AI proposal drafting from a precedent library: the build

AI proposal drafting from a precedent library works when the library is small, current, and labelled, and it disappoints when the library is a shared drive of final PDFs. The build is mostly curation: choosing which past documents represent how the firm writes and prices today, breaking them into parts, and telling the AI which parts it may reuse word for word.

Partners have always drafted from precedent. They open the last proposal for a similar client, save it under a new name, and edit. The trouble is that the last one is not always the best one, and a proposal carrying a limitation of liability clause the firm retired two years ago looks exactly like the one counsel reviewed last month. An AI pointed at that drive will draw on both with equal confidence.

What is the difference between an archive and a precedent library?

A folder of sent documents is an archive, and a library is the smaller set someone has chosen, cleaned, and agreed to stand behind. The difference matters more with AI than with a person, because a senior associate knows which partner's proposals to copy and which clauses changed after the last insurer review. The AI knows only what it is given and what it is told about it.

Which past documents should go in, and which should be retired?

Start with one service line and ask the partners who sell it for the three to five proposals they would hand a new manager as the model. Add the current engagement letter and statement of work templates, in the versions counsel last approved. Everything else stays in the archive and out of the library.

Three checks decide whether a document is kept:

• It is current. The terms, the fee approach, and the service description match what the firm sells now.
• It is clean. Client names, fees, and identifying facts are replaced with placeholders, so a draft for a new prospect cannot carry another client's details.
• Someone owns it. A named partner or practice leader has said this is how the firm writes this kind of document.

Breaking precedents into reusable parts

Whole documents are hard to reuse well, and parts are easy. Split each kept document into the sections that recur, and record for each part whether its language is approved for reuse as written or is only an example of how the firm sounds.

Part	What it holds	How the AI may use it
Scope descriptions, by service	What the firm will do, stated the way the firm states it	Example of structure and tone; rewritten for each client
Assumptions	What the fee depends on (records condition, client staff time, timing)	Approved list; selected and lightly adapted
Exclusions	What is not included	Approved list; selected, not rewritten
Fee structures	Fixed, hourly, phased, retainer, and how each is presented	Structure reused; numbers come from the pricing model
Terms and conditions	Billing, termination, records, dispute resolution, liability	Approved; inserted verbatim, never paraphrased
Bios	Current team biographies	Approved; verbatim
Relevant experience	Past work, in the form each client agreed could be shared	Approved; verbatim or shortened

Approved language versus example language

This label does more work than anything else in the library. Without it, the AI treats a carefully negotiated indemnity paragraph and a partner's warm opening line as the same kind of text, free to blend and rephrase. With it, the instructions can say plainly: insert approved clauses exactly as written, treat example passages as a guide to voice, and mark every sentence that came from neither.

It is also what keeps drafts from sounding alike. The voice comes from the examples, and the substance comes from the call notes, so two proposals for two different clients share terms and bios and differ everywhere the client differs.

Proposals and engagement documents are different workflows

A proposal is written to persuade, and an engagement letter or statement of work is the document the firm will rely on if the engagement goes wrong. The Tax Adviser, citing a risk control director at CNA (the underwriter for the AICPA Professional Liability Insurance Program), reports that about 75% of claims asserted against CPA firms in the program in 2023 stemmed from tax services, and that more than half of those had no engagement letter. AICPA & CIMA describes the letter as the place that details the scope of services, the duration of the relationship, and sometimes prices or rates, often paired with a terms and conditions addendum, and notes that any task not cited in it is an expansion of service that requires a contract modification.

Those two purposes call for two different ways of using AI.

	Proposal	Engagement letter or statement of work
Starting material	Call notes plus the closest two or three precedents	The approved template and clause set
What the AI does	Writes a first draft in the firm's voice	Assembles approved clauses and fills in engagement details
Fresh writing	Expected, especially the client's situation and the approach	Kept to the scope paragraph and flagged wherever it occurs
Partner's review	Is this the right pitch, scope, and price for this client	The scope paragraph, the flagged text, and any departure from standard terms

One drafting run, from intake notes to reviewed draft

Here is an illustrative run for a proposal, followed by the engagement letter once the client says yes. It shows the sequence only and does not describe a particular engagement.

1. Intake. The partner pastes in notes from the discovery call: who the client is, what they asked for, what worried them, timing, and anything said about budget.
2. Precedent selection. The workflow names the two or three library proposals closest in service and client type and says why, so the partner can swap one out before drafting starts.
3. Draft. The AI writes the client's situation and the proposed approach from the notes, in the voice of the example passages, then inserts approved assumptions, exclusions, bios, and experience.
4. Pricing. The fee section is filled from the firm's pricing model, described below.
5. Source notes. The draft comes back in the firm's Word template with a short list beside it: which parts were inserted verbatim, which were adapted, and which were written fresh.
6. Engagement letter. After acceptance, a second workflow assembles the letter from approved clauses, carries the scope across from the accepted proposal, and flags every sentence that is not from the clause set.

Pricing from the firm's own model

A language model asked for a fee can produce a plausible number with nothing behind it. The fee in a draft should come from wherever the firm already prices work, whether that is a rate card, a fee schedule by service, or a staffing workbook with hours by level. The workflow reads the inputs from the notes, applies the firm's model, and shows the working so the partner can see how the figure was reached. Where the notes do not supply an input the model needs, the right behaviour is to leave the fee blank and ask.

What the partner reviews before anything goes out

The source notes exist so review can be aimed. A partner who knows the terms were inserted verbatim does not need to reread them, and can spend the time where judgment is required:

• The scope paragraph, read against what the client asked for on the call.
• Every passage marked as written fresh.
• The fee and the assumptions it rests on.
• Names, numbers, and facts, checked for anything that belongs to a different client.
• For engagement documents, any departure from the standard clause set, which goes to whoever the firm has decided approves exceptions.

We cover how to design that step in more detail in how to design the human review step in an AI workflow. The terms themselves are not the AI's to set, or ours. The firm's counsel and professional liability insurer decide what the clauses say, and the workflow's job is to reuse them faithfully.

One topic deserves its own conversation with counsel. Writing in the Journal of Accountancy, a CNA risk consulting director advises firms to use straightforward language to tell clients what generative AI the firm uses and how it will be used in the services, and, where tax return information is involved, to consider Regs. Sec. 301.7216-2 and whether consent is required. Whether that disclosure belongs in the engagement letter, and what it says, is for the firm's counsel and insurer to decide. Our article on the Safeguards Rule, Publication 4557, and Section 7216 goes through what those documents say.

Should you build this in your AI workspace or buy proposal software?

Drafting from a library does not require a separate product. In Claude, a project is a self-contained workspace with its own knowledge base and project instructions, and on Team and Enterprise plans it can be shared with view or edit permissions. A skill packages instructions with optional resources such as scripts and templates, and bundled reference files are read only when a task needs them, which suits a clause set and a Word template well. Anthropic's documentation lists a pre-built Word skill for creating and editing documents on the same page. How a custom skill is distributed to colleagues depends on the product and plan, so check the current help article before planning a rollout. In Microsoft 365 Copilot, Agent Builder creates agents with dedicated knowledge sources, including content on SharePoint, with the available capabilities depending on the licences in the firm's tenant. If skills are new to you, what Claude skills are explains them without the jargon.

Built this way, the library, the instructions, and the output all sit in the firm's own environment, under the access controls the firm already runs, and the firm owns every piece. That is how we build it for clients. Dedicated proposal software is worth evaluating when the firm needs things beyond drafting in one product, such as e-signature, payment collection, or pipeline tracking. If you look at one, ask where the library lives, whether it can be exported in a usable form, and whether approved and example language can be told apart.

Either way, the library holds firm and client information, so the workspace it lives in should meet the standard described in setting up secure AI workflows for confidential client data.

Keeping the library current

A library decays the day after it is finished unless someone owns it. Name one person per service line, usually the practice leader or an operations lead working with them, and give them three recurring jobs. After a win, decide whether the proposal is better than one already in the library and replace it if so. After a loss, note what the client said and whether any language should change. When counsel or the insurer revises the terms, swap the clause in the library the same day and retire the old one, so the next engagement letter cannot pick it up.

If you would like help building this against your own documents, get in touch.

Common questions

How many past proposals does a firm need before AI drafting is useful?

In our view a handful of current, well-written proposals per service line is a better starting library than a full archive, because every weak or outdated document in the library is something the draft can copy. Quality and currency matter more than volume.

Can AI draft an engagement letter?

It can assemble one from clauses the firm has already approved, fill in the client and engagement details, and flag any sentence it had to write fresh. The scope paragraph and any departure from standard terms still need a partner's review, and the terms themselves should come from the firm's counsel and professional liability insurer.

How do we stop AI from reusing another client's confidential details in a new proposal?

Clean the library before the AI ever reads it. Replace client names, fees, and identifying facts in precedents with placeholders, keep experience write-ups only in the form the client agreed could be shared, and make a check for stray names and numbers part of the review step.

Do we need proposal software, or can this run in Claude or Copilot?

Drafting from a library can run inside the AI workspace the firm already pays for, using packaged instructions, a knowledge base, and the firm's Word templates. Dedicated proposal software earns its place when the firm also needs things outside drafting, such as e-signature, payment collection, or pipeline tracking in one product.

Should the engagement letter mention the firm's use of AI?

A CNA risk consulting director writing in the Journal of Accountancy recommends telling clients, in straightforward language, what generative AI the firm uses and how it will be used in the services, and says to consider Regs. Sec. 301.7216-2 where tax return information is involved. Whether that disclosure sits in the engagement letter, and how it is worded, is a question for your own counsel and insurer.