Relativity aiR vs Everlaw vs DISCO: Real AI eDiscovery Cost per 1 TB Matter in 2026
The short answer: eDiscovery pricing is a stack of separate meters, and the one that grabs attention, AI review, is rarely the largest. Using a published service-provider rate card for RelativityOne, a 1 TB matter with £50-per-GB processing, twelve months of active hosting and AI review of a million documents costs about £210,000 in year one, or £210 per collected GB and £0.21 per document, and AI review is £76,000 of that. Everlaw and DISCO do not publish rates; both use per-gigabyte models that bundle processing and users, and third-party survey data puts typical hosting at $10–$20 per GB per month. The pricing question that matters most is not the AI fee. It is what a gigabyte means (collected or processed), which storage tier it sits in, and how much human review the AI genuinely removes.
What each vendor bills
RelativityOne (REPORTED, a UK government G-Cloud rate card from a professional-services provider, January 2026, priced in pounds). Processing is a one-off £50 per GB ingested; OCR and de-duplication are free. Hosting is £7.00 per GB per month for review, £3.50 for repository storage and £1.50 for cold storage. User licenses were listed as free of charge. aiR for Review is billed per document request: the first 50,000 requests are free and each further request is £0.08. Analytics and audio/video transcription were free. Restoring an archived workspace was £1,500. Another cost tracker (REPORTED) cited processing at $35 per GB, storage of $1–$4 per GB per month in lower tiers, user licenses of $75 per user per month and a warning that aiR requests can be 100–200% higher than expected because long documents are split by token limits. Its scenarios implied committed hosting of $20–$25 per GB per month, well above the rate card. The two sources disagree, and rate cards from service providers may not match direct Relativity contracts.
Everlaw (REPORTED for ranges; packaging per its published pages). Everlaw's pages describe a data-and-usage subscription with unlimited users, with legal holds, processing, predictive coding, analytics, unlimited productions and many single-document AI actions included in the per-GB rate. Batch AI actions require purchased credits that expire at the end of the term. Exact rates, platform minimums and credit prices are not published (QUOTE-ONLY / UNKNOWN); third-party reports cited a platform minimum of a few thousand dollars a month and roughly $15–$35 per GB, none of which is an official list price.
DISCO (REPORTED). A February 2026 announcement was reported as moving to an all-inclusive per-GB price assessed on processed data, bundling ingestion, users and the Cecilia AI assistant. Billing uses a high-water mark of hosted data in each month, advance billing with true-ups, and a highest-applicable-rate rule when data sits in more than one review state during a month. A competitor's comparison page states that DISCO's Auto Review is billed per document while chat and summaries are bundled, which conflicts with descriptions of fully bundled AI. Dollar rates were not publicly listed (QUOTE-ONLY / UNKNOWN).
Market benchmark (REPORTED). A 2026 pricing survey found 54.7% of respondents reporting basic hosting below $10 per GB per month and 30.2% between $10 and $20. Processed data was reported to be 1.5 to 3 times larger than collected data.
Model: three matters
The matters are ILLUSTRATIVE: 100 GB and 100,000 documents; 1 TB (1,000 GB) and 1 million documents; 10 TB and 10 million documents, with active review for 12 months and AI first-pass review of every document.
RelativityOne rate-card model (£): processing = GB × £50; hosting = GB × £7 × 12; aiR = (document requests − 50,000) × £0.08.
| Matter | Processing | Hosting, 12 months | aiR for Review | Year-one total | Per GB | Per document | Cold storage per year after |
|---|---|---|---|---|---|---|---|
| 100 GB / 100,000 docs | £5,000 | £8,400 | £4,000 | £17,400 | £174 | £0.174 | £1,800 |
| 1 TB / 1 million docs | £50,000 | £84,000 | £76,000 | £210,000 | £210 | £0.210 | £18,000 |
| 10 TB / 10 million docs | £500,000 | £840,000 | £796,000 | £2,136,000 | £213.60 | £0.214 | £180,000 |
Per gigabyte, the total converges near £210 because processing and hosting scale linearly and the free 50,000 aiR requests matter only for the smallest matter. Year-one cost is not steady-state: after review ends, moving data to cold storage lowers the annual cost to £1.50 per GB per month (£18 per GB a year), a 79% reduction from active hosting.
Per-GB platforms without published rates (Everlaw, DISCO) at survey bands ($, not a quote): cost = collected GB × expansion × $ per GB per month × 12.
| Matter | $10 per GB per month | $20 per GB per month | Same, if hosting bills processed data at 2x |
|---|---|---|---|
| 100 GB | $12,000 | $24,000 | $24,000–$48,000 |
| 1 TB | $120,000 | $240,000 | $240,000–$480,000 |
| 10 TB | $1.2 million | $2.4 million | $2.4–$4.8 million |
Those bands are market benchmarks, not Everlaw or DISCO rates, and they exclude batch AI credits at Everlaw and any per-document Auto Review charge at DISCO. Pounds and dollars are not converted here.
Cost per document reviewed
Formula: total cost ÷ documents. The Relativity model is £0.17–£0.21 per document including AI. For $10–$20 per GB per month at 1,000 documents per GB, the per-gigabyte benchmark works out to $0.12–$0.24 per document at 1.0x and $0.24–$0.48 at 2x processed-data expansion. Against a human first-pass review at 50 documents per hour and £50 an hour, which is £1.00 per document (ILLUSTRATIVE), the platform cost is 17–21% of manual review cost.
Break-even: reviewer hours saved
Fee break-even. aiR costs £0.08 per document request against £1.00 of human cost per document, so it breaks even if it removes 8% of documents from human review, or 16% if long documents double the number of requests. For 1 million documents that is 1,520 reviewer hours (152,000 documents at 50 an hour, for £76,000) or 3,040 hours with doubled requests.
Saving is not one-for-one. The scenario below separates first-pass savings from legally required human validation. If humans still review a share v of documents after AI, hours saved are (1 − v) × documents ÷ 50. For 1 million documents, manual review is 20,000 hours (£1,000,000):
| Share still reviewed by humans | Hours saved | Value at £50 | Net of aiR fee (£76,000) |
|---|---|---|---|
| 35% | 13,000 | £650,000 | £574,000 |
| 50% | 10,000 | £500,000 | £424,000 |
| 75% | 5,000 | £250,000 | £174,000 |
AI stops paying for itself only if humans still review more than about 92% of documents (84% with doubled requests). Nothing here assumes AI replaces attorney review one-for-one: privilege review, validation sampling and second-level review remain human, and the 35–75% range is ILLUSTRATIVE. Hosting and processing costs are incurred whether or not AI is used, so they are not part of the AI break-even.
Sensitivity
- Collected versus processed GB. If hosting bills processed data at 2x, Relativity hosting doubles to £168,000 at 1 TB and year one rises to £294,000 (with aiR unchanged) or £374,000 (with doubled aiR requests).
- aiR request inflation. Doubling requests takes 1 TB from £76,000 to £156,000.
- Storage tier discipline. Cold storage at £1.50 versus review hosting at £7 is a 79% difference per GB per month.
- Human validation share. Moving from 35% to 75% cuts net AI savings by 70%.
Budgeting traps
- Highest-rate rule. Data in more than one review state in a month may be billed at the highest rate for the whole month.
- High-water mark. Peak hosted volume sets the invoice, so mid-month uploads matter even if data is later deleted.
- Processing multipliers. Processed volume can be up to three times collected volume.
- Credit expiry. Everlaw's batch AI credits expire at the end of the term.
- Seat fees. A Relativity rate card listed free licenses while another source listed $75 per user per month; confirm which applies.
- Long-term retention. A matter that stays in active hosting for years costs far more than the same data archived.
What to ask before you buy
Ask each vendor whether hosting applies to collected or processed data, what tiers exist and when data moves between them, how AI is metered per document and per request, and what happens to unused credits. Then model year one and the retention years separately.