We've spent three posts on the mechanics of getting AI document review cheap and accurate: which model to pick, how to architect a pipeline that gets both checks under a nickel, and how to clear most of a collection deterministically before a model ever runs. Each of those posts ends in roughly the same place: the review architecture sets the ceiling, model choice is a cost decision within that ceiling, and a small model fine-tuned on the right labels can match a much larger one on this specific, narrow task.
This post asks the question those findings point toward but don't answer: if a small, fine-tuned model can match a frontier model on privilege and responsiveness review, who should own that model? Right now, for almost every firm, the answer is a vendor or a foundation lab — you pay per document, per token, for the life of every matter, forever, and the intelligence you're renting is identical to what your opposing counsel's firm is renting. We think that's backwards, at least for firms running enough volume to make it worth fixing. Large firms are sitting on the raw material to train something they'd actually own. Almost none of them are collecting it.
1.The Rented Intelligence Problem
Every firm calling a frontier model API for document review is buying the same commodity intelligence as every other firm calling that same API, at the same list price, subject to the same rate limits and the same deprecation calendar. There is no version of "our AI review is better than theirs" available to a firm whose entire review stack is a pass-through to someone else's model — the underlying intelligence is, by construction, identical. Differentiation has to come from somewhere else: better pre-filtering, better escalation logic, better QC sampling. All real levers, and all the subject of our prior posts. But none of them touch the one input that actually is proprietary to a firm: the judgment calls its own attorneys have already made, on its own matters, under its own risk tolerance.
That rented-intelligence position also carries a cost structure a firm doesn't control. Per-token pricing moves at a foundation lab's schedule, not the firm's. Models get deprecated and replaced with successors that score differently on the same task — our own benchmark work shows F1 swinging by double digits between models from the same family. A litigation budget set at matter intake, priced against today's API rate card, is exposed to a line item that can move for reasons that have nothing to do with the matter.
2.What Your Review Process Already Produces — and Throws Away
Here is the raw material a large firm already has, generated as a byproduct of work it was doing anyway. Every document a well-architected review pipeline escalates for a closer look is a labeled example of a case the cheap first pass got wrong or wasn't confident about. Every QC decision an attorney makes on a sampled document is a gold-standard human label, tied to that firm's specific interpretation of a specific privilege definition, for a specific client, in a specific jurisdiction. A firm running review across dozens of concurrent matters generates thousands of these decisions a month — and today, when the matter closes and the file is archived, that signal is discarded along with it.
A generic model, or even a generic fine-tune, is trained to approximate some firm's average judgment on some version of a privilege definition. Your firm's escalations and QC decisions encode something more specific: how your attorneys actually resolve the close calls, under your client's specific privilege posture, in the practice areas you actually handle. That specificity is exactly what a public benchmark or a general-purpose model cannot give you — and exactly what supervised fine-tuning is built to capture.
This is not a hypothetical resource. It's the same labeling motion behind the benchmark methodology in our model selection research — a fixed, gold-labeled document set run through a controlled pipeline to measure precision and recall against ground truth. A firm doing privilege and responsiveness review at volume is producing that same kind of gold-labeled data continuously, as an unavoidable output of doing the review correctly. The only thing missing is a mechanism to capture it and turn it into a model.
3.Why a Small, Fine-Tuned Model Wins This Specific Task
The instinct to reach for the biggest, most capable model is understandable but, per our own benchmark data, wrong for this task. Across nine models spanning a 68× cost range, eight landed within an 11-point F1 band — the most expensive model tested scored below models costing 30× less. The reason is structural: privilege and responsiveness classification is a narrow, repeatable task against a fixed written definition, not an open-ended reasoning problem. Once a review architecture handles document understanding and criteria evaluation correctly, a bigger model has very little marginal signal left to contribute.
That same narrowness is precisely what makes this task an unusually good fit for supervised fine-tuning of a small, open-weight model. Distillation works best when the target behavior is well-defined and the training examples are numerous and consistent — both true of a firm's own escalation and QC history, and much less true of open-ended chat. We're not describing this in the abstract: it's an active research track for us, running supervised fine-tuning and DPO (direct preference optimization) on open-weight 7–9B models — Qwen2.5-7B-Instruct, DeepSeek's 7B chat model, GLM-4-9B — using LoRA adapters to distill a flagship model's judgment into a fraction of the parameter count, then validating the result against the same kind of gold-labeled benchmark set that produced the model selection data above.
4.The Compliance and Sovereignty Case
Privilege review has a distinct error asymmetry we've written about before: a false negative can waive privilege under FRE 502, and a false positive builds a privilege log opposing counsel can challenge entry by entry. A model that runs entirely inside a firm's own environment removes a third category of risk that neither error type captures — the question of whether privileged content should be transmitted to a third party's inference endpoint at all, on every single document, for the life of the engagement. Client outside counsel guidelines are moving toward restricting exactly this, and a firm whose review stack depends on a call to a public model API has no answer to a client that says no.
DecoverAI's standard policy, stated on our security page, is that customer documents are never used to train, fine-tune, or improve our models. That default doesn't change for firms that don't ask for this. What we're describing here is a separate, opt-in engagement: a firm directs us, under its own instruction and its own data processing agreement, to build a private model trained exclusively on that firm's own reviewed matters, deployed exclusively for that firm, and never shared with or accessible to any other client. It is closer to commissioning custom software than to a platform-wide feature — your data trains your model, full stop.
5.The Economics: From Variable Line Item to Owned Asset
Hosted-API review cost is a metered expense that scales linearly with volume and is set by someone else's price list — our own benchmark shows a 68× spread between models on that list, with no guarantee the ranking holds after the next model release. A private, fine-tuned model running on infrastructure a firm controls turns that into a largely fixed compute cost. The economics improve with scale in a way rented intelligence never does: every additional matter adds training signal at close to zero marginal cost, while a hosted-API bill adds a dollar figure at exactly the same per-document rate as the first matter did.
The comparison that matters to a managing partner or GC isn't this year's invoice — it's the trendline. A hosted-API cost structure is flat or rising for the life of the firm. An owned-model cost structure gets cheaper and more accurate the more the firm uses it, because every matter it runs is training data for the next one.
Gen AI Use Cases for In-House Counsel
Join DecoverAI CEO Ravi Tandon for a live session on how in-house legal teams are using Generative AI to accelerate investigations, compliance reviews, and due diligence — including how we post-train large language models to run Document Review at under 5¢ a document, the architecture underneath the ownership case in this post.
Save Your Seat →6.The Competitive Case for the Firm, Not Just the Budget
Clients pushing for alternative fee arrangements on discovery-heavy matters are asking a version of the same question we're asking here: why does review cost scale linearly with document count when the marginal cost of running a model against one more document is close to zero? A firm that can say "we run privilege and responsiveness review on a model trained on our own attorneys' judgment, deployed in our own environment" has an answer that a firm quoting a per-document rate tied to a foundation lab's list price does not. That is a differentiator in an RFP, and it is a differentiator that competitors cannot copy by switching vendors, because the model is trained on judgment specific to the firm that built it.
There's a second, quieter effect. Senior associate and partner judgment on close privilege calls is, today, tribal knowledge that leaves the firm when the person does. A fine-tuning pipeline built on that judgment turns some of it into an institutional asset that persists past any one attorney's tenure — not a replacement for that judgment, but a record of it that the firm actually owns.
7.Why This Is Hard to Do Alone
None of this is a reason for a firm's IT department to stand up a distillation pipeline on its own. The work looks straightforward from a distance and is not:
- Label curation, not just collection. Escalations and QC decisions have to be deduplicated, checked for internal consistency, and balanced across practice areas so a fine-tune doesn't overfit to one client's document style or one matter's unusually easy fact pattern.
- An evaluation harness that catches regressions before they ship. A fine-tuned model has to be checked against a held-out gold-labeled set for precision and recall against the teacher model it was distilled from — the same discipline behind our nine-model benchmark — before it ever touches a live matter. Skipping this step is how a firm ends up with a model that looks fine in a demo and quietly under-recalls on the doctrine that matters most to its practice.
- Getting the precision-recall tradeoff right per doctrine. As we've written before, privilege review and responsiveness review have different error asymmetries. A generic fine-tuning recipe that doesn't account for that will optimize the wrong metric.
- A continuous retraining loop, not a one-time project. The value of this compounds only if new matters keep feeding the model. That requires infrastructure most legal IT teams aren't staffed to build or maintain: versioning, scheduled retraining, and rollback if a new fine-tune underperforms its predecessor.
- Secure, production-grade self-hosted inference. Standing up a model behind the same access controls, redaction pipeline, and production formatting as the rest of a review workflow is an integration problem, not a research problem — and it's the part most likely to get shortcut.
Most large firms are good at security, good at legal judgment, and increasingly good at buying AI tools. Very few have an ML engineering function that does distillation, evaluation science, and MLOps for a living. That gap is the actual argument for a technical partner, not a reason to avoid model ownership altogether.
8.Why DecoverAI Is Built to Do This With You
Everything in the previous section is infrastructure we've already built, because we needed it to run our own platform before this idea was ever a product conversation. Our two-step review architecture — document understanding, then criteria evaluation, with escalation on uncertainty — is what generates the exact labeled data described in Section 2, as a byproduct of running review correctly rather than as a separate collection effort. Our benchmarking discipline — the same gold-labeled, controlled methodology behind Working Paper 2026-02 — is the evaluation harness a fine-tuned model has to clear before deployment. And our supervised fine-tuning and DPO research on open-weight 7–9B models is the distillation pipeline itself, not a roadmap slide.
We've also already built the integration point. DecoverAI's chat review architecture supports customers bringing their own model into the classification pipeline — the same message-level tagging, thread reconstruction, redaction, and production formatting underneath, with the model choice left to the customer. A firm-owned, fine-tuned model plugs into that same production pipeline; it doesn't require re-architecting ingestion, chain-of-custody handling, or production formatting to use it.
The honest way to describe most AI-in-legal vendors is that they sell a well-designed interface around someone else's frontier model. We think the more durable position is the one we've built toward across our research: an architecture that generates the training signal, a benchmarking discipline rigorous enough to validate a fine-tune before it touches a live matter, and an active research program already running the distillation experiments a firm would otherwise have to build from nothing. That combination is the actual case for DecoverAI as the technical partner for this, not just a vendor selling access to a model somebody else trained.
9.A Framework: Is Your Firm Ready to Own a Model Now?
Model ownership isn't the right move for every firm today. Here's how we'd think about the decision:
This isn't a pitch to train a large language model from scratch, or to feed data into a fine-tune the firm doesn't already have the rights to review. It's supervised fine-tuning of a small, open-weight model on labels a firm's own review process already produces, validated against a gold-labeled benchmark, and deployed privately. The model doesn't replace attorney judgment — it operates inside the same escalation and QC framework described throughout our research, tuned to a specific firm's decisions instead of a generic teacher's.
10.Conclusion
The three posts before this one established that model choice is a cost decision once the review architecture is right, and that a small model can match a much larger one on this specific classification task. The conclusion those findings point to is that the intelligence behind privilege and responsiveness review doesn't have to be rented from a foundation lab for the life of a firm's practice. It can be built from labels the firm is already generating, validated with the same rigor we apply to our own model benchmarks, and deployed inside an environment the firm controls.
That is a harder thing to build than calling an API, which is exactly why most firms haven't done it, and exactly why doing it requires a technical partner who has already built the pipeline that generates clean labels, the benchmark discipline to validate the result, and the fine-tuning research to produce it. That's the partnership we're describing here, not a hypothetical one — ask us for the same kind of controlled validation run behind our public benchmarks, applied to your own matters.
To talk through what a private, fine-tuned model would look like for your firm's own document volume and practice mix, book a session with our technical team.