[Legal](https://bfcs.ai/industry/legal/) · [Cheaper](https://bfcs.ai/cheaper/)

# Can AI Reduce Discovery Costs?

Discovery is the rare case where AI has already won on cost, quietly, and with the courts' blessing, for over a decade. The cautionary tale sits one desk away: the same firms saving millions on review are being sanctioned for AI that invents case law.

Updated **July 2026** · 3 min read · Evidence-graded · 6 sources

## The BFCS verdict

Every article scored on the same four dimensions

| Dimension | Verdict | Why |
| --- | --- | --- |
| Better | Partly · Moderate confidence | More consistent culling; generative drafting less so · Technology-assisted review is often more consistent than fatigued manual review. Generative tools layered on top can inject error, so the two must not be conflated. |
| Faster | Yes · High confidence | Millions of documents culled to thousands · First-pass classification compresses review timelines dramatically and is standard practice in large matters. |
| Cheaper | Yes · High confidence | The most proven cost lever in law · NIST-linked research shows review productivity gains of ~80%; predictive coding is mature, court-accepted, and demonstrably cheaper than exhaustive manual review. |
| Safer | Conditional · Governance-gated | Privilege leakage and hallucination draw sanctions · The generative wave next door carries real penalties. The verification duty stays with the lawyer. A citation the signer has not read does not go in. |

## The win that already happened

Long before generative AI, litigation had a data problem: matters routinely turn on millions of documents, and reviewing each by hand costs a fortune. Technology-assisted review (predictive coding) solved much of it. A reviewer tags a training set; the model then ranks the rest, culling collections from millions to thousands. NIST-linked research puts the productivity gain at roughly 80%, and the landmark Grossman and Cormack study found TAR could be _more_ effective and more efficient than exhaustive manual review. (Source 1), (Source 2)

The uncomfortable finding underneath: manual review was never the gold standard. In a classic study, skilled reviewers were convinced they had found 75% of relevant documents when they had actually found 20%. (Source 3) Against that baseline, a well-run TAR process is cheaper _and_ more thorough, and courts have accepted it for years where the process is properly documented.

### The evidence, in four numbers

~80% · gain in review productivity · predictive coding vs manual review (NIST-linked research)

20% vs 75% · relevant docs found vs believed · manual reviewers, landmark Blair & Maron study; manual is not the gold standard

millions → thousands · document set after TAR culling · typical reduction in large matters

$145k · US AI-hallucination sanctions, Q1 2026 · fabricated citations in filings, the governance counterweight

> Classifying documents you already have is a solved, court-tested problem. Generating documents you don't is where the sanctions begin. The word "AI" hides the difference.

## Two very different AIs

The critical distinction, endlessly blurred by the single word "AI": predictive coding _classifies documents that exist_, while generative models _produce text that may not_. The first is mature and defensible. The second is drawing a wave of penalties: at least US$145,000 in sanctions for fabricated citations in the first quarter of 2026 alone, and in one Nebraska appeal, 57 of 63 citations were defective, ending in the first indefinite suspension over AI filings. (Source 4)

## Where governance decides the outcome

Discovery's specific governance burdens are privilege and defensibility: an AI process that inadvertently surfaces or leaks privileged material, or that cannot be explained to a court, forfeits the cost saving many times over. The signing lawyer remains accountable for every citation and production choice; as one appellate panel put it, a citation the lawyer has not personally read does not go in, whatever produced it. (Source 5) The ABA's 2024 guidance is equally blunt: competence, candour and confidentiality do not transfer to a tool. (Source 6)

## How to think about adoption

Use mature, documented TAR for review with confidence; it is the clearest cost win AI offers any profession. Keep generative drafting on a much shorter leash: verify every citation, wall off privileged and client data, and record the process so it is defensible. Treat "AI" not as one capability but as two with opposite risk profiles, and govern each on its own terms.

## How we read this evidence

BFCS.ai does not fabricate figures. Technology-assisted review and generative drafting are distinct technologies with distinct evidence; this article keeps them separate deliberately.

## Sources

1.  NIST-linked research (via industry reporting):
    
    **What we took from it:** predictive coding can increase document-review productivity by as much as ~80%.
    
2.  Grossman, M. R. & Cormack, G. V., “Technology-Assisted Review in E-Discovery Can Be More Effective and More Efficient Than Exhaustive Manual Review,” Richmond Journal of Law & Technology 17 (2011), the landmark peer-reviewed comparison; TAR is routinely court-accepted where the process is documented.
    
3.  Blair & Maron (1985):
    
    **What we took from it:** skilled reviewers believed they had retrieved ~75% of relevant documents when they had in fact found only ~20%, evidence that exhaustive manual review is not the gold standard it is assumed to be.
    
4.  ComplexDiscovery / GC AI sanctions tracking (2026):
    
    **What we took from it:** US courts imposed at least US$145,000 in sanctions for AI-generated fake citations in Q1 2026 alone; in one Nebraska appeal 57 of 63 citations were defective (20 hallucinated), leading to the first indefinite suspension over AI filings.
    
5.  Whiting v.
    
    **What we took from it:** City of Athens: a US$15,000 federal appellate sanction; the panel held that no filing should contain any citation, from AI or any other source, that the signing lawyer has not personally read and verified.
    
6.  ABA Formal Opinion 512 (July 2024):
    
    **What we took from it:** lawyers using generative AI retain their duties of competence, candour and confidentiality, and the obligation to verify the work remains with the lawyer.
    

Putting this to work

## Working through this in your own organisation?

BFCS.ai is published by Inhouse CX. We help small and mid-sized organisations decide where AI earns its place, and put the governance in place that makes it safe to use.

[Talk to Inhouse CX →](https://inhouse.cx/engage/)

Continue the thread

- [Can AI Make Contract Review Faster?](https://bfcs.ai/articles/can-ai-make-contract-review-faster/): More in Legal · LEGAL · FASTER · 7 min · Read →
- [Can AI Keep Client Data Safer?](https://bfcs.ai/articles/can-ai-keep-client-data-safer/): More in Legal · LEGAL · SAFER · 7 min · Read →
- [Can AI Make Small Firms More Competitive?](https://bfcs.ai/articles/can-ai-make-small-firms-more-competitive/): Also Cheaper · SMALL BUSINESS · CHEAPER · 6 min · Read →

---

Canonical page: https://bfcs.ai/articles/can-ai-reduce-discovery-costs/
This is the Markdown version of the page, generated from its HTML at build time.
Site guide for agents: https://bfcs.ai/llms.txt · Every page: https://bfcs.ai/sitemap.xml