From the journal

A practical guide to contract drafting with AI

More clients are drafting contracts themselves, often with AI, before sending them to lawyers. A few simple checks can help turn a polished first draft into an agreement that actually reflects the deal, the risks, and the law that applies.

Illia ProkopievCo-Founder and CEO9 min read

Start with the actual transaction

Before requesting or revising a draft, establish whose interests it should represent and what the parties have agreed. Identify the commercial decisions that remain open. Give the drafter enough information to distinguish an agreed term from an option still under discussion.

Instructions to produce a “strong,” “balanced” or “standard” agreement leave important choices unspecified. Explain the intended arrangement in concrete terms: the services to be provided, the permitted uses of the deliverable, or the circumstances in which a party should be able to withdraw.

A hypothetical supply agreement illustrates the point. Permission to substitute equivalent components may suit a buyer that values flexibility. The same permission would depart from instructions requiring prior approval of every substitution. Assess the wording against the relevant instructions before deciding whether it is good drafting.

Make the language easier to understand without changing the bargain

Use direct sentences and identifiable actors. Give the person responsible for performance a clear account of the obligation. Remove unnecessary repetition and elaborate wording that adds no precision.

Simplification requires attention to meaning. When shortening a provision, preserve the conditions that activate it and the exceptions that limit it. Review changes to words expressing obligation, permission and discretion. A shorter sentence is useful only when it still communicates the intended arrangement.

Keep defined terms consistent. Replacing a repeated term with a synonym for stylistic variety can introduce doubt about whether the wording concerns the same subject. Necessary repetition is preferable to an unexplained distinction.

Readability should remain a separate assessment from legal effectiveness. A clear sentence can express an unauthorised obligation. A difficult sentence can contain a protection that must survive editing.

Ask for the basis of a proposed provision. It may implement an agreed term, address an applicable legal requirement or reflect a preferred negotiating position. Keep those explanations distinct.

A recommendation to adopt a clause should identify the relevant reason. “Market standard” needs support about the relevant market and transaction. Familiar wording in a precedent does not establish that support.

For legal propositions, consult the applicable source and confirm its relevance to the transaction and date. Check the jurisdiction before borrowing terminology or provisions from another agreement. Where the existing instrument specifies governing law, preserve that choice unless a change is authorised. Separately assess any legal requirements that may apply despite that choice.

A model’s explanation can identify a question worth researching. Treat the explanation as a starting point for the legal assessment, rather than confirmation of the rule.

Read the agreement as a connected document

Review the main terms together with the schedules and incorporated materials. A clause can appear satisfactory by itself while creating uncertainty when read alongside another provision.

Consider a hypothetical agreement that requires payment after acceptance. If the invoicing schedule instead requires payment on delivery, determine whether the difference is intentional and how the provisions interact. Resolve the allocation through the agreed terms; do not select whichever sentence appears more familiar.

Check whether references lead to the intended provisions and whether definitions remain consistent throughout the document. Where general terms and a schedule differ, establish which governs the affected issue. Avoid assuming that document order supplies the answer.

Involve the people expected to administer the agreement where operational facts matter. Their input can establish whether the proposed process fits the transaction. Legal review must still address whether the wording expresses that process appropriately.

Keep revisions within the authorised scope

Distinguish correcting a defect from proposing a different bargain. A request to improve clarity does not settle whether a party should receive an additional right or assume a further obligation.

Suppose the agreed warranty concerns a product’s specifications at delivery. Extending that warranty throughout the product’s operating life is a substantive proposal. Its commercial appeal does not make it an editorial correction.

When reviewing changes, explain their effect on the transaction. Preserve agreed protections on either side unless the instructions or an applicable legal requirement call for a change. Put optional alternatives before the person authorised to decide them.

Restraint also applies to additions. Include a provision because it serves an identified purpose in the transaction. Avoid expanding a document merely to make it look more complete. Necessary detail should remain even when removing it would produce a shorter draft.

Keep assumptions visible and preserve deliberate discretion

Separate confirmed facts from assumptions and unresolved questions. Do not fill a gap with a plausible name, date, approval or commercial term.

A missing approval record leaves an evidential question. It does not establish that approval was refused or that no approval exists. Equally, the absence of contrary evidence does not justify stating that approval has been confirmed.

Identify what information is missing and which decision depends on it. Continue work that does not depend on the unresolved point, while retaining the limitation for the affected part.

Some uncertainty is deliberate. Parties may choose a reasonableness standard or leave a decision within defined discretion. Do not replace that choice with an invented threshold merely to obtain a definite answer. Explain what the standard leaves open and assess whether that uncertainty is acceptable for the transaction.

Check the conclusion as well as its supporting source

Finding a relevant source is only part of the work. Examine whether the proposed conclusion follows from the rule and the established facts.

A source may describe when a right becomes available. Before advising that the right is available in the transaction, establish whether its conditions are satisfied. Preserve any exception that could change the result.

Give serious consideration to a supported alternative interpretation. Identify the fact, instruction or authority that would change the answer. Where a material premise remains uncertain, keep the conclusion conditional.

The same discipline applies to AI explanations. Repeated answers and persuasive reasoning do not supply missing evidence. Ask for the basis of a material conclusion and examine that basis directly. An accurate quotation cannot repair an inference that goes further than the quoted material permits.

Use AI for a defined task and assess the resulting work

Specify the assignment before assessing whether AI performed it well. Extracting a deadline, summarising a clause and drafting an agreement require different outputs. Success on one assignment does not establish competence across the others.

Give clear constraints for the requested work. State the relevant document, instructions and permitted scope of change. Require unresolved commercial choices to remain identifiable instead of allowing the system to decide them silently.

Inspect the output against those constraints. Compare proposed revisions with the agreed position and examine their practical effect. A general explanation of why the revision is “safer” or “clearer” is insufficient when the change reallocates a material risk.

Use examples of the actual work when assessing suitability. Retain unsuccessful outputs as part of that assessment. Selecting only the strongest draft would leave the frequency and cost of correction unexamined.

Measure usefulness through completion and correction

Consider the work required to reach the intended endpoint. A quickly generated first draft may still require substantial review. Include that review and the resulting revisions when comparing drafting arrangements.

Record material errors separately from stylistic preferences. An omitted protection that defeats an agreed objective should not disappear within a favourable average writing score. Identify what the error would change and whether it was found before the document was used.

Also record tasks that remain unfinished. An arrangement that withholds every draft avoids issuing a defective clearance, but has not demonstrated useful completion. Assess whether withholding clearance was appropriate and what prevented the task from progressing.

Avoid reducing an assessment to an unexplained reliability percentage. State what was examined, what counted as an error and how unresolved cases were treated. Those details determine what the result can support.

Protect confidential methods while explaining the work product

Keep proprietary instructions, internal templates and confidential client materials outside public demonstrations. Use purpose-built examples or materials cleared for the intended disclosure.

When anonymising a transaction, consider whether distinctive commercial terms could identify it even after party names are removed. Include only the detail needed to explain the point.

Confidentiality should coexist with an understandable account of the result. A reviewer or client should be able to identify the task, the scope of assessment and any unresolved limitations. Disclosing those matters need not reveal the internal method used to produce the draft.

Keep claims proportionate to the evidence available. A successful example can illustrate an outcome. It cannot establish general superiority of an undisclosed method or identify which part of that method produced the result.

Describe review accurately

State what review occurred and who performed it. Distinguish an additional reading by the drafter from an assessment by another reviewer. Describe the relevant expertise and review scope without implying checks that were not performed.

A review label should help someone understand the basis for reliance. “Reviewed” is incomplete when it does not identify whether the assessment concerned language, commercial instructions, legal requirements or readiness for a particular use.

Where a necessary source or record was unavailable, identify the affected issue. Preserve completed findings within their supported scope. Repeating the same assessment without resolving the missing input does not strengthen the unsupported conclusion.

Choose review arrangements proportionate to the intended use and consequences of error. An automated check may contribute to that work. Its output still needs assessment against the function it is being asked to perform.

Release the document for its actual next use

Identify the next step before describing a draft as ready. An internal discussion draft can properly contain open choices that would prevent circulation for signature.

Communicate material open items with the draft. State what requires confirmation, who must decide it and which next step depends on that decision. Keep working commentary separate from the contractual text, while preserving required disclosures and other content that belongs in the instrument.

Release the version that was assessed. If a material term, fact or instruction changes afterward, reconsider the affected conclusions before reusing the earlier approval.

Where the outstanding issue concerns a commercial choice, obtain the authorised decision. Where it concerns missing evidence or a legal requirement, complete the relevant assessment. Until then, describe the document’s permitted use and limitations precisely.

***

And, of course, whenever AI helps generate a contract, talk to your lawyer before anyone signs, relies on, or acts on it.

For this guide, “AI slop” means plausible legal text whose apparent usability exceeds the support for relying on it. Treat that description as a warning about work quality, without assuming that authorship determines quality. Apply the same substantive expectations to human drafting and AI-assisted work.

Illia Prokopiev

Written by

Illia Prokopiev

Co-Founder and CEO

Illia is the Managing Partner and founder of Licentium. With over 11 years of practice, he has guided innovators through cross-border M&A deals and the disputes that follow, combining transactional skill with courtroom resolve. Admitted to the bar in 2017, he pivoted early to Web3, serving as legal advisor to prominent crypto projects and carrying AML/MLRO duties that anchored complex token, DAO, and compliance questions on solid regulatory ground. Certified in money laundering prevention and an active crypto investor, Illia blends market intuition with a global network of specialists, enabling Licentium to untangle licensing knots for crypto and AI ventures anywhere in the world.