When we say August drafts from your precedent bank, we mean something specific. It does not mean the system has read some contracts and will make educated guesses. It means the first draft of every clause in a new agreement is drawn from the specific deals your firm has actually closed, in comparable transaction types, with the language your partners chose to use.
That distinction matters because the alternative, which is what most AI drafting tools on the market do today, is something quite different. Understanding the gap between the two helps explain why we built August the way we did.
What a Generic AI Drafting Tool Actually Does
Generic large language models are trained on vast corpora of publicly available text. That includes model contracts, legal textbooks, court opinions, bar journal articles, and the full range of transactional agreements that have been published, shared online, or included in regulatory filings. The model develops statistical associations between clause types, transaction contexts, and language patterns across all of that material.
When you ask a generic AI to draft an indemnification clause for a technology services agreement, it produces something that looks like a reasonable indemnification clause because it has seen many of them. The output reflects an average of common practice across the broad base of text it was trained on, weighted toward whatever patterns appeared most often in that training data.
This produces serviceable boilerplate in many cases. The problem is that it has nothing to do with how your firm has handled indemnification in comparable deals. It does not know that your practice group routinely negotiates mutual indemnification caps at two times annual fees, or that your clients in a particular industry segment have consistently accepted a carve-out for IP infringement claims. Those positions live in your closed deal files, not in public text.
What a Precedent Bank Is
A precedent bank, in the August context, is the indexed body of your firm's closed deal library. Not a curated subset selected for a form book. Not a collection of template agreements approved by a practice group committee. The actual signed agreements from completed transactions, tagged by deal type, industry, governing law, counterparty profile, and the specific clause provisions that were negotiated to closing.
When August drafts a new agreement, it queries that library. For each clause type, it identifies the deals in your bank that most closely resemble the current transaction, by type, by the parties' relative positions, by industry context, by deal size where that information is available. It then drafts from the language those comparable deals used, in the specific variations your partners chose, with the fallback positions that appeared most frequently when initial positions were not accepted.
The output is not an average of public contracts. It is your firm's actual practice, made visible and queryable.
A Concrete Example: the Limitation of Liability Clause
Consider limitation of liability in a recurring revenue software agreement. A generic AI will draft something that looks standard. Depending on which training data dominated, it might default to excluding consequential damages and capping direct damages at fees paid in the preceding twelve months, or it might produce some variation of that. The output will be defensible. It will not be wrong.
What it will not do is reflect that your firm's technology practice group has, over the past several years, consistently pushed for a mutual cap structure rather than a unilateral cap, and that in deals where the client had meaningful indemnification exposure, your partners accepted a carve-out from the cap for certain specified damages categories. A new associate drafting from the generic AI output will produce a clause that does not start from your firm's actual negotiated position. A senior associate will then spend time walking that clause back to where the firm typically lands, and marking up a draft that should have started there.
August drafts the limitation of liability clause from the comparable deals in your bank. If those deals reflect the mutual cap structure and the carve-out pattern, that is where the draft starts. The associate's time goes toward the exceptions and the client-specific negotiation, not toward re-establishing the baseline.
What This Requires to Work
Drafting from your precedent bank requires that the bank actually contain enough comparable deals to be informative. We are sometimes asked what the minimum number of agreements is to make August useful. The honest answer is that it depends heavily on deal type and how narrow the comparable transaction category is.
For practices with a focused deal flow in one or two transaction types, even a modest archive of twenty to thirty closed agreements per deal type can surface meaningful patterns. For practices with more varied deal flow, larger libraries produce more reliable results because the comparable-deals query has more to work with.
We also want to be direct about what the system is not doing: it is not deciding which past deals are legally authoritative or what language is optimal. It is identifying what your firm has actually used in comparable contexts and presenting that as the starting point. The legal judgment about whether that starting point is correct for this particular deal still belongs to the attorney.
The Relationship Between Drafting and Review
One thing we noticed early in building August is that drafting from precedent and reviewing against precedent are not really separable activities. When you draft from your bank, every clause that appears in the initial draft carries an implicit citation: this is where this language came from. That citation is made explicit in August's interface, so the reviewing attorney can see not just the drafted language but the specific prior deal it reflects.
This changes what a review comment means. Instead of marking up a clause because it does not look right, the reviewer can mark it up because it departs from the comparable deal in a specific, nameable way. That is a more useful comment. It is also a better record of why the redline was made, which matters if the deal comes back for amendment or renegotiation years later.
The same logic applies to counterparty redlines. When a counterparty proposes a change to a clause that August drafted, the system can immediately surface how often your firm has seen similar language proposed, what positions your firm took in response, and which concessions were ultimately accepted. The associate reviewing the redline has your firm's actual negotiating history available at the moment they need it, not hours later after digging through the deal management system.
Why We Built It This Way
We are not arguing that generic AI drafting has no place in legal work. For matters where a firm's historical practice is not relevant, or where the deal type is genuinely novel and no comparable precedent exists, a good generic tool can produce a useful starting point. The question is whether that starting point serves your clients well when comparable firm precedent does exist.
Most transactional practices at mid-size firms do the same kinds of deals repeatedly. The variance is in the parties, the deal-specific economics, and the particular negotiated points that come up in each matter. The underlying clause framework, the standard positions, the fallback language: that knowledge lives in the firm's closed deal history. Making that knowledge available at the drafting stage is what we set out to build.