How we use AI — and where it stops
Most firms use AI and don’t say how. We’d rather you knew exactly what touches your documents: the machines do the reading at scale, the lawyer does the judging, a layer of standards-as-code gates what may be issued, and the signature at the end is human, accountable and insured.
Four moves, in order
Exact match over our own playbooks — no free-floating output reaches a deliverable.
A qualified lawyer reads every flag in context and decides.
Standards written as code are built to stop a known breach being issued — a check, not just a promise.
A human signature that carries professional liability and indemnity cover.
What the machines do
Search, first-pass drafting, clause comparison, consistency checks — at machine speed, under instructions we wrote and control. AI does the reading at scale that used to make legal work slow and expensive.
What only counsel does
Judgment. Every flag is read in context, every conclusion is decided by a qualified lawyer, and nothing leaves under our name without a signature that carries professional liability and indemnity cover. AI never advises you, never decides, never signs.
Grounding, not generation
Reviews run against our own playbooks — positions distilled from live practice — and every finding is pin-cited to your document and to our stated position. No free-floating model output reaches a deliverable.
Confidentiality
Matter documents are handled under the profession's secrecy rules, processed in controlled environments, and never used to train any model. Where practice becomes a template or an article, it is anonymised first — the pattern, never the matter.
Gold standards, grep, and frontier models
We do not fine-tune models on anyone’s data — we ground them. Our gold standards live as playbooks: positions distilled from live practice, anonymised, and maintained by counsel. At run time, deterministic pattern search — the grep family of tools, in the lineage of Thompson’s regular-expression search algorithm (Communications of the ACM, 1968) — pulls the exact relevant standard into the model’s working context. The most advanced publicly available models are then instructed to compare your document against that standard, clause by clause, and to quote both sides of every difference.
Computer science calls the wider family retrieval-augmented generation (Lewis et al., NeurIPS 2020). We deliberately work at its deterministic end: exact match over curated standards, pin-cited, auditable — because in legal work a finding you cannot trace is a finding you cannot sign.
The discipline answers a now well-documented problem. The canonical survey of generative-model hallucination, in ACM Computing Surveys, frames it as a structural failure mode of ungrounded generation, not a passing bug (Ji et al., 2023). Two Stanford studies made the legal stakes concrete: frontier models hallucinate on legal queries roughly half the time (Dahl, Magesh, Suzgun & Ho, Journal of Legal Analysis, 2024), and even the purpose-built commercial legal-AI tools — some marketed as “hallucination-free” — produce incorrect or unsupported assertions in 17–33% of queries (Magesh, Surani, Dahl, Suzgun, Manning & Ho, Journal of Empirical Legal Studies, 2025). Grounding, pin-citation and human sign-off are not over-engineering; they are the minimum honest response.
A standard you can prove, not merely promise
“A human reviews it” is true and not enough — a person can be rushed, a checklist skipped. So between the work and any act that leaves under our name sits a small enforcement layer we call a governance kernel. It does no legal thinking; it only checks that what should have happened, did, before a door opens.
There are three gates, and each rule names the engagement term or professional duty it enforces — so the code and the letter are the same standard, written twice. The default is denial: an action no rule yet covers does not pass.
Conflicts & KYC/AML
No matter is accepted until both are clear.
Scope · PII · no-retention
The machine never reaches the act of signing.
Sign-off, scope, cap, disclaimer
The gate we design not to be skipped.
When you want the model itself, on the same discipline
Some clients ask us to install the harness, not just apply it. We will — on a clean scope and the same transparency promise. Two shapes, depending on the structure.
Single model, grounded
A frontier model wired to your gold-standard playbook through the grep-grounded pattern we run internally. You get the playbook, the prompt architecture, the retrieval rules, and the user-facing notices the GDPR and AI Act expect to see when a machine is part of the decision.
Convergent orchestration
For structures where one model is not enough, several models read the same input under the same ruleset and the output is what they agree on. The empirical basis is published: sampling multiple reasoning paths and taking their consensus measurably improves accuracy (Wang et al., ICLR 2023), and structured debate among models reduces factual errors and hallucinations (Du, Li, Torralba, Tenenbaum & Mordatch, ICML 2024). Disagreement is logged, not hidden, and routed for human review — this is how we shorten the gap between machine speed and counsel-grade certainty when the stakes ask for both.
Scope, stated plainly
What we own: the legal-governance layer — playbook, grounding rules, prompt architecture, evaluation criteria, user notices, the policy under which the model may and may not act. What we do not own: production engineering, hosting, MLOps, model security at infrastructure level. Those stay with your team or a coordinated partner; we work to their interface, not in it. Every deployment ships with the same transparency you see on this page — including the link back to it, on the surface where end-users meet the model.
The positions on this page are not just our preference. The short list below is the peer-reviewed and academic spine they stand on — survey, evidence, method.
- Ji, Z., Lee, N., Frieske, R., Yu, T., Su, D., Xu, Y., Ishii, E., Bang, Y. J., Madotto, A. & Fung, P. (2023). Survey of Hallucination in Natural Language Generation. ACM Computing Surveys 55(12), Art. 248. doi.org/10.1145/3571730
- Dahl, M., Magesh, V., Suzgun, M. & Ho, D. E. (2024). Large Legal Fictions: Profiling Legal Hallucinations in Large Language Models. Journal of Legal Analysis 16(1), 64–93. Oxford Academic
- Magesh, V., Surani, F., Dahl, M., Suzgun, M., Manning, C. D. & Ho, D. E. (2025). Hallucination-Free? Assessing the Reliability of Leading AI Legal Research Tools. Journal of Empirical Legal Studies. Wiley
- Wang, X., Wei, J., Schuurmans, D., Le, Q. V., Chi, E. H., Narang, S., Chowdhery, A. & Zhou, D. (2023). Self-Consistency Improves Chain of Thought Reasoning in Language Models. ICLR 2023. arXiv:2203.11171
- Du, Y., Li, S., Torralba, A., Tenenbaum, J. B. & Mordatch, I. (2024). Improving Factuality and Reasoning in Language Models through Multiagent Debate. ICML 2024. arXiv:2305.14325
- Choi, J. H., Monahan, A. B. & Schwarcz, D. (2024). Lawyering in the Age of Artificial Intelligence. Minnesota Law Review 109(1), 147–218. SSRN
- Background:Thompson, K. (1968). Regular Expression Search Algorithm. Communications of the ACM 11(6), 419–422. · Lewis, P., Perez, E., Piktus, A., Petroni, F., Karpukhin, V., Goyal, N., Küttler, H., Lewis, M., Yih, W., Rocktäschel, T., Riedel, S. & Kiela, D. (2020). Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks. NeurIPS 2020.
Unchanged by AI
The professional frame does not bend to the tooling. Every matter still opens with a conflict check, identity verification and an engagement letter before advice is given; Romanian and EU law is advised by a practising avocat; English law is comparative analysis by a solicitor (non-practising). AI changes the speed of the reading — never the locus of responsibility. The first randomised controlled trial of legal-AI assistance — published in the Minnesota Law Review (Choi, Monahan & Schwarcz, 2024) — found exactly that pattern: large, consistent gains in speed; quality gains that were “slight and inconsistent” without lawyer judgment. The signature at the bottom is doing what no model can. And where an engagement is agreed as human-only, we drop AI for that matter entirely — no model touches the file.
Ask us about it directly