OCR Software Comparison: Building a Side-by-Side Table
By Schema LawPublished Updated
An OCR software comparison done before any trial is a table of facts copied from each vendor's own public pages: limits, formats, terms for uploaded files, user model and price. Give each product a row, give each fact a column, record where you read it, and strike any product that fails a requirement. Three finalists go on to testing.
The table costs no money, and it stops you from spending a trial week on a product whose page limit or terms would have ruled it out on the first day.
Build the table with the people who touch client intake and the case management system. Ask them which discovery sets they convert each month and how long records retention says you may keep an upload. Client confidentiality belongs in a column you can defend to a partner, so quote the vendor's own sentence on who can see a file and when you can delete it.
What belongs in the columns of an OCR software comparison for a law firm?
The columns of a law firm's OCR software comparison are the facts a vendor publishes and your work depends on: accepted file types, size and page limits, export formats, terms for uploaded documents, who may use the account, where it runs, and how it is priced. Leave out anything that can only be judged by using the product.
Start from your requirements
Each column should trace to a line on your requirements list. If nobody at the firm needs spreadsheet output, a column for it adds noise. The post on writing a requirements list for OCR software for legal documents shows how to produce that list from a count of your own intake. With it in hand, the columns write themselves.
The column groups
| Group | Column | What to record |
|---|---|---|
| Intake | File types | The exact list the vendor gives |
| Intake | Largest file | The limit in MB |
| Intake | Longest document | Pages per file, by plan |
| Intake | Volume | Pages per month or documents per day, by plan |
| Output | Export formats | Every format, and which plan includes each |
| Client files | Training | The vendor's sentence on whether uploads train a model |
| Client files | Retention | How long files are kept, and who sets it |
| Client files | Deletion | Whether you can delete, and what is removed |
| Access | Users | Per person, per device, or unlimited on a plan |
| Access | Platform | Browser, Windows, Mac, mobile |
| Price | Model | Per page, per seat, flat rate, or a mix |
| Price | Billing terms | Monthly or annual, minimum term |
| Evidence | Source and date | The page you read and the day you read it |
Record words, not ticks
For the client-file columns, copy the vendor's sentence instead of writing "yes". "We do not train on customer content" and "We may use content to improve our services" would both earn a tick from a hurried reader, and they mean different things. A quoted sentence can be re-read by the partner who signs off.
Leave a column for "not stated"
When a vendor's public pages say nothing on a point, write "not stated" in the cell. Do not leave it blank and do not guess. A row with several "not stated" cells tells you something about the vendor, and it gives you a list of questions to send.
Where to find each comparison fact: pricing pages, terms of service and product documentation
Each comparison fact lives on one of four kinds of vendor page. Prices, limits and plan features are on the pricing page. Rights over uploaded files are in the terms of service and privacy policy. Formats, file types and platforms are in the product documentation or help center. Security statements sit on a security or trust page.
A map of fact to page
| Fact | Look first | Then check |
|---|---|---|
| Price and billing period | Pricing page | Terms of service, for renewal and minimum term |
| Page and document limits | Pricing page | Help center article on limits |
| File size limit | Help center or upload screen | Product documentation |
| Accepted file types | Product documentation | Upload screen |
| Export formats by plan | Pricing page | Product documentation |
| Training on uploads | Privacy policy, terms of service | Security page |
| Retention and deletion | Privacy policy | Help center article on deleting files |
| User model | Pricing page | License terms |
| Platform | Product documentation | System requirements page |
When two pages disagree
A pricing page may say one thing and a help article another. Record both statements, each with its source, and ask the vendor in writing which one governs. For anything about rights over your documents, read the terms of service and any agreement you would sign, and do not rest the cell on a sentence from a marketing page alone.
An example of reading one vendor
To show the method on a real set of pages, here is what Schema Law's public pages yield. The pricing page lists four plans: Free with 1 document a day and up to 3 pages per document, Solo with 1,000 pages a month and up to 250 pages per document, Team with 10,000 pages a month and up to 1,000 pages per document, and Enterprise with 100,000 pages a month and no fixed page limit per document. Solo adds email support, Team priority support, and Enterprise a security questionnaire, contract review and a named contact. The price of each plan is on the same page, and the post on OCR cost per page turns them into Schema Law plans as a per-page figure. The homepage states that uploads are a PDF, PNG or JPEG of up to 20 MB. The pricing page lists six lines under every plan, among them "Your documents are never used to train an AI model" and "Delete any document, at any time", along with all eight export formats, and the homepage's security commitments set each one out in a sentence or two.
That is one row filled in, with its sources. Every other row deserves the same treatment from the other vendor's pages, read on the day you build the table.
Date every cell
Vendors change prices and terms. Write the date beside each source. If the decision slips by a quarter, re-read the pages before relying on the table. For contract terms that matter beyond the first invoice, such as renewal and what happens to documents when you leave, the post on legal OCR subscription terms lists what to read.
Putting per-page, per-seat and flat-rate OCR quotes on one line of the table
Put different OCR pricing models on one line by converting each to the same figure: what a typical month would cost your office at your own page volume and headcount. A per-page price, a per-seat license and a flat plan only become comparable once all three are expressed as a monthly total for the same workload.
Fix the workload first
Write down three numbers before you touch a price:
- Pages you convert in an ordinary month.
- Pages in a heavy month.
- People who need to convert documents.
Use the same three numbers for every vendor. A comparison in which the workload shifts between rows compares nothing.
Convert each model
| Pricing model | How to reach a monthly total |
|---|---|
| Per page | Monthly pages multiplied by the page price |
| Per seat, monthly | Number of people multiplied by the seat price |
| Per seat, annual | Annual price multiplied by people, divided by 12 |
| Perpetual license | Purchase price divided by the months you expect to use it, plus any maintenance fee |
| Flat plan with a page allowance | The plan price, using the smallest plan whose allowance covers your ordinary month |
| Flat plan plus overage | The plan price plus the overage rate on the pages above the allowance |
Add two columns beside the monthly total: the total in your heavy month, and the limit you would hit first. A plan that is cheapest in an ordinary month can be the most expensive in a heavy one.
A worked example with invented vendors
Suppose a fictional office, Brandt Osei Law, converts 2,400 pages in an ordinary month and 6,000 in a heavy one, with four people uploading. Three invented vendors quote as follows.
| Vendor | Quote | Ordinary month | Heavy month | First limit hit |
|---|---|---|---|---|
| Vendor P | 4 cents a page | $96 | $240 | None stated |
| Vendor S | $30 per seat a month | $120 | $120 | 100 pages per file |
| Vendor F | $150 flat for 5,000 pages, 3 cents a page beyond | $150 | $180 | 5,000 pages a month |
The vendors and prices are made up. The table shows what the conversion reveals: Vendor P is cheapest in an ordinary month and most expensive in a heavy one, and Vendor S has a steady price but a per-file limit that may matter more than the price.
Do not copy another vendor's prices into a document you circulate without the date and the source page, and re-check them before you decide. The arithmetic for turning any plan into a cost per page, including staff time, is set out in the section on how to work out a real cost per page.
Which OCR marketing claims cannot be compared on paper?
Accuracy percentages, speed figures and words such as "advanced" or "AI-powered" cannot be compared on paper. Each vendor measures on its own pages by its own method, so two numbers that look alike describe different things. Keep these claims out of the table, or park them in a column marked "to test".
Why accuracy figures do not line up
An accuracy percentage depends on what was counted (characters or words), on which pages, and against what reference text. A vendor that tests on clean printed pages and a vendor that tests on mixed scans can publish figures that tell you nothing about each other. The post on how OCR accuracy is measured explains the measures and why a vendor's figure does not transfer to your documents.
Why speed figures do not line up
"Pages per minute" depends on the page, the file size, the hardware or the service load, and whether the clock includes upload and export. A figure measured on one-page letters says little about a 600-page production.
Claims to move to the test column
| Claim on the page | Why it cannot be compared | What to do with it |
|---|---|---|
| "99% accurate" | Unknown pages, unknown method | Measure on your own pages |
| "Reads handwriting" | No statement of which handwriting or how well | Test on your own handwritten pages |
| "Preserves formatting" | No statement of which formatting | Export a page and look |
| "Handles tables" | Could mean detection, or export, or neither | Export a ledger to a spreadsheet |
| "Fast" | No workload stated | Time your longest file |
| "Secure" | An adjective, not a term | Record the specific commitments instead |
What a vendor declining to claim tells you
A vendor that publishes no accuracy figure has given you nothing to compare, which is no worse than a figure you cannot interpret. Schema Law, for one, publishes none. Its homepage says recognition is strongest on clean print and weaker on faint faxes, stamps over text and handwriting. Treat any vendor's statement of its own limits as a fact for the table, and the rest as something to test.
How do you cut a long list of OCR software down to three finalists?
Cut a long list of OCR software to three finalists in passes: remove every product that fails a must-have, remove those with unanswered questions on client files, then rank the rest on your preferences and monthly cost. Stop at three. Each finalist has to be tested by hand on your own pages, so every extra name delays the decision.
The passes
- Gate on must-haves. Go down each column that maps to a must-have. A product that cannot take your largest file, lacks a format you need, or will not run on your computers is struck. Do not weigh it against a low price.
- Gate on client-file terms. Strike any product whose terms permit training on uploads if your firm's policy forbids it. For "not stated" cells, send the vendor the question in writing and set a date. No answer by that date is an answer.
- Rank on preferences. For the survivors, count how many of your preferences each meets.
- Rank on cost. Put the ordinary-month and heavy-month totals beside the preference count.
- Pick three. Take the best on preferences, the best on cost, and one more that differs in kind from the other two, such as an installed program beside two browser tools.
A worked elimination
Suppose the fictional Brandt Osei Law starts with seven products.
| Pass | Struck | Reason | Left |
|---|---|---|---|
| Must-haves | 2 | One caps files at 100 pages, one has no spreadsheet export | 5 |
| Client-file terms | 1 | Terms allow use of content to improve services | 4 |
| Unanswered questions | 1 | No reply on retention after ten days | 3 |
Three remain without a single trial having been run. If more than three remain after the gates, the ranking passes decide. If fewer than two remain, look again at whether every must-have is truly one.
Keep the struck rows
Do not delete eliminated products from the table. Grey them out and keep the reason. In six months someone will ask why a well-known product was not chosen, and the row answers the question.
A blank OCR software comparison table to copy
A blank OCR software comparison table has one row per product and one column per fact, with a source and date for each row. Copy this table into a spreadsheet, delete the columns your requirements do not call for, and mark each must-have column so a failing cell is obvious.
| Product 1 | Product 2 | Product 3 | Product 4 | |
|---|---|---|---|---|
| Intake | ||||
| File types accepted | ||||
| Largest file (MB) | ||||
| Pages per document | ||||
| Pages per month or documents per day | ||||
| Output | ||||
| Export formats | ||||
| Formats limited to higher plans | ||||
| Client files | ||||
| Training on uploads (quote) | ||||
| Retention (quote) | ||||
| Deletion (quote) | ||||
| Access | ||||
| User model | ||||
| Platform | ||||
| Price | ||||
| Pricing model | ||||
| Plan that fits our ordinary month | ||||
| Ordinary month total | ||||
| Heavy month total | ||||
| Billing period and minimum term | ||||
| To test | ||||
| Claims we could not compare | ||||
| Evidence | ||||
| Pages read | ||||
| Date read | ||||
| Questions sent, and date | ||||
| Status | ||||
| Finalist, struck, or waiting | ||||
| Reason |
Filling it in
- One person fills in the whole table, so the same reading is applied to every vendor.
- Cells hold the vendor's words or numbers. Opinions go in the "Reason" row only.
- "Not stated" is a valid entry. An empty cell is not.
- A second person spot-checks the client-file rows against the source pages before the table is used to decide anything.
What happens next
The table produces finalists, not a winner. Nothing in it says how any product reads your pages. Take the three that survive to a hands-on trial, as the guide to testing OCR software on your own documents sets out, and add the results as new rows under "To test".