Contracts and Proposals Need Opposite Voices
Same AI. Completely different personality.
The Architecture Series · 7 minute read · By Ed | Founder & CEO, DealiOS
The same AI. The same persona. The same 25-year sales veteran. The same conversation.
And then the document type changes from proposal to contract, and everything inverts.
The proposal says "approximately 40% improvement." The contract says "not less than forty percent (40%)." The proposal bans semicolons because they're stiff. The contract reinstates them for enumerated clause lists because that's legal convention. The proposal uses your brand colors throughout. The contract uses zero color. Black and white. No logos. No workspace identity.
These aren't different settings on the same dial. They're different dials entirely.
These aren't different settings on the same dial. They're different dials entirely. And if your architecture treats them as the same dial with a template swap, you're building a system that writes contracts like proposals and proposals like contracts. Both are wrong. Both are dangerous. A contract that reads like marketing copy scares the legal team. A proposal that reads like a contract scares the buyer. Both get thrown out, just by different people in different conference rooms.
The Inversion
Let me make this concrete with a side-by-side.
The proposal AI writes "approximately 40% improvement in deal velocity." The contract AI writes "not less than forty percent (40%) improvement in deal velocity." Same fact. Entirely different epistemic posture. The proposal rounds and hedges because precision signals stiffness in a sales context. The contract spells out numbers in words and numerals because ambiguity is a liability in a legal context.
The proposal uses the workspace's display font. The contract overrides to Georgia serif, because legal documents have different typographic conventions and a contract rendered in Poppins Light looks like a design student's homework, not a binding agreement.
The proposal uses brand colors throughout. Headers in the primary color. Accents in the secondary. Visual identity reinforcing that this is from you, about your offering, expressing your brand. The contract uses zero color. Black text on white paper. No logos. No workspace identity. Because a contract isn't marketing material. It's a legal instrument. Brand identity is irrelevant and potentially distracting when someone is reviewing indemnification clauses.
The proposal uses standard 1-inch margins. The contract uses 1.25 inches on the left side for binder-punching, because physical contracts still get printed, punched, and filed. An architect who hasn't worked with legal teams doesn't know this. A legal team that receives a contract with 1-inch margins notices immediately.
"SHALL" means something legally enforceable that "WILL" does not in many jurisdictions. "LIMITATION OF LIABILITY" must appear in ALL CAPS for enforceability in certain states. These aren't stylistic choices. They're functional requirements that determine whether a clause holds up in court.
The ABA issued Formal Opinion 512 in July 2024, establishing national ethical guidelines for lawyers using generative AI. [1] Multiple states including California, Florida, New Jersey, New York, and Pennsylvania have issued their own guidance. [2] The consistent message: AI can draft legal language, but it cannot advise. The distinction between drafting and advising isn't a footnote. It's the ethical boundary that separates a useful tool from a liability.
The "You Are Not a Lawyer" Constraint
This phrase isn't a footer disclaimer. It's not a legal CYA added after the contract is generated. It's an architectural constraint that changes the AI's behavior at the behavioral layer of the prompt stack.
When the AI is generating a contract and it encounters a gap, the question is: what does it do? Most AI systems do the worst possible thing. They fill the gap with plausible-sounding language. They generate a governing law clause that sounds right. They pick a liability cap that seems reasonable. They invent an insurance requirement because the training data contained thousands of contracts with insurance requirements.
This is where hallucination stops being an academic problem and becomes a legal one. In a Stanford study examining over 200,000 legal questions, general-purpose large language models hallucinated at least 75% of the time when answering questions about a court's core rulings. [3] The AI doesn't know it's guessing. It generates with the same confidence whether the clause is pulled from retrieved context or fabricated from statistical patterns.
The architectural fix: when the AI lacks a governing law, a liability cap, or an insurance requirement, it doesn't guess. It inserts [TO BE CONFIRMED] and flags the gap. The placeholder propagates downstream. The export pipeline sees it. The approval chain sees it. The DocuSign flow sees it. You can't sign a contract with [TO BE CONFIRMED] in it. That's the point. The gap is surfaced, not suppressed.
[TO BE CONFIRMED] works like a null type in programming -- it forces every consumer of the document to confront the gap and resolve it.
This is the same principle as a null type in programming. In a well-typed system, you can't accidentally use a null value. [TO BE CONFIRMED] works the same way for legal uncertainty. It forces every consumer of that document to confront the gap and resolve it. The AI's job is to generate the language it can support and mark the language it can't. The human's job is to fill the marked gaps. The architecture ensures neither party forgets.
Amendment Mode
Here's a subtlety that reveals whether a platform actually understands legal documents or just generates them.
When a contract needs revision after negotiation, what happens? Most AI systems regenerate the entire contract with the changes incorporated. This is catastrophic for two reasons.
First, the unchanged sections change. Not intentionally. But regeneration means the AI produces new text for every section, and the new text will differ slightly from the original even if the instructions are identical. Different word choices. Different sentence structures. Maybe a clause that was carefully negotiated now uses slightly different language. In a proposal, nobody notices. In a contract, a lawyer notices. And they start the review from scratch, because they no longer trust any section.
Second, the unchanged sections carry legal weight. If a counterparty has already reviewed and approved the indemnification clause, regenerating that clause with different language means they need to re-review it. Even if the change is cosmetic. The legal standard is what the document says, not what you intended it to say.
The fix: amendment mode. When a contract is revised, the AI generates only the changed sections and attaches a "Continuing Effect" clause that explicitly states all other provisions remain in force. The unchanged sections aren't touched. The counterparty's review scope is limited to the amendments. Trust in the unchanged language is preserved.
This connects to a broader principle about edit granularity matching document maturity. A first draft can be regenerated freely. A document that's been shared with a client should be edited section by section. A negotiated contract must be edited clause by clause. The further a document progresses through its lifecycle, the more surgical the edits need to be, because the stakes of unintended changes increase at every stage.
Think of it as a trust gradient. At the beginning, the AI has full creative authority. Generate, restructure, rephrase. Nobody has seen this yet. No expectations exist. As the document moves through review and negotiation, the AI's authority narrows. It can edit what you tell it to edit. It cannot touch what's been approved. By the time a contract reaches amendment stage, the AI's authority is restricted to the specific clauses being revised. Everything else is frozen.
A system that treats a first draft and a signed amendment with the same regeneration approach doesn't understand what it's building. It doesn't understand that documents accumulate trust through their lifecycle, and regeneration destroys that accumulated trust in a single operation.
Documents accumulate trust through their lifecycle. Regeneration destroys that accumulated trust in a single operation.
The Universal Principle
This isn't just about proposals and contracts. It's about any system where the same AI produces artifacts with different functional requirements.
Medical: a patient-facing summary should use accessible language, empathetic tone, and clear action items. A clinical note for the patient's chart should use precise terminology, standard abbreviations, and objective observations. Same information. Different functional requirements. Different AI behavioral profiles.
Financial: an investment proposal can use projective language and ranges ("we anticipate 8-12% returns"). A regulatory filing must use exact figures, defined terms, and prescribed formats. Same underlying analysis. Different epistemic postures.
Education: student-facing content should engage, explain, and encourage. An accreditation report should document, measure, and verify. Same institution. Different audiences. Different standards for what "good writing" means.
The architectural question is always the same: does document type change the AI's identity, or just its template?
If you're swapping templates, you're solving the surface. Same voice, same vocabulary, same epistemic posture, different header image. The output will feel wrong to anyone who reads both document types, because the person reading a contract has different expectations than the person reading a proposal. Both will feel like the AI doesn't understand what it's writing.
If you're changing the AI's behavioral profile at the feature module layer, you're solving the foundation. Different voice, different vocabulary, different epistemic rules, different formatting conventions, all triggered by the document type. The output feels right because the system understands that producing a contract isn't the same task as producing a proposal, even when the underlying deal is identical.
The Brand Decision
The deepest version of this principle shows up in the decision about brand.
In proposals, brand is everything. The workspace's colors, fonts, logo, and visual identity should permeate every page. Brand consistency builds trust. Brand expression builds recognition. The proposal should look and feel like it came from your company, because it did, and that identity is part of the argument.
In contracts, brand is nothing. The workspace's colors, fonts, and logo should be absent. Not reduced. Absent. Because a contract isn't about your brand. It's about the agreement between two parties. Visual branding in a legal instrument is, at best, decorative clutter and, at worst, a subtle signal that this document serves the vendor's interests rather than capturing a mutual agreement.
The most interesting thing an AI architecture can do with brand isn't expressing it. It's knowing when to remove it entirely.
A system that blindly applies brand to every document doesn't understand brand. A system that applies brand to proposals and removes it from contracts understands what brand is for. And that understanding, encoded architecturally rather than left to a rep's judgment, is the difference between a template engine and a platform.
The takeaway: Contracts and proposals need opposite AI personalities: different vocabulary, formatting, epistemic posture, and even brand treatment, because the same word that builds trust in a sales context destroys it in a legal one.
References
1. American Bar Association. Formal Opinion 512: "Generative Artificial Intelligence Tools." ABA Standing Committee on Ethics & Professional Responsibility, July 29, 2024.
2. State bar guidance issued by California, Florida, New Jersey, New York, Pennsylvania, Texas, Kentucky, North Carolina, and others. Compiled by Steno, "Legal AI: ABA & State Legal Ethics Guidance on Artificial Intelligence," updated 2025.
3. Stanford University legal AI research, referenced in Oregon State Bar Formal Opinion 2025-205. General-purpose LLMs hallucinated at least 75% of the time on questions about court core rulings.
This paper is part of The Architecture Series by Ed, Founder & CEO of DealiOS. For the weekly series and more technical deep-dives, visit dealios.app/library.
See how these principles come together. Explore the platform →