Cheaper Models Are Coming for E-Discovery's AI Bill — But Not for Every Task
In an August 11, 2026 guest column on LawSites, Ken Crutchfield — founder and CEO of Spring Forward Consulting, a longtime contributor to the site — argues that "open-weight" AI models are about to become a real budget line for legal tech buyers planning for 2027. His case rests on three data points: OpenAI has released open-weight models that run on a personal computer and support distillation via API; Thomson Reuters has built its own language model rather than renting one; and Crutchfield says he spoke with a legal tech vendor that now writes its software using an open-weight model run entirely on local machines, paying no per-token fees at all. His broader argument is that legal work carries a 100% accuracy expectation, and that standard doesn't require the newest frontier model for every task — only for the hardest ones.
Should e-discovery buyers expect cheaper AI processing in 2027?
For narrow, repetitive tasks — not for review of the full evidentiary record — yes, and vendors should be able to explain which is which.
Crutchfield's point is that clause labeling, term extraction, and similar pattern-matching work were already solved by pre-ChatGPT models like Google's BERT, well before frontier LLMs existed. If a vendor is still routing that kind of work through an expensive, metered frontier model in 2027, that's a vendor choice, not a technical necessity — and it's a choice that shows up in your invoice.
Which e-discovery tasks can actually move to a smaller, local model?
Tagging, metadata extraction and issue-coding of routine documents are candidates; complex chronology-building and legal reasoning likely still need a frontier model.
Crutchfield draws the same line: "complicated legal problems and deep research-style tasks will continue to require comprehensive foundation models," while narrower jobs don't. For an evidence platform that has to issue-code every document in a collection and then answer plain-language questions against the full record with citations, the coding pass and the reasoning pass are different jobs with different accuracy stakes — and arguably different model requirements.
What should you ask your own vendor about model choice?
Ask which specific tasks run on which model, who verified the accuracy claim, and whether switching models changes where your data goes.
This is the gap in the vendor-numbers conversation generally: "good enough" is not a number anyone has published, it's a judgment call the vendor is making on your behalf. If a platform is single-tenant or runs inside your own cloud account under your own keys — as document review tools increasingly are — ask explicitly whether a move to a local or open-weight model happens inside that same boundary, or introduces a new one. Distillation and open licensing are not the same thing as "your privileged data never left your environment," and the column doesn't address custody at all.
Does the 100% accuracy standard actually change with a cheaper model?
No — the standard stays the same; only the cost and location of hitting it changes.
Crutchfield poses this as an open question rather than an answer: "how much should an organization pay for greater model accuracy?" That's a fair question for a budget memo, but for a court-ready record, the honest follow-up is who is measuring accuracy against what benchmark before a cheaper model gets anywhere near a production coding run.
Frequently asked questions
Is this the same DeepSeek story from last year?
It's the backdrop, not the news. Crutchfield references DeepSeek's low-cost model and OpenAI's distillation allegations as context for why open-weight options are now on the table for 2027 planning.
Did any e-discovery vendor announce a switch to open-weight models?
No. The column cites one unnamed legal tech provider building its software on an open-weight model, plus Thomson Reuters building its own model — neither is an e-discovery-specific announcement.
Should firms wait for lower AI pricing before signing 2027 contracts?
No — but they should push vendors now to unbundle pricing by task, since the columnist argues that pricing is already achievable with current open-weight options.
Sources: LawSites, guest column by Ken Crutchfield.
The original report
Industry headlines from other publications. Each links to the original reporting on the publisher's own site.