question.

## Questions about staff access to put to an OCR vendor in writing

Send the questions by email, to a named contact, and keep the reply with your vendor file. A written answer is a record you can keep and point to. It also shows the inquiry you made before sending client files. Ten questions cover the ground.

1. Which roles at your company can open a customer's uploaded document or its converted text?
2. For what purposes may they do so? Please list them.
3. Does a person need approval before opening customer content? Who gives it?
4. Is the customer asked or notified before staff open a file? After?
5. Is every staff access recorded, with the person, time and reason? Can we obtain the record for our account?
6. Do contractors or other third parties have the same access? Under what confidentiality terms?
7. Do people review customer pages or output for quality, on any plan? Is it on by default? Can it be turned off?
8. Are customer files ever labeled or annotated by people, for any purpose?
9. When we send a file to support, where is it stored, who can see it, and when is it deleted?
10. After we delete a document, can any staff member still retrieve it? For how long?

### Reading the reply

- **Specific beats reassuring.** "Two on-call engineers, with manager approval, logged" tells you more than "we take privacy seriously".
- **Ask where the answer is written down.** A sentence in the terms or a signed agreement is part of your contract with the vendor. A support agent's email is not, but it is still worth having.
- **Check the plan.** An answer about an enterprise tier may not describe the tier you pay for. Schema Law, for instance, lists "Security questionnaire and contract review" on its Enterprise plan only; the [pricing page](https://www.schemalaw.com/pricing) shows what each plan includes.
- **Record a refusal.** If a vendor will not answer, write that down with the date. Then decide which documents, if any, you are willing to send it.

Match the answer to the file in front of you. A vendor whose staff may open content for support, with a log, may be acceptable for routine records and not for privileged or sealed material. The decision belongs to the lawyer responsible for the matter, made with the vendor's written answers in hand and, where your rules call for it, the client's informed consent.

## Try it on a real document

Upload a scanned PDF and export it as Word, a searchable PDF or plain text on the free plan. Your documents are never used to train an AI model. [Start free](https://www.schemalaw.com/api/auth/google)