Small firms pay for e-discovery software per GB a month, per case size or per user. Published prices are rare and modest, such as $27 a month for a 3 GB case and $76 for 10 GB; most vendors quote. AI review cuts reading time but can misjudge short replies, poor scans and privilege, so validate it by sampling.
Hosting is rarely the big number. In the worked example below, five months of hosting costs $380 while reviewer time costs nearly $20,000, so the software choice matters mainly through the review hours it removes and the meters that keep running after a matter closes. Everything here is arithmetic on published list prices or on stated assumptions, and since most vendors publish no price at all, you will need quotes for the rest.
Six pricing models, and who each one suits
The vendor pages read in November 2026 sort into six ways of charging. The same matter can cost very different amounts under each, which is why a per-GB quote and a per-user quote cannot be compared until you know your data volume and how many people will review.
| Model | How you are billed | What the vendor's page says | Suits |
|---|---|---|---|
| Case-size plan | A flat monthly fee for the case's maximum volume, whatever you actually store | GoldFynch: a 3 GB case is $27 a month and a 10 GB case $76; unlimited users, processing and productions; only the case owner pays | One-off matters whose data you can estimate |
| Per GB, pay as you go | Monthly, no commitment, on data stored or processed | Logikcull: priced on data stored, with a 10 GB minimum threshold and unlimited users. DISCO: a per-GB structure billed on processed data, no prices published | Irregular matters of unknown size |
| Subscription with client bill-back | Annual plan with reusable storage, or a set number of matters | Logikcull: storage-based plans include reusable blocks of storage; matter-based plans charge once for uploaded data; most subscribers bill clients per GB per month | A steady flow of matters |
| Per user, hosting included | A flat monthly rate per user | Nextpoint: Essential, Advanced and Apex plans, unlimited cases and free hosting, no rate published | A small team running several matters at once |
| Committed volume, AI included | Pay as you go, or a one- or three-year commitment on data volume | Relativity: the AI tools come with committed volume, with no separate AI licences and no per-user fees; price on request | Firms with large, AI-heavy matters |
| Quote only | Demo, then a quote | Everlaw: unlimited user licences listed, no prices or tiers published | Mid-size firms comfortable negotiating |
Two details decide how those models behave on a real matter. First, what counts as volume. GoldFynch counts an uploaded ZIP or PST file and its extracted contents, while DISCO says it bills on processed data. Second, who pays. On GoldFynch only the case owner is billed and sharing is free, which matters when a paralegal creates the case and then leaves.
Meters beyond the headline price: AI, collection and services
The headline plan is only one meter. Several vendors charge separately for the AI, for collection or for services, and the pages do not always make clear which plan includes what.
- AI review can be a separate per-document charge. DISCO says its Cecilia AI (question answering, summaries, timelines) carries no extra charge, but its Auto Review product, which suggests tags with explanations, is priced per document on top of the per-GB rate. No figure is published.
- AI can run on credits. Everlaw's feature list marks coding suggestions, its natural-language question tool, machine translation and custom extractions as using credits, and does not say how many credits a plan includes. Coding suggestions, translation and extractions use no credits when run on a single document, so the meter starts when you run them across a set.
- AI can be an add-on. Reveal describes its Ask feature as an add-on, and Logikcull's comparison table puts add-on markings against Reveal AI fact-finding. Ask which tier includes it and what it costs.
- AI can be bundled. Relativity says its AI tools come with the committed data volume, with no per-user fees.
- Collection sits in higher tiers. Logikcull describes its Corporate Suite as the plan that adds data collection from 31 or more sources and legal holds. Everlaw lists cloud connectors for Microsoft 365, Zoom and Google Workspace on its platform. On a lower tier, expect to collect by export, which is your time.
- Services are usually unpriced. DISCO lists four support levels, from self-service to a managed service, without prices. Nextpoint offers forensic collection, managed review and production support on demand, also unpriced.
- Promotions end. Logikcull's pricing page shows promotion pricing until 31 December 2026. Ask what the same plan costs after that date, and what it costs at renewal.
The Complete Excel Course with AI Integration
From formulas and pivot tables to Power Query, DAX, VBA and AI with Copilot, ChatGPT and Claude — hands-on from the first lesson.
By Sagnik Bhattacharya and Virginia Thorn, Primetime Emmy Award winner
Buy nowCosting a commercial dispute from your own data
Take a four-lawyer firm acting in a commercial dispute: the client says a contractor finished a shop fit-out 11 weeks late, and the contractor blames missing sign-offs. The firm collects from three mailboxes and one shared folder, using targeted date ranges to keep the upload to about 4 GB of PST and ZIP files. Extraction takes the case to roughly 9 GB once containers and contents are both counted, which fits GoldFynch's 10 GB plan. Everything below is illustrative.
Culling happens inside the platform, and each step removes a different kind of waste.
| Stage | Items left | What removed them |
|---|---|---|
| Collected from three mailboxes and a shared folder | 62,000 | Nothing yet |
| After removing exact duplicates | 41,000 | The same email sitting in several mailboxes |
| After the date range (contract signature to two months after handover) | 33,500 | Items before and after the dispute period |
| After removing newsletters, calendar notices and system files | 24,000 | An exclusion list the supervising lawyer approved |
The 24,000 items are what people may need to read. Now compare reading all of them with reading a ranked subset, using two stated assumptions: a reviewer reads 55 documents an hour, and an hour of reviewer time costs the firm $45. Both are placeholders, so time a sample of your own before trusting the result.
| Approach | Documents read | Hours at 55 an hour | Cost at $45 an hour |
|---|---|---|---|
| Read everything | 24,000 | 436 | $19,620 |
| AI-ranked: read the top 9,000, plus a 1,500-document validation sample | 10,500 | 191 | $8,595 |
| Difference | 13,500 fewer | 245 | $11,025 |
Three conclusions follow. Hosting the 9 GB case for five months at $76 comes to $380, about 1.9% of the read-everything cost, so hosting price should not drive the decision. The AI-assisted route saves $11,025 of time, which means any AI charge above about 46 cents a document across the 24,000 (that is $11,025 divided by 24,000) makes it more expensive than reading everything on these assumptions. And the saving exists only if you can show that the 13,500 unread documents were safe to leave, which is what the validation section below is for. Privilege checks on the responsive set cost the same either way and are left out.
Where the hosting meter keeps running after the matter ends
Small meters turn into real money when nobody watches them. Two examples, both driven by rules on GoldFynch's own pages.
The first is container files. GoldFynch's example is a 2 GB ZIP that expands to 2.5 GB on processing: the case volume becomes 2 + 2.5 = 4.5 GB, so the case moves up to the 10 GB plan at $76 a month. Delete the ZIP after processing and the case holds 2.5 GB, which fits the 3 GB plan at $27. The difference is $49 a month, and the page itself says it does not recommend deleting the originals, so the better habit is to estimate expanded volume before you choose a plan.
The second is a case left open. Billing runs on the plan, not on use, and the only way to stop it is to delete the case. A 10 GB case forgotten for eight months after settlement costs $608 for nothing. A closing checklist stops that:
- Export the production set, the load files and your coding decisions, and confirm you can open them without the platform.
- Ask the client in writing how long the data must be kept and where.
- Record who owns the case, since only the owner is billed.
- Delete the case, then check that the next invoice drops it.
A review workflow for a four-lawyer firm, stage by stage
The software handles stages three to six. The first two, and the last, decide whether the output can be defended. These planning allowances are for a matter of the size above and should be replaced with your own measurements after the first one.
- Preserve and scope (days 1 to 2). A lawyer writes the hold notice, lists custodians, sources and the date range, and agrees the exclusion list with the client. Allow two to four hours of lawyer time.
- Collect (days 2 to 5). Take targeted exports from the client's systems, keep the originals untouched and write down who collected what, when and how. Never forward emails to a personal account to "gather" them.
- Process and cull (days 5 to 7). Upload, let the platform extract and run OCR, then apply de-duplication, date and custodian filters. Spot-check what the filters removed before you rely on them.
- First-pass review (weeks 2 to 3). Either read linearly or let ranking and AI sort the set. A lawyer writes the instructions and a second person checks a sample of calls each day.
- Privilege and confidentiality pass (weeks 3 to 4). A lawyer works through the responsive set, redacts and logs. AI can raise candidates, but the decision stays with the lawyer.
- Production (week 4). Produce in the format agreed with the other side, check redactions on a sample, and keep a copy of exactly what went out.
- Close-out (week 5 onwards). Archive or delete as the client instructed, and stop the billing meter.
One culling trap is worth testing for. Imagine a keyword search for "delay" that returns 31,400 hits because it also matches a boilerplate footer about delayed parcel deliveries. Adding the date range and a second term (handover or completion) might bring it down to 9,100. The numbers here are made up; the habit is real. Test every search term against a sample of 100 hits before you cull on it, and keep a note of the final terms and the count each one returned, because you may have to explain them later.
Where AI review helps and where it falls short
"AI review" covers two different things, and vendors use the label for both. Predictive coding, or active learning, watches your reviewers' decisions and ranks everything unread by likely relevance. Generative review reads each document against instructions you write and returns a call with a reason. Everlaw lists predictive coding and a natural-language question tool. Relativity and DISCO describe generative review with explanations.
It helps in these ways, according to the vendors' own descriptions:
- Ordering the reading. People see the most promising documents first, which is where the time saving in the worked example comes from.
- First-pass calls with reasons. Relativity's documentation says aiR Review returns a recommendation (relevant, not relevant or borderline), citations with a rationale and notes on possible prediction errors. DISCO describes tag suggestions with full explanations.
- Questions and summaries. Everlaw's question tool, DISCO's Cecilia and Relativity's Assist let you question a set in plain language, useful for early case assessment before a single document is coded.
- Translation. Everlaw lists machine translation in more than 100 languages, free on a single document and billed in credits across a set.
It falls short in these ways, some stated by the vendors and some my reading of how the tools work:
- One document at a time. Relativity's documentation says each document is reviewed separately. A short reply such as "agreed, go ahead" means nothing without the thread above it, so review at thread level or give the tool the context.
- Extracted text only. The input is the extracted text. A badly scanned page or a photographed handwritten note arrives as noise, and the call on it is noise too.
- Language and informal text. The same documentation says the model was evaluated for 83 languages but primarily tested on English, and was not specifically tested on emojis. Chat and text-message exports are where that bites.
- Privilege. Vendors sell privilege tools separately or alongside, and a lawyer still decides. Treat any "not privileged" call as a suggestion.
- Instruction quality. Vague issue definitions produce inconsistent calls, exactly as they do with human reviewers.
- Cost per document or per credit. It may cost more than the time it saves on a small set, as the 46-cent break-even showed.
Writing review instructions an AI can follow: a filled-in example
Whatever the tool calls it (prompt criteria, review instructions, a protocol), the quality of a generative review rests on a page of plain writing. This version is for the fit-out dispute above. Adapt the headings to your matter and test it on 200 documents you have already coded yourself before running the whole set.
MATTER: The claimant says the contractor handed over a shop fit-out 11 weeks late.
The contractor says the delay was caused by missing client sign-offs.
ISSUES
1. Agreed or revised completion and handover dates
2. Requests for sign-off or approval, and the replies
3. Causes of delay, including supplier problems
4. Payments withheld or released because of delay
RELEVANT: the document addresses any issue above. Read the whole thread,
including quoted text lower down.
NOT RELEVANT: newsletters, calendar invitations with no discussion, personal
messages, invoices for other projects.
BORDERLINE: the project is mentioned only in passing. A lawyer will look.
PRIVILEGE CHECK: the document is to or from a lawyer, or mentions legal advice.
Use this label whatever the relevance.
NEEDS HUMAN: the text is missing, garbled or in a language other than English.
RETURN FOR EACH DOCUMENT: the label, the issue number(s), one sentence of
reasoning, and the exact words from the document that support it.Here is how the first results might look. They are invented to show what to check, not output from any real product.
| Document | Call returned | What you would do |
|---|---|---|
| Site manager asks whether "the 14 March handover is still on" | RELEVANT, issue 1, quotes the date | Accept. The quote can be checked against the text. |
| Reply in the same thread: "Agreed, go ahead" | NOT RELEVANT, no content | Wrong in isolation. Add the "read the whole thread" line, or review at thread level. |
| Photographed site note, handwriting | RELEVANT, issue 3, no quote available | Reject until a person has read it. The text came from OCR of handwriting. |
| Email from the client to its lawyer about the contractor's claim | PRIVILEGE CHECK | Send to the lawyer. Never produce on the machine's say-so. |
| Supplier email in another language | NEEDS HUMAN | Route to translation, then recode. |
The second row is the instructive one. A call can look confident and be wrong because the model saw one short reply with no thread. Fixing the instruction, not the individual call, is what improves the next batch. Two or three rounds of this on a 200-document sample, comparing the tool's calls with your own, is the "training" a generative tool needs.
Proving it worked: an elusion sample you can run in an afternoon
Skipping documents is only defensible if you can show how few relevant ones were left among them. The standard check is an elusion test: draw a random sample from the documents nobody read (the "null set"), have a lawyer code it, and see how many relevant ones turn up.
Take the earlier matter and a bigger collection to show the arithmetic: 120,000 documents, of which the AI and reviewers put 18,000 forward. Reviewers confirmed 15,300 as relevant, so precision was 85% (15,300 of 18,000). That leaves a null set of 102,000. A lawyer codes a random sample of 1,500 of those and finds 9 relevant.
| Step | Sum | Result |
|---|---|---|
| Elusion rate | 9 / 1,500 | 0.6% |
| Relevant documents estimated in the null set | 0.6% x 102,000 | about 612 |
| Estimated recall | 15,300 / (15,300 + 612) | about 96% |
| Plausible range (95% confidence) | 0.32% to 1.14% elusion | about 322 to 1,159 missed; recall 93% to 98% |
Report the range, not the single figure. Nine hits in 1,500 is a small number, and a different random sample of the same set could plausibly have found as few as five or as many as seventeen. Then apply a rule you agreed before you began: for example, if the top of the range drops recall below the figure your supervising lawyer set, review a further slice of the next-ranked documents and test again. The right target depends on the matter, the other side and the court, so take advice rather than borrowing a number from this tutorial. Record the sample size, the date, who coded it and the result in the matter file.
A trial set that tests each platform on your worst documents
Demos use clean data. Both GoldFynch (a free case limited to 512 MB uploaded and 90 days, with watermarked productions) and Logikcull (a sample matter on pay-as-you-go accounts) let you try a platform without a quote. Build one test set of 1,000 documents from an old, closed matter, with the answers already known, and load the same set everywhere.
| Group in the test set | Documents | What it tests |
|---|---|---|
| Clearly relevant emails with attachments | 150 | Baseline recall for search and ranking |
| Exact and near duplicates | 120 | Whether de-duplication and threading work |
| Scanned PDFs, including 10 handwritten notes | 60 | OCR quality: can you find a known phrase? |
| Spreadsheets with budgets and snag lists | 50 | How numbers and sheets are shown and searched |
| Chat or text exports with emojis and one-word replies | 100 | Informal language and missing context |
| Emails to or from a lawyer | 40 | How many privilege candidates are flagged |
| Non-English emails | 30 | Language handling and translation |
| Noise: newsletters, calendar notices | 450 | Precision, and how much junk reaches a person |
Score each platform on the same five numbers: time from upload to first search, how many of the 10 handwritten notes and 50 scanned pages the search finds, how many of the 40 privileged emails are flagged, the AI's precision and recall against your own coding, and the minutes taken to produce a set. If you are testing AI in several tools at once, the scoring approach in scoring accuracy, privacy and cost in a short trial adapts well. For the supplier questions that sit around it (where data is hosted, who can see it, what happens on exit), use the data, supervision and supplier checks before any real client data goes in.
Choosing a low-cost host, an AI platform or a service provider
Match the route to how often you face document-heavy matters, not to the most impressive demo.
| Your situation | Route | Why |
|---|---|---|
| One or two small matters a year, under about 10 GB | Case-size plan or pay as you go, after a free trial case | Cost tracks the matter, and the meter stops when you delete the case |
| Several matters a year that you bill back | Subscription with a per-GB bill-back rate | Predictable cost and a simple line on the client invoice |
| A few large matters a year with heavy email | An AI platform on committed volume, or a service provider for the matter | The AI and analytics earn their cost only at scale |
| A matter of a few thousand documents | Search in your case system, or pay a provider once | A new platform costs more in set-up time than it saves |
Whichever route you pick, agree how the technology is billed to the client before data goes in. Logikcull's pricing FAQ says most subscribers bill clients per GB per month, and the same practice works with any hosting plan: state the hosting rate, the review rate and the point at which hosting stops in the engagement letter, and have a qualified adviser check the wording. Many firms also ask whether a general assistant could do part of this job on a business plan; the tutorial on ChatGPT Business for law firms covers the seats, privacy settings and cost, and what to test in AI paralegal software covers the tools that sit between a chat assistant and a full platform.
Before committing to anything with a quote attached, put the same questions to every vendor: the unit of volume, whether AI is included or metered, what happens to the price at renewal, what you can export and when. Most of the questions a small firm should put to CoCounsel before buying carry over to e-discovery vendors, and the breakdown of per-seat and pay-as-you-go legal AI pricing shows how to compare the two billing styles side by side.
E-discovery questions small firms ask
Is e-discovery software worth it for a small firm?
Usually when a matter holds more documents than one lawyer can read in a week. Hosting is a small cost beside review time, so the saving comes from culling and search. For a few thousand documents, a service provider for that one matter, or the search in your case system, may be enough.
Can ChatGPT or Claude replace e-discovery software?
No. General assistants do not process mailbox exports, de-duplicate, keep an audit trail, redact or produce documents. They can help draft review instructions or summarise a few documents, but only on a plan that does not train on your data and with client consent where your duties require it.
What is the difference between TAR and generative AI review?
Technology-assisted review learns from your reviewers' coding decisions and ranks the remaining documents. Generative review follows instructions you write and returns a call and a short reason for each document. The first needs good seed decisions, the second needs good instructions, and both need validating by sampling.
How big a sample do I need to check AI review?
About 1,500 documents drawn at random from the set nobody read gives a margin of roughly 2.5 percentage points at 95% confidence in the worst case. When few documents are relevant, as in a well-culled set, the range is tighter. Agree the target with your supervising lawyer first.
Further reads
- Generative AI for Law Firms: Which Matters Suit It and Which Do Not — Which matter types suit generative AI and which do not.
- AI Contract Review for Small Firms: What It Catches and Misses — What AI contract review catches and misses.
- Best AI Platform for Law Firms: Clio, Smokeball and Lawmatics Compared — Practice platforms with AI built in, compared.
- Law Firm AI Usage: How to Audit What Your Fee Earners Really Use — Find out what fee earners really use.
- AI Implementation Plan for a Small Law Firm: The First 90 Days — A 90-day plan for introducing AI safely.
- Conflict Check Software for Small Law Firms: Where AI Helps — Another matter-opening tool where AI helps.
- AI Discovery Review for Defence Firms: Index, Search and Cross-Check — Index discovery, map each statement to the disputed issues, build a contradiction table with page cites, then verify every cite by hand before filing.
- Family Law AI Software: What a Small Practice Should Check Before Buying — A scored checklist for a small family practice buying AI software: confidentiality, data location, conduct rules, citation checks and exit terms.
- AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.
Sources: vendor pricing and help pages for GoldFynch, Logikcull (Reveal), DISCO, Everlaw, Nextpoint, Reveal and Relativity, read in November 2026, including Relativity's aiR Review documentation.