AI Can Draft a Clause. Can It Draft the Agreement?

Why substantive legal knowledge, disciplined contract architecture and a tested drafting playbook remain essential

The danger of drafting without understanding

Generative AI can compare clauses, suggest wording and produce a first draft within seconds. Those capabilities are useful, but they do not remove the need to understand the type of agreement, the transaction it records, the applicable law and the reason why particular provisions have been drafted in a particular way.

An agreement is not a collection of independent clauses. It is a single risk-allocation system. Definitions, warranties, indemnities, termination rights, remedies and liability provisions interact. A change made to one clause can alter the meaning or commercial effect of several others. A user without substantive knowledge may accept an AI suggestion because it sounds comprehensive, without recognising that it duplicates, contradicts or displaces the agreement’s existing structure.

The recurring AI problem: every clause tries to solve the whole agreement

AI commonly reviews provisions one at a time. When it encounters a risk, it tends to address that risk where it appears. It may therefore add an explanation, exception or qualification under each relevant clause instead of identifying the correct “home” for the issue and dealing with it once.

This drafting method creates a document that appears thorough but is difficult to interpret and administer. Typical consequences include:

• the same rule being stated more than once in different language;
• multiple carve-outs that are similar but not identical;
• cross-references that multiply with each revision;
• uncertainty about which provision prevails;
• negotiated positions being changed unintentionally; and
• longer agreements that are less coherent and more likely to generate disputes.

Example: limitation of liability

Limitation of liability is a clear example. In a properly structured commercial agreement, the liability regime is ordinarily addressed in one dedicated clause. That clause should establish the cap or caps, excluded categories of loss, claims that fall outside the cap, and any special treatment agreed for indemnities, confidentiality, data protection, intellectual property infringement, fraud or wilful misconduct.

An unstructured AI review may instead insert a confidentiality carve-out in the confidentiality clause, an intellectual property carve-out in the intellectual property clause, a separate data-protection exposure in the data clause and another exception under each indemnity. The general liability clause then contains its own list of exclusions and exceptions.

The result is not additional protection. It is an interpretive problem. Does an uncapped obligation in the confidentiality clause override the general cap? Is an indemnity subject to the cap if the indemnity clause is silent but the liability clause is not? If two carve-outs use different wording, was the difference deliberate? The agreement may no longer communicate one clear liability allocation.

Why the agreement must be read as a whole

South African contract law reinforces the importance of coherent drafting. In Capitec Bank Holdings Ltd and Another v Coral Lagoon Investments 194 (Pty) Ltd and Others (470/2020) [2021] ZASCA 99, the Supreme Court of Appeal reiterated that interpretation concerns the relationship between the words used, the concepts they express and the place of the provision within the scheme of the agreement as a whole. Text, context and purpose must be considered together, with the language of the provision remaining the point of departure.

That principle has a practical lesson for AI-assisted drafting: the legal effect of a provision cannot reliably be assessed in isolation. If AI adds overlapping exclusions, qualifications or remedies across an agreement, a court will interpret the resulting text as part of the entire contractual scheme. The explanation that “the AI inserted it” will not resolve the ambiguity. The parties remain responsible for the language they adopt.

AI must work from a playbook, not a blank page

The safer model is to use AI within a tested drafting methodology. A good contract playbook should record more than preferred wording. It should explain the purpose of each provision, where that topic must be addressed, the approved fallback positions, the circumstances in which a departure is permitted, and the provisions that must be checked when a change is made.

The AI should also be instructed not to repeat a rule across clauses. Where an issue is governed centrally, other clauses should cross-refer to that provision only where the cross-reference genuinely assists interpretation. This converts AI from an unrestricted clause generator into a tool operating within an established legal architecture.

Practical safeguards

• Begin with the transaction and risk allocation. Identify the parties, services, data flows, intellectual property, payment mechanics, regulatory constraints and material failure scenarios before drafting.
• Use an approved precedent for the relevant agreement type. AI should adapt a known architecture rather than assemble a contract from unrelated clauses.
• Give AI a clause map. Specify where liability, indemnities, confidentiality, data protection, intellectual property, suspension, termination and dispute resolution are to be dealt with.
• Require a whole-document consistency review. Check defined terms, hierarchy, cross-references, remedies, survival, liability treatment and duplication after each material revision.
• Require reasons for proposed changes. The output should identify the risk, the existing provision that addresses it, why that provision is inadequate and where the change belongs.
• Prohibit invented authority. Every case, statute or regulatory statement used in the review must be verified against an authoritative source.
• Retain human legal review. The responsible lawyer must decide whether a clause reflects the law, the commercial bargain and the intended architecture of the agreement.

Conclusion

AI can make contract drafting faster, but speed is not the same as quality. The real skill in drafting lies in understanding why a provision exists, how it interacts with the rest of the agreement and where each risk should be addressed.

Without substantive knowledge and a disciplined playbook, AI tends to over-explain, over-qualify and repeat. With appropriate training, approved precedents and experienced human oversight, it can instead support the drafting of agreements that are shorter, clearer and internally consistent.

Source note

Capitec Bank Holdings Ltd and Another v Coral Lagoon Investments 194 (Pty) Ltd and Others (470/2020) [2021] ZASCA 99 (9 July 2021), Supreme Court of Appeal of South Africa. The article also reflects DKVG’s established drafting principle that each rule should be stated once and cross-referenced rather than duplicated.