Draft 3.4: voice modelling on your last 200 closed deals

Two hundred is not a magic number. It is the point in our testing at which adding more closed deals to the voice model stopped changing the output in ways a partner could detect. Below about eighty, the model’s drafts still read as competent but generic. Between eighty and two hundred, they progressively acquire the firm’s habits. Above two hundred, the gains are within the noise of partner disagreement about what the firm’s habits are.

Draft 3.4 ships voice modelling built on that finding. It is available from today for accounts with a connected document management system.

What changed

• Per-firm voice model. Draft now builds a voice profile from up to 200 of the firm’s most recent executed agreements, selected by the administrator or by matter tag. The profile is used when composing any new clause or document.
• Per-practice-group profiles where the firm has enough closed deals in that group. Where it does not, the group inherits the firm profile.
• Voice is separate from positions. The voice model governs how a clause is written. The playbook governs what it says. Changing one does not change the other, and the administrator console shows them as separate objects.
• Profile inspection. Administrators can open the profile and see, in plain language, what Draft has inferred: defined-term formatting, numbering conventions, preferred modal verbs, average sentence length by clause type, the firm’s conventions for cross-references and schedules.
• Override at the convention level. Any inferred convention can be pinned or overridden. If Draft has inferred “will” but the firm has decided to move to “shall”, one setting fixes it without retraining.

What the model learns, and what it does not

It learns form. Whether the firm writes “the Purchaser” or “Purchaser”, whether definitions are collected in clause 1 or scattered, whether the firm numbers to three levels or four, whether a limitation clause opens with the cap or with the exclusions, how long a typical boilerplate sentence runs, whether “including” is always followed by “without limitation”.

It does not learn positions. If the firm’s last 200 deals all conceded a twelve-month cap, Draft does not conclude the firm wants a twelve-month cap. The playbook says what the firm wants. The closed deals say how the firm writes. We were careful about this separation because closed deals contain the other side’s wins, and a drafting tool that absorbed those as preferences would be drafting against its own client.

It also does not learn counterparty paper. Where a closed deal was drafted on the other side’s template and marked up by the firm, the voice model uses only the firm’s insertions, identified through the document’s revision history where available and through a style-consistency check where not. This is imperfect, and we describe the limitation below.

A firm’s voice is in its habits, not its concessions.

How to check its work

We recommend the following in the first fortnight.

1. Open the profile and read it. It is two pages. Most of it will be obviously right. Something will be wrong, usually because a particular associate’s quirk appears often enough in the sample to look like a convention. Override it.
2. Draft one document you know well. A standard NDA or a short services agreement. Read it as if a trainee had written it. The positions should be the playbook’s; the prose should be yours.
3. Compare against the pre-3.4 output for the same document. Draft keeps the previous version’s output available under “compare with generic voice” for thirty days after upgrade.
4. Check the sample. The administrator console lists which 200 documents were used. If any are counterparty paper that slipped through, or deals from a practice the firm has since exited, remove them and rebuild. A rebuild takes under an hour.

Known limitations

• Counterparty-paper detection is heuristic. Where the firm’s DMS does not retain revision history, Draft falls back to a consistency check that compares each document against the emerging profile and down-weights outliers. This catches most counterparty templates but not all, and a firm whose own drafting is highly variable will see more false negatives. Firms with clean revision histories get materially better profiles.
• Languages. Voice profiles are built per language. A firm drafting in English and German gets two profiles, each built only from documents in that language. There is no transfer between them. A firm with 200 English closed deals and 30 German ones will have a thin German profile.
• Recency weighting is fixed. The 200 most recent documents are weighted equally. A firm that changed its house style eighteen months ago will have a profile that blends old and new. The workaround is to restrict the sample by date in the administrator console.
• Scanned documents are excluded. Only documents with a text layer contribute.
• The profile does not yet capture conventions in schedules and annexes separately from the main body. Schedule drafting uses the main-body profile, which is sometimes too formal for a commercial schedule.

Smaller changes in 3.4

• Clause insertion into an existing document now matches the surrounding numbering depth automatically.
• The definitions panel flags any defined term used before it is defined.
• Exported documents retain the firm’s Word styles rather than Draft’s defaults.
• Generation time for a full first draft of a 40-page agreement is reduced by roughly a quarter.

Read the profile. It is the quickest way to find out what your firm’s drafting habits actually are, as opposed to what the style guide says they are.