# OCR for Legal Process Outsourcing Teams at Volume

- Author: Schema Law (https://www.schemalaw.com/)
- Published: 2026-10-10T04:32:25Z
- Updated: 2026-10-11T19:22:20Z
- Category: Choosing legal OCR
- Canonical: https://www.schemalaw.com/blog/ocr-for-legal-process-outsourcing

> How to run OCR for legal process outsourcing across several clients: who picks the tool, written approval, client separation, sampling and plan sizing.

OCR for legal process outsourcing has to work across several clients at once. The provider should run one approved tool where a client allows it, get written approval before any upload, keep every client's documents in a separate space, check output by sampling with a recorded sign-off, and size the plan to its busiest project.

An outsourcing provider is not a law firm with more seats. It works under other people's engagement letters, on other people's deadlines, with staff who rotate between clients. Each of those facts changes how OCR should be chosen and run. This post follows a document from the statement of work to the sign-off sheet.

You work under someone else's engagement letter, so get the hiring attorney's written approval before a page goes up. Keep each client's discovery, pleadings, and medical records in a separate workspace, and run a conflict check when staff rotate onto a new matter. Client confidentiality fails if two clients' files share a folder. Record the sign-off in the project log the same day you deliver.

## Where does OCR fit in outsourced legal work?

OCR fits at the very start of outsourced legal work, as the step that turns the client's scans into text before review, abstraction, coding or summarizing begins. Almost every task an outsourcing team bills for is slower or impossible on image-only pages, so conversion quality sets the ceiling for everything after it.

| Outsourced task | What arrives as scans | What the team needs from conversion |
| --- | --- | --- |
| Document review and coding | Productions, client paper files | Text that can be searched and loaded into the review tool |
| Contract abstraction | Signed agreements, amendments, side letters | Clause text an abstractor can copy and check |
| Medical record summaries | Charts, billing ledgers | Searchable pages and tables as rows |
| Deposition digests | Older paper transcripts, marked exhibits | Text with page references intact |
| Due diligence indexing | Data room exports of mixed quality | A text layer so an index can be built by search |

### Check whether the text already exists

Before conversion, test what the client sent. Sets from a litigation platform often arrive with extracted text files beside the images. Data room downloads are frequently a mix of text-based and scanned PDFs. Converting pages that already carry text uses allowance for nothing, and can replace good text with worse.

### Conversion is a billable step, so define it

Write OCR into the statement of work as its own line: what will be converted, into which format, who checks it and what the acceptance test is. A project that treats conversion as invisible plumbing has no answer when the client asks why a keyword search missed forty documents.

## Who chooses the OCR tool, the outsourcing provider or the client?

The client decides whether an OCR tool may touch its documents, and the provider can propose which one. In practice the choice falls into one of three patterns: the work happens inside the client's own systems, the client approves the provider's standard tool, or the client names a tool the provider must use.

### Three patterns

**Client environment.** The team logs in to the client's platform and uses whatever conversion it offers. The provider chooses nothing and controls nothing. Quality problems go back to the client's administrator.

**Provider's standard tool, approved by the client.** The provider runs one tool across engagements and asks each client to approve it. This is the pattern where the provider's choice matters, and where staff build real skill with one product.

**Client-mandated tool.** The client's security team has approved a short list and the provider must buy or be given access to something on it. Price the learning time into the engagement.

### Why the client holds the veto

The client's lawyers answer for the people they engage. [Rule 1.6 of the D.C. Rules of Professional Conduct](https://www.dcbar.org/for-lawyers/legal-ethics/rules-of-professional-conduct/client-lawyer-relationship/confidentiality-of-information), for example, requires a lawyer to exercise reasonable care to prevent "others whose services are utilized by the lawyer" from disclosing or using a client's confidences or secrets. An outsourcing team is squarely in that group, and so is any tool the team adds. Rules differ by jurisdiction, and a client outside the United States will have its own.

### What the provider should standardize anyway

Even when clients choose, the provider can fix its own practice: one preferred tool, one written description of how it is used, one test set of difficult pages, and one person who owns the relationship with the vendor. A provider that can hand a client a two-page description of its conversion process gets approvals faster than one that improvises each time.

## Getting client approval before an outsourcing team uploads documents

Getting client approval means a written yes, from someone with authority, that names the tool and the documents, before the first page is uploaded. A general clause permitting "standard software" is not enough for a service that receives the client's files, and silence is not approval.

The D.C. Bar's comments to its [Rule 5.3 on nonlawyer assistants](https://www.dcbar.org/for-lawyers/legal-ethics/rules-of-professional-conduct/law-firms-and-associations/responsibilities-regarding-nonlawyer-assistants) describe a lawyer using outside nonlawyers, with a document management company and an internet-based storage service among the examples, and say the lawyer must make reasonable efforts and give appropriate directions. Your approval request is how the client's lawyer gives those directions. Make it easy to answer.

### What the request should contain

1. The tool's name and the vendor's legal name.
2. The vendor's published terms on training, storage, access and deletion, attached or linked, not summarized.
3. Which document sets will be uploaded, and which will not.
4. Who on your team will have access.
5. How long documents will stay in the tool and who deletes them.
6. What happens at the end of the engagement.

### Clauses to read in the master agreement first

- **Subprocessors or subcontractors.** An agreement may require notice or consent before a new one is added. An OCR vendor may count.
- **Location and access.** Some clients restrict where data may be handled or who may see it. Ask the vendor directly. If its published terms do not answer, treat the question as open.
- **Return and destruction.** Your obligations at the end of a project extend to copies in any tool you used.
- **Audit and questionnaires.** A client may send its security questionnaire to you and expect you to pass relevant parts to the vendor.

Schema Law lists the same lines for every plan, among them "Your documents are never used to train an AI model", "Documents stay inside your own account", "Encrypted in transit and at rest" and "Delete any document, at any time". They are set out in the [security section of the homepage](https://www.schemalaw.com/#security). A security questionnaire and contract review is listed as part of the Enterprise plan. Put those statements in front of the client as written and let the client judge them.

### When the answer is no

Some clients will refuse any online tool for some sets. Plan for it. Keep those sets out of the shared process entirely, mark them in the project tracker, and agree with the client how they will be converted, if at all.

## Keeping each client's documents apart in an outsourcing operation

Keeping each client's documents apart starts with one rule: one client, one space, with no shared folders, and access limited to the people staffed on that client. Add separation between matters inside a client where the client has ethical walls of its own.

### Layers of separation

| Layer | What to set up | What to test |
| --- | --- | --- |
| Client | A separate workspace or account for each client | A reviewer staffed on Client A cannot open Client B's space |
| Matter | A workspace per matter inside the client | Names follow the client's own matter numbers |
| People | Access granted by staffing list, removed when someone rolls off | Monthly comparison of access against the current roster |
| Files | Client and matter code at the start of every file name | A stray file can be traced to its owner at a glance |
| Exports | Downloads saved only to that client's project share | No exports on desktops or in personal folders |

How a given tool divides work matters here. Schema Law has no shared workspaces and no matters: each account is one person, and a document can be seen only by the account that uploaded it. Separation there is by person, so a reviewer who converts for two clients holds both clients' documents in one account. Assign conversion with that in mind. Before you rely on any tool's workspaces to separate clients, ask the vendor how access to a workspace is granted and removed, then test it with two accounts. The post on [organizing converted documents by matter](https://www.schemalaw.com/blog/ocr-for-mid-size-law-firm#organizing-converted-documents-by-matter-across-practice-groups) covers naming and structure inside a single firm.

### The mistakes that actually happen

Software settings are only part of separation. Plan as well for the human slips: a reviewer with two browser tabs open, a batch dragged into the wrong space at the end of a shift, or a team lead who reuses last month's upload folder. Three habits help:

- Staff people on one client per shift where you can.
- Have the uploader read the client code on the first file aloud to a second person on large batches. It sounds silly and it works.
- Give each client's space a name that starts with the client code, so the wrong one looks wrong.

### When a file lands in the wrong place

Delete it from the wrong space at once, record what it was and who could have seen it, and tell your own compliance lead the same day. Whether and how the client must be told is governed by your agreement with that client. Do not decide that on the floor.

## How does an outsourcing team check OCR output at volume?

An outsourcing team checks OCR output at volume in three tiers: an automatic count of every file, a human sample of pages from every batch, and a full check of every field or quotation that goes into a deliverable. A named person signs each batch, and a failed sample sends the batch back.

### Tier one: every file

Compare page counts between source and output for every document. A missing page is the worst error and the cheapest to catch. Flag any file whose output is nearly empty, since that usually means a blank, rotated or unreadable scan.

### Tier two: a sample from every batch

Agree the sample plan with the client before the project starts. A workable plan has four parts.

1. **Sample size.** A fixed number or share of pages per batch, chosen with the client. Write it down.
2. **Selection.** Pick pages at random, then add the pages most likely to fail: faxes, tables, stamped pages, handwriting.
3. **What counts as an error.** A wrong character in a name, date, amount or identifier is an error. A stray mark in white space usually is not. Define both.
4. **The threshold.** How many errors in the sample fail the batch, and what happens then: reconvert, rescan or route to manual keying.

### Tier three: everything that leaves the building

Any value abstracted into a client deliverable is checked against the page image by a second person. That covers contract dates, party names, amounts, record dates in a medical summary and every direct quotation. Sampling does not apply here.

### A sign-off line

```text
Client/matter: NRT-0417 Halvorsen lease portfolio
Batch 12 of 30   Files: 214   Pages: 3,880   Page counts match: yes
Sample: 80 pages (60 random, 20 targeted)   Errors: 2 (both in fax pages)
Result: PASS   Checked by: P. Anand   Date: 2026-10-08
Notes: 6 files handwritten throughout; routed to manual abstraction
```

The names in that example are invented. Keep one line like it per batch and hand the sheet to the client with the deliverable.

### What sampling cannot do

A passed sample tells you the batch is broadly sound. It does not tell you a particular page is right. Say so in the project report, so the client does not read "QC passed" as "every word verified."

## Sizing an OCR plan when page volume swings from project to project

Size an OCR plan for swinging volume by the busiest month you can foresee, not the yearly average. Outsourced work arrives as projects, so a provider can convert almost nothing for six weeks and then a quarter of a million pages in three. Monthly allowances punish averages.

### A worked year

Take a fictional provider, Northgate Legal Services, with four projects on the calendar.

| Months | Project | Pages needing conversion per month |
| --- | --- | --- |
| January to February | Lease abstraction, steady | 8,000 |
| March | Second-request review lands | 95,000 |
| April to August | Lease work plus small matters | 12,000 |
| September to October | Medical summaries for a mass tort | 40,000 |
| November to December | Wind-down | 5,000 |

Averaged over the year, Northgate converts roughly 22,000 pages a month. No month looks like the average. Three of them exceed 10,000 pages by a wide margin, and one approaches 100,000.

### Ways to handle the swing

- **Buy for the peak.** A plan built for OCR at 100,000 pages a month covers March. It sits mostly idle in December. This is the simplest choice when peaks are frequent.
- **Change tier by project.** Move up when a large project starts and down when it ends. Before counting on this, ask the vendor how plan changes work mid-month and whether workspaces and settings survive a move down.
- **Stage the work.** Convert in the order review needs it, and spread a set across two billing months where the deadline allows. The guide to [batch OCR for discovery documents](https://www.schemalaw.com/blog/batch-ocr-discovery-documents) covers splitting and ordering large sets.
- **Push conversion upstream.** For a client-mandated platform, the client's allowance, not yours, absorbs the pages.

### Charge the pages to the project

Whatever the plan costs, attribute pages to the engagement that used them. Track pages converted per client per month, and quote conversion in the statement of work. The post on [OCR cost per page](https://www.schemalaw.com/blog/ocr-cost-per-page) has the arithmetic, and the section on [planning volume at one hundred thousand pages a month](https://www.schemalaw.com/blog/enterprise-ocr-for-law-firms#planning-volume-at-one-hundred-thousand-pages-a-month) covers running a plan of that size once you are on it. Current allowances for each tier are on the [pricing page](https://www.schemalaw.com/pricing).

### Do not forget per-document caps

A monthly allowance is one limit. Pages per document and file size are others, and a 3,000-page production volume can break them even when the month has room. Check the cap against your largest expected file during scoping, not on delivery day.

Providers whose work is order-by-order delivery to many firms, with a price list instead of statements of work, will find a closer match in the guide to [OCR for records retrieval companies](https://www.schemalaw.com/blog/ocr-for-records-retrieval-companies).

## 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)