What separates an AI draft from a template.
A template fills blank spaces. A draft makes a case. The distinction is small enough to miss in casual conversation and large enough to determine, on most days, the difference between a document that survives review and one that does not. The discipline of drafting at the standard of a senior practitioner is the discipline of refusing to confuse the two operations. It is the work that legal professionals are most jealous of, most exacting about, and most reluctant to delegate. It is also, increasingly, the work that legal artificial intelligence is asked to do.
This essay is about what it takes for a system to draft at a standard that a partner would sign. The standard is high. It is not unreachable. It is not, however, achieved by templates with sophisticated fillers, by language models trained on legal text, or by retrieval systems that surface the right precedent. Each of these helps. None of these is sufficient. The drafting that earns the partner's signature is drafting that does several things at once, in a particular order, with a particular orientation toward the reader, and with a particular willingness to take a position. We will set out, in turn, what those things are.
Begin with structure. A document that has been well drafted is, before anything else, structured. The structure is not the imposition of headings on an underlying flow of prose. It is the deliberate arrangement of the matter so that the reader, moving through the document in the order it was written, encounters the propositions in the sequence that builds the case.
The architecture of a partner-quality draft has several stable features. It opens by orienting the reader: who is writing, who is being addressed, what the document is, and why it has been produced. It states, near the beginning, the principal proposition the document advances. It then provides the support for the proposition, in an order that respects how legal arguments are built. The factual matrix precedes the legal framework; the legal framework precedes the application; the application precedes the conclusion. Each of these sections, in turn, has internal structure. The factual matrix is organised chronologically, or by issue, or by party, depending on what the matter requires. The legal framework is organised by issue, with each issue addressed in its own section. The application returns to the issues in the same order, applying the framework to the facts. The conclusion summarises and, where appropriate, requests action.
This architecture is not a formula. It is an idiom. Its rhythm is felt by experienced readers, and its absence is felt as well. A document that begins without orientation, that buries its principal proposition halfway through, or that organises its facts by importance rather than by chronology, makes the reader work harder than necessary. The reader will, often, do the work, but the document has lost something the partner-quality draft would have preserved. It has lost the reader's confidence that the writer knows what she is doing.
An artificial intelligence drafting at the partner standard must, accordingly, internalise the architecture rather than apply it. Internalisation looks like this: when the system drafts, it does not begin with prose and then impose a structure. It begins with the structure, derives the substantive content of each section from the preceding sections, and produces the prose only after the structure has been stabilised. The structure is the skeleton; the prose is the flesh. The order of operations is not negotiable.
Beyond structure lies voice. A document does not only say something; it says it in a particular register. The register is calibrated to the audience, to the matter, and to the institution. A submission to a court is written in a register that differs from a memorandum to a client. A regulatory filing is written in a register that differs from an internal opinion. The differences are not stylistic flourishes. They are part of the meaning of the document.
An AI drafting system that produces, by default, a uniform register across these contexts has not drafted at the partner standard. It has produced legal text. The two are not the same. The partner who reviews the draft will, often within the first paragraph, decide that the register is wrong, and will rewrite the document into the right register before going further. The drafting effort, at that point, has been wasted. Worse, it has imposed a cost on the partner's time that the absence of the system would not have imposed.
The discipline of register is therefore part of the discipline of drafting. The system must know, before producing a single sentence, what kind of document it is producing, who the reader is, what relationship the writer stands in to the reader, and what register the matter calls for. These are not technical parameters. They are the questions a senior partner would ask of a junior associate before the associate began. A system that does not ask them, because they have not been provided in the prompt, is a system that has, at best, guessed.
This is why the better systems include explicit clarifying interfaces at the front end of the drafting process. Before drafting begins, the system establishes the basic parameters of the document. The clarifying questions are not invitations to elaborate; they are precise requests for the specific information the drafting requires. The user who supplies the answers receives a draft calibrated to her matter. The user who skips the answers receives a draft calibrated to a default that may or may not be appropriate. The discipline is to make the default visible rather than silent.
A subtler aspect of voice is what might be called authorship. A document is read, by experienced readers, as the work of a particular author. The author may be the firm rather than the individual lawyer; she may be the client rather than the lawyer; she may be an unnamed institutional voice. In each case, the document conveys, through its choices of phrasing, emphasis, and sequence, who is speaking. A partner-quality draft is one in which the authorship is clear and consistent. A weak draft is one in which the authorship drifts: the document is partly the firm's voice, partly the client's, partly nobody's.
Drafting systems that operate at the partner standard preserve authorship by holding the author's voice constant across the document, by avoiding the linguistic patterns characteristic of automated text, and by producing prose whose rhythm matches the practised legal hand. None of this is achieved by post-processing. It is achieved by drafting, from the start, with the author in view. The author is a parameter of the system. The system does not produce text and then attribute it; it produces text from a defined authorial perspective.
A draft, in the legal sense, is rarely neutral. It advances a position. The position may be the client's position in a dispute, or the firm's view of a question, or the writer's recommendation in a memorandum. In each case, the document is not a survey of possibilities. It is the marshalling of facts and authorities behind a specific stance. To draft is to decide, on behalf of the client or the firm, what stance the document will take, and to commit to it.
This commitment is exactly what most artificial intelligence systems struggle to make. The default behaviour of language models, when faced with an open question, is to enumerate the considerations on each side. The enumeration is informative. It is not, however, drafting. A document that enumerates without committing has performed analysis, not advocacy. A partner reviewing the document will recognise it as analysis, will accept it as analysis, and will then ask the question the analysis has not answered: which side are we taking?
The drafting system must, accordingly, be able to take a position. The position is sometimes given by the user, in the form of an instruction: draft a notice asserting that the contract was breached. The position is sometimes given by the matter, in the form of facts that admit only one reasonable view. The position is sometimes given by the framework of authority, in the form of a holding so clear that no other position is defensible. In each case, the system must hold the position throughout the draft, not only in the conclusion but in the structure, the language, and the choice of authorities. The opening paragraph of a partner-quality draft already announces the position; the closing paragraph completes the line of argument that the opening began.
Holding a position does not mean ignoring the opposing view. The strongest drafts engage the opposing view explicitly, address it directly, and explain why the writer's position prevails despite it. This is the discipline of advocacy that combines candour with commitment. It is also the discipline most easily lost in automated drafting, where the temptation to absorb the opposing view into a balanced enumeration is strong. The system that resists the temptation, that surfaces the opposing view as something to be answered rather than something to be acknowledged, has drafted at the partner standard.
Facts, in legal drafting, are not raw material. They are crafted. The selection of which facts to include, the order in which to present them, the level of detail at which to recount them, and the language in which to describe them are all decisions the writer makes, deliberately, in service of the position the document advances.
This is the place where automated drafting most often falls short. A system that has been given a fact pattern can reproduce the fact pattern, summarise it, or rephrase it, but to craft it is something different. Crafting requires understanding which facts will move the analysis and which will not, which facts the opposing side will rely on and how to address them, and which facts can be left implicit because the reader will supply them.
The crafting of facts is the senior practitioner's discipline that most distinguishes the partner-quality draft from the competent one. The partner reads the matter, identifies the dozen facts that will determine the outcome, and constructs the factual narrative around them. The competent draft, by contrast, often recounts the facts as they appeared in the file, with the determinative facts buried among the rest.
For an AI system to draft at this standard, it must do more than ingest the fact pattern. It must analyse the fact pattern, identify the facts that move the analysis, structure the factual narrative around them, and produce a recital of the facts that is, simultaneously, accurate and persuasive. The recital is not a free-form retelling; it is a choreography of detail. We have come to think of this as the most demanding internal capability of a drafting system. The system that performs it well produces drafts that read as written; the system that performs it poorly produces drafts that read as assembled.
Authority, in legal drafting, performs several functions. It supports propositions of law that the writer asserts. It establishes the framework against which the facts are evaluated. It provides the persuasive force that the bare assertion of a proposition cannot supply. It signals, to the reader, that the writer has done the research that earns the right to advance the position.
The use of authority in a partner-quality draft is, like the use of facts, crafted. The drafter does not cite every relevant authority. She cites the authorities that, in combination, build the strongest case for the position. She prefers binding authority to persuasive authority where the choice is available. She prefers controlling authority to general authority where the matter requires precision. She quotes the cited authority sparingly, where the language of the court is more powerful than her own paraphrase, and paraphrases otherwise. She introduces each authority with the proposition it supports, rather than dropping the citation at the end of an unsupported sentence.
Each of these is a craft decision that an automated system must learn to make. The system that cites all available authority, indiscriminately, produces a document that is dense and unpersuasive. The system that cites authority sparingly, but does not match each authority to the proposition it supports, produces a document that is incoherent. The system that quotes too much produces a document that reads as a copy. The system that quotes too little produces a document that loses the persuasive force the quotation would have supplied.
The drafting system that operates at the partner standard, accordingly, treats authority not as decoration but as architecture. Each authority appears at the place in the argument where it does the most work. Each is introduced in the way that a careful drafter would introduce it, with the proposition it supports and the relationship it bears to the position the document advances. The verification discipline that ensures each authority remains good law sits behind the drafting; the craft discipline that places each authority in its proper place sits in front of it.
Beyond the visible discipline of structure, voice, position, facts, and authority sits a less visible layer that is, in some matters, the determinative one. Strategy is the lawyer's judgement about what to advance, what to defer, and what to leave unstated. It is the calculation that some arguments, however available, are best held back; that some concessions, however reluctant, are best made early; that some positions, however tempting, are best not taken.
Strategy is the senior partner's domain. It is the part of legal practice that is least amenable to automation, because it depends on the reading of the other side, the reading of the bench, the reading of the regulator, the reading of the client's underlying interest, and the reading of the wider context in which the document will be received. A drafting system that attempts to perform strategy unsupervised will, at best, do so by approximation. The approximation is rarely good enough.
The honest position for a drafting system is therefore that strategy remains the partner's call, supported but not supplied by the system. The system can produce drafts that reflect the strategy the partner has decided. The system can surface considerations the partner may wish to weigh. The system cannot, and should not, take the strategic decision in the partner's place. Drafting at the partner standard, in this sense, includes the discipline of stopping at the boundary that the partner's judgement properly occupies.
This discipline has practical consequences. The drafting interfaces we provide make explicit room for strategic input. The user is asked, before drafting begins, to specify the position to be advanced, the concessions to be made, the arguments to be reserved, and the tone to be struck. The system drafts within those constraints. A user who supplies them gets a draft that reflects her strategy. A user who omits them gets a draft that reflects a default which may or may not match her view. The clarity of the interface is part of the discipline.
No partner-quality draft is produced in a single pass. The discipline of drafting is the discipline of iteration. A first draft is produced, reviewed, revised, reviewed again, and refined until the document is ready. The iteration is part of the work, not a sign that the first draft was inadequate.
An AI drafting system, accordingly, is not designed to replace iteration. It is designed to participate in it. The first draft is produced quickly, in a form that is structurally complete and substantively reasonable. The reviewer reads the draft, identifies the changes, and supplies them. The system applies the changes, with awareness of the structural implications, and produces the second draft. The cycle continues until the reviewer is satisfied.
The participation of the system in the iteration loop changes the economics of drafting. The first draft, which would have taken hours to produce manually, is produced in minutes. The reviewer's time is then concentrated on the work the reviewer is best positioned to do: the strategic refinement, the substantive correction, the calibration of voice. The system does not replace the reviewer; it amplifies her, by producing the first complete draft fast enough that the iteration cycle can be conducted within a single working session rather than across days.
This change in economics is, we suspect, the most consequential commercial implication of artificial intelligence in legal drafting. The partner who reviews drafts has more drafts to review, more matters to handle, more attention to give to the matters where her attention is most needed. The associate who produced the drafts can, instead, supervise the system's drafting and learn from the iteration. The training pipeline of the firm is preserved, the productivity of the firm is increased, and the standard of the work product is held.
Throughout this essay we have spoken of partner-quality drafting as a standard. It is worth being precise about the role of the system in the production of such drafts. The system is not the partner. The partner reviews, judges, decides. The system produces drafts that meet the standard the partner is willing to put her name to, after the partner has reviewed them and made the corrections that her judgement requires. The system is, in functional terms, a senior associate: technically competent, trained in the drafting conventions, capable of producing complete first drafts at speed, requiring supervision but not requiring re-doing.
The senior associate framing is, in our experience, the most useful way to think about the system's role. It sets the right expectations on the firm's side. The firm does not expect the system to produce documents that are ready to file without review. The firm does expect the system to produce documents that are ready to review, that contain the structural work, the factual recital, the legal framework, and the application that a senior associate would supply. The partner's role is to apply judgement on top of competent drafting, not to do the drafting from scratch.
This framing also sets the right expectations on the system's side. The system is not asked to perform the partner's strategic judgement; it is asked to perform the senior associate's drafting work. The asking is calibrated to what the system can do well, with the boundary at what the system cannot do well held firm. Within that boundary, the system can be very strong. Outside it, the system declines to overreach.
The drafting pipeline that emerges from these disciplines is, in our system, a sequence of skills rather than a single act. The intake skill captures the matter, the position, the audience, the strategy, and the constraints. The factual analysis skill identifies the determinative facts and structures the recital. The framework skill assembles the controlling authorities and orders them by issue. The application skill applies the framework to the facts. The drafting skill produces the prose, in the appropriate register, holding the position. The review skill examines the draft for structural integrity, citation discipline, and consistency. The verification skill, as discussed elsewhere in our writing, examines each citation for currency and accuracy.
The pipeline is configurable. A firm that has its own drafting conventions can encode them as a configuration. A senior who has her own preferences can apply them. A matter that requires special handling can be routed accordingly. The configurability is what allows the system to draft at the partner standard for many partners, in many firms, on many matters, without imposing a single house style on all of them.
The result is, on the user's side, a draft that arrives complete, structured, voiced appropriately, holding the position, supported by authority that has been verified, and ready for the partner's review. On the system's side, the result is a sequence of operations that has been performed, logged, and made available for inspection. The work of drafting at the partner standard is, finally, the integration of these disciplines into a single arriving document. The discipline is not optional. The integration is not automatic. The result, when it arrives, is a draft that the partner can sign, after the partner has done what the partner does, with the system having done what the system does.
— Editorial