By Sandeep Devanda, IT Expert, HAS Law Firm
Generative AI has moved from novelty to necessity in legal practice with unusual speed. Most firms have now given lawyers access to some form of large language model, whether through a dedicated legal AI product or a general-purpose assistant. Far fewer have given those lawyers a shared, practical understanding of how the technology actually behaves, where it helps, and where it must never be trusted unsupervised. That gap, not the absence of tools, is what determines whether AI adoption in a law firm produces genuine efficiency or a new category of professional risk.
Closing it does not require a large training budget or a proprietary tool. It requires a shared vocabulary, a simple structure for prompting, and a small set of firm-wide rules that everyone follows regardless of seniority. The framework below sets out a practical starting point that any firm can adapt, whatever its size or jurisdiction.
The junior colleague, not the database
The most useful mental model for a large language model is a very fast, very articulate junior colleague, not a legal database. An LLM generates the most statistically plausible next word based on patterns in its training data; it is not retrieving verified facts from an authoritative source. That distinction matters enormously in legal work, where a fluent, confident answer can be entirely wrong, including fabricated case citations that read as genuine.
The practical consequence is straightforward: supply the AI with the actual documents relevant to the task, the contract, the judgment, the client’s own file, rather than relying on what it may “recall.” This is commonly called grounding, and it is the single most effective way to reduce error. Even grounded output still requires human verification before it is relied upon, in the same way a partner would review a junior associate’s first draft rather than file it unread.
Briefing AI like an unsupervised junior associate
No competent lawyer would hand a first-year associate a one-line instruction and expect a court-ready memorandum. AI deserves the same courtesy, with one important difference: it will not ask clarifying questions unless invited to. A vague brief produces a generic, one-size-fits-all answer, because the model has no choice but to guess at intent. A specific brief, one that states the context, the task, the intended audience, and the desired format,produces a genuinely useful first draft. Either way, that draft is a starting point for a qualified reviewer to refine, never a finished work product.
A working framework: CLAIM
A simple, repeatable structure for legal prompting is worth more than individual habit or intuition. One useful checklist, which can be summarised as CLAIM, covers the elements a well-formed legal prompt needs:
- Context – the matter, document, or transaction being addressed.
- Legal task – whether the request is to summarise, draft, compare, or analyse risk.
- Audience – whether the output is destined for a partner, a court, a client, or an internal file note.
- Instructions – jurisdiction, constraints, tone, and the review standard expected.
- Mode of output – the specific format required, such as a memo, a table, or a redline.
The difference this makes in practice is best shown by example. “Review this contract” invites a generic summary, because it states no audience, no jurisdiction, and no requested format. “Review the attached NDA from a licensor’s perspective under UAE law, flag risk in the indemnity and liability clauses, and deliver a prioritised table of clause, issue, and suggested edit” gives the model a job it can actually do well, and gives the reviewing lawyer a first-pass output worth their time.
Precision can be increased further through examples. A zero-shot instruction with no example suits a simple, well-understood task, such as listing the termination clauses in a lease. Providing one worked example helps the model match a desired pattern, such as summarising a document in a specific house style. Providing several examples produces consistent output across a batch of similar items, useful when the same edit needs to be applied across a stack of contracts.
Operationalising it across a firm
Individual prompting skill does not scale on its own. Firms get more consistent value from three habits: setting standing instructions once, so a lawyer’s role, jurisdiction, and quality standard need not be repeated in every prompt; packaging recurring tasks into reusable, sharable instructions, so a good practice becomes firm-wide rather than one person’s habit; and connecting the AI, through properly secured channels, to the firm’s own current files, rather than leaving it to work from guesses or outdated memory.
Applied well, this approach supports several concrete use cases: first-pass flagging of unusual or missing clauses in contract review; surfacing research starting points and drafting routine correspondence; condensing long judgments, filings, or reports into working briefs; and producing draft translations and structured meeting notes for multilingual teams. In each case, the benefit is time returned to the lawyer for higher-value judgment, not a replacement for that judgment.
The non-negotiables
None of this is workable without governance. Four rules should be treated as non-negotiable, whatever the size or structure of the firm: every citation and statutory reference is personally verified before it is relied upon or filed; privileged and sensitive client information stays out of unvetted tools; every AI-assisted draft requires sign-off from a qualified human reviewer; and partner or senior approval is mandatory for anything court-bound or high-value. These rules exist precisely because AI is fast and fluent enough to be persuasive when it is wrong.
Literacy as the differentiator
The firms that get the most durable value from AI will not be the ones with the most licences, but the ones whose lawyers understand, at a shared and practical level, what the technology is good at, what it is not, and how to brief it accordingly. That is a training and culture challenge as much as a technology one, and it is one every firm, regardless of size or jurisdiction, can start addressing now.
Disclaimer – This article is for general informational purposes only and does not constitute legal advice.