Complex deals rarely fail because nobody could produce a summary. They fail because the summary quietly mixes facts, seller optimism, old pricing rules, unsigned promises, and one heroic spreadsheet nobody wants to question.
AI deal desk prompts can help organize verified opportunity data, identify missing inputs, compare commercial options, and prepare approval questions. They cannot approve a discount, interpret a contract as legal advice, confirm a buyer’s authority, promise a product capability, or decide how much risk the company should accept.
AI can prepare a deal for judgment. It cannot supply the judgment, evidence, or authority.
The ten templates below help deal desk teams, revenue operations leaders, sales managers, finance and legal partners, and founders review complex deals without promoting autocomplete to chief commercial officer.
What a deal desk actually controls
A deal desk is a decision system. It brings the right evidence and owners together before the company commits price, terms, scope, capacity, security obligations, or risk. The useful version answers:
- What has the buyer actually requested?
- Which details are verified, and where?
- Is the proposed price allowed under current policy?
- Which terms are standard, negotiable, or prohibited?
- What delivery, security, legal, finance, or product dependencies exist?
- Who has authority to approve each exception?
- What remains unknown before anyone signs?
The bad version is an administrative obstacle course performed at quarter-end while everyone types “urgent” in increasingly creative ways.
Keep these categories separate in every AI-assisted review:
- Verified fact: supported by a current, dated source.
- Customer statement: what the buyer said, not proof that it is true or approved.
- Seller interpretation: a reading of the situation, not the situation itself.
- Requested exception: a deviation awaiting an authorized decision.
- Scenario: a possible outcome used for comparison.
- Unknown: information still requiring discovery or specialist review.
- Decision: an explicit choice made by a named human owner.
- Restricted data: information that must not enter the selected AI tool.
A polished deal memo is still dangerous if it turns “customer asked” into “company agreed.” Language models are unusually talented at sanding off exactly that distinction.
For adjacent work, use AI sales pipeline review prompts to find stalled opportunities, AI sales forecast prompts to build a grounded number, and AI account planning prompts to prepare the wider relationship strategy.
The evidence-first deal desk prompt formula
Add this instruction to every prompt below:
“Act as a deal desk preparation assistant. Use only the sanitized, approved sources I provide for [deal, product, region, customer segment, review stage, and decision]. Cite a source label and date for every price, discount, term, capability, deadline, commitment, dependency, and policy statement. Separate verified facts, customer statements, seller interpretations, requested exceptions, scenarios, decisions, and unknowns. Mark unsupported or stale items [VERIFY]. Do not invent approvals, policy, precedent, legal interpretations, buyer motives, product capabilities, implementation capacity, or probability. Route each open issue to the named human owner.”
This makes the output less magical and considerably less likely to become evidence in an unpleasant meeting.
Never paste customer names, personal contact details, contracts, credentials, payment information, private CRM notes, confidential margins, unannounced pricing, security findings, regulated data, or legally sensitive information into an unapproved AI tool. Use approved systems, minimum necessary data, de-identification, access controls, retention limits, and human legal, finance, security, privacy, product, delivery, and commercial review.
What to collect before prompting
Build a controlled source pack. Do not paste an entire CRM export into a chatbot and call the resulting privacy incident “revenue intelligence.”
| Input | Why it matters | Human owner |
|---|---|---|
| Current opportunity record | Grounds stage, amount, dates, and scope | Account owner |
| Customer request in writing | Separates buyer language from seller memory | Account owner |
| Approved price book | Establishes valid products and list prices | Finance or commercial owner |
| Discount and approval policy | Defines thresholds and authority | Revenue operations |
| Proposed order form and terms | Shows the actual commitment | Legal and commercial owners |
| Product capability source | Prevents accidental roadmap promises | Product owner |
| Delivery and implementation plan | Tests whether the company can perform | Delivery owner |
| Security and privacy requirements | Routes specialist obligations | Security and privacy owners |
| Margin or unit-economics model | Shows the commercial trade-off | Finance owner |
| Decision deadline and reason | Distinguishes urgency from theater | Deal sponsor |
| Source dates and versions | Exposes stale evidence | Document owners |
| Restricted-data rules | Controls what may enter the tool | Security or privacy owner |
If two sources conflict, preserve the conflict. Do not ask the model to pick whichever number makes the opportunity look healthier. Named owners resolve it.
This came from a book.
Don't Replace Me
200+ pages. 24 chapters. The honest version of what AI means for your career, written by someone who actually builds this stuff.
Get the Book →10 AI deal desk prompts
Replace brackets with sanitized, approved information. Every output is a draft until authorized humans verify and approve it.
1. Check deal intake completeness
“Review this sanitized deal intake against the supplied intake requirements: [paste]. Create a table with required field, supplied value, source, source date, confidence, conflict, missing information, owner, and next action. Separate verified facts from customer statements and seller interpretations. Do not fill blanks by inference. Mark every unsupported field [VERIFY].”
This is the sensible first use of AI: finding empty boxes before twelve people join a call to discover them together.
Require sources for the buying entity, products, quantities, term length, start date, billing schedule, requested discount, implementation scope, security needs, legal changes, decision process, and deadline. “It’s in Salesforce” is not a source if the field has not been updated since the previous account owner left.
The human deal owner should resolve missing inputs with the buyer or relevant specialist. The model can identify the hole; it cannot pour truth into it.
2. Build a verified deal summary
“Using only these labeled sources: [paste], draft a one-page deal summary. Include customer request, verified business need, proposed products and quantities, list price, proposed price, term, payment schedule, delivery scope, requested exceptions, dependencies, risks, unknowns, decision deadline, and required approvers. Cite every material statement. Put seller interpretations in a separate section and mark unsupported content [VERIFY].”
A useful summary lets an approver understand the decision without excavating chat threads. It should also make uncertainty visible rather than laundering it into confident prose.
Do not describe a deal as strategic unless the organization defines that word and the evidence meets the definition. “Large logo” is not a complete business case. Neither is “the rep says they will expand later.”
For customer-facing language after approval, use AI sales proposal prompts. The proposal must reflect the approved deal, not become a parallel negotiation invented by a text generator.
3. Review a discount request
“Evaluate this requested discount against the supplied current price book, discount policy, margin guardrails, and approved precedent register: [paste]. Show list price, proposed price, effective discount, term, volume, payment timing, included services, margin impact supplied by finance, threshold triggered, required approvers, evidence offered, and missing evidence. Compare approve, revise, and decline scenarios without recommending a decision unless criteria are explicitly supplied.”
Discounts are not isolated percentages. A smaller discount with expensive custom work, generous payment terms, and broad termination rights may be worse than the bigger number everyone is arguing about.
Require finance to validate calculations and assumptions. Require commercial owners to decide whether the trade is worth making. The model should never invent a precedent because another deal “sounds similar.” Precedent needs a source, context, and authorized interpretation.
Also avoid using AI to infer how much a buyer can afford from company size, location, demographics, or personal information. Price decisions belong to approved commercial policy and actual negotiation evidence, not algorithmic fortune-telling.
4. Map the approval path
“Map this proposed deal against the supplied approval matrix: [paste]. For each price, discount, term, legal change, security requirement, data-processing obligation, implementation exception, product commitment, and payment condition, identify the rule triggered, evidence, required approver, sequence, deadline, and unresolved question. Do not treat silence, attendance, or a previous approval as approval for this deal.”
This prompt turns a messy request into a routing plan. It does not grant authority.
Approval matrices often contain thresholds, regional differences, and combinations. A discount may be acceptable until paired with nonstandard payment terms. A standard contract may stop being standard when a data-processing addendum changes liability. Ask owners to confirm the current matrix before trusting the output.
Record decisions in a durable system. AI decision log prompts can help structure the record, but the approval evidence must come from the actual approver.
5. Compare commercial scenarios
“Using these verified inputs and finance-approved calculations: [paste], compare [scenario A], [scenario B], and [scenario C]. For each, show price, discount, term, payment timing, included scope, excluded scope, margin input, delivery demand, customer concession, company concession, dependencies, risks, approval path, and unknowns. Keep calculations traceable. Do not assign probabilities or declare a winner unless explicit decision criteria are provided.”
Scenario comparison helps teams negotiate bundles rather than obsess over one number. Perhaps the buyer receives a lower price in exchange for a longer term or faster payment. Perhaps custom work is removed. Perhaps the responsible answer is no.
Do not let the model fabricate buyer acceptance. A scenario is an internal option until the buyer accepts it and authorized humans approve it. Clearly label assumptions and expiry dates because product availability, staffing, and pricing can change.
6. Surface contract, security, and delivery dependencies
“Review these sanitized requirement summaries from legal, security, privacy, product, and delivery: [paste]. Create a dependency register with requirement, source, status, owner, due date, blocker, proposed response, approval needed, and effect on scope, timing, or price. Quote no legal conclusion beyond the supplied source. Route interpretation to the relevant specialist.”
The phrase “legal is reviewing” hides a small universe. Which clause? Which entity? What decision is needed? Does security require architecture work? Does implementation have capacity on the promised date?
AI can organize specialist inputs and show where one answer affects another. It cannot provide legal advice, conduct a security assessment, or commit delivery resources. Treat generated interpretations as drafts and have each specialist validate their row.
For broad third-party evaluation, AI vendor evaluation prompts provide a useful mirror image: buyers need evidence too.
7. Separate facts from deal optimism
“Classify every statement in this sanitized deal narrative: [paste] as verified fact, customer statement, seller interpretation, requested exception, scenario, decision, or unknown. Add source, date, owner, and verification action. Flag certainty words such as ‘will,’ ‘committed,’ ‘approved,’ ‘easy,’ ‘standard,’ and ‘guaranteed’ when the evidence does not support them. Preserve useful hypotheses but do not promote them to facts.”
This is the anti-vibes prompt.
A champion saying “procurement should be fine” is a customer statement. It is not procurement approval. A salesperson believing expansion is likely is a hypothesis. It is not contracted value. A product manager saying something may fit the roadmap is not a delivery commitment.
Use this classification before forecast calls too. The goal is not to punish optimism; it is to keep optimism from impersonating evidence.
8. Prepare an exception memo
“Draft an exception memo from these verified sources: [paste]. Include the exact policy exception requested, business rationale supplied, customer concession, financial effect supplied by finance, operational effect, precedent references, risks, mitigations, expiry or one-time boundary, alternatives considered, owner, required approvers, and unresolved questions. Mark proposed language DRAFT. Do not invent precedent, mitigation, approval, or strategic value.”
A good exception memo makes the trade visible. It does not bury a permanent operational headache beneath “logo value.”
If an exception is approved, state whether it creates precedent. State who owns fulfillment. State when the approval expires. One-time custom work has a charming habit of becoming standard customer expectation unless somebody labels it aggressively.
Run an AI risk assessment when the exception affects security, privacy, delivery, finance, customer commitments, or multiple teams.
9. Draft the approval meeting brief
“Turn this verified deal record into a meeting brief for [approvers]. Include the decision required, deadline and source, five essential facts, requested exceptions, scenario comparison, material risks, specialist positions, unresolved questions, recommendation supplied by the accountable human, and explicit decisions to record. Add a parking-lot section for issues outside scope. Do not invent consensus or mark attendance as approval.”
The brief should enable a decision, not narrate the opportunity from the first cold email.
Send it early enough for specialists to inspect their sections. In the meeting, record decisions, conditions, dissent, owners, and expiry dates. If facts change afterward, reopen the relevant approval rather than assuming the old decision covers the new deal.
A summary generated after the meeting must be checked against the actual record. Fast notes are useful. Fictional consensus is not.
10. Run the final human sign-off checklist
“Review the final proposed deal against the supplied approved sources and approval matrix: [paste]. Produce a sign-off checklist covering entity, products, quantities, price, discount, term, billing, scope, implementation capacity, product claims, legal terms, security, privacy, data processing, support, renewal, cancellation, customer commitments, internal obligations, approvals, document consistency, and restricted-data handling. Cite evidence for every completed item. Mark an item complete only when explicit proof is supplied.”
This is the gate before signature, not a decorative table appended to an email.
Compare the CRM record, quote, order form, contract, implementation scope, and approval record. A correct approval attached to the wrong document version is not a correct deal. Ask the named owners to sign off in their domains.
Use an AI QA checklist for the document package, then let authorized humans make the final decision. The model gets no vote and, conveniently, accepts no liability.
A practical deal desk operating rhythm
AI-assisted deal review works best inside a visible sequence:
- Intake: collect the minimum required, current, sanitized sources.
- Classify: separate facts, statements, interpretations, exceptions, scenarios, decisions, and unknowns.
- Validate: ask source owners to resolve missing or contradictory information.
- Compare: examine commercial options using finance-approved calculations.
- Route: send legal, security, privacy, product, delivery, and finance issues to specialists.
- Decide: record explicit human approvals, conditions, and rejections.
- Reconcile: confirm every final document matches the approved position.
- Maintain: preserve the decision record and update policy from real lessons.
Start with one bounded workflow, such as intake completeness or approval routing. Do not begin by giving a model access to every contract, CRM note, and pricing workbook in the company. That is not a pilot. It is a resignation letter for your security team.
If your team needs the basics first, read the no-BS guide to using AI at work and what AI can and cannot do. The useful pattern is boring: controlled evidence, explicit labels, narrow outputs, and accountable humans.
Frequently asked questions
What are AI deal desk prompts?
They are structured instructions that help an AI system organize verified deal information, identify gaps, compare scenarios, and draft review artifacts. Good prompts define approved sources, classifications, prohibited assumptions, output format, and human owners. They do not ask the model to approve a deal from vibes.
Can ChatGPT approve a sales discount?
No. It can compare a proposed discount with a supplied policy and show which threshold appears to be triggered. Finance, sales leadership, or another authorized owner must validate the calculations, interpret the policy, assess the trade, and approve or reject the request.
Can AI review a customer contract?
AI may help organize clauses or questions inside an approved legal workflow, but it should not be treated as legal advice or an authorized interpretation. Contracts can contain confidential and legally sensitive information. Qualified counsel must review the actual language and own legal decisions.
What information should never go into an AI deal review?
Do not put personal data, credentials, payment details, private CRM notes, contracts, confidential margins, security findings, regulated information, or legally sensitive material into an unapproved tool. Follow company data-classification, access, retention, privacy, security, and legal rules. Use minimum necessary and de-identified inputs where possible.
How do we stop AI from inventing approvals or policy?
Supply the current policy and approval matrix, require citations for every rule, label missing evidence [VERIFY], and forbid the model from marking approval complete without explicit proof. Then have the named human owners inspect the output. Prompting reduces risk; it does not guarantee truth.
Is an AI-generated deal score reliable?
Not by default. Scores can hide stale CRM fields, biased history, inconsistent definitions, and missing context. Prefer transparent evidence, decision criteria, and human review. If a score is used, document its inputs, limitations, validation, and prohibited uses. Never use it as a substitute for accountable judgment.
Can AI replace a deal desk team?
No. AI can reduce administrative work around intake, classification, comparison, and documentation. It cannot negotiate trust, interpret novel risk responsibly, balance company priorities, authorize exceptions, or own legal and commercial outcomes.
What is the safest first deal desk use case?
Start with intake completeness using sanitized data and a current checklist. It has clear sources, limited scope, and an obvious human owner. Measure whether reviews arrive more complete and whether corrections decrease before expanding into scenario comparison or exception preparation.
Speed up preparation, not authority
Deal desk work contains exactly the kind of structure AI handles well: repeated fields, classification, comparison, summaries, checklists, and missing-information detection.
It also contains exactly the kind of judgment AI should not own: trust, negotiation, policy interpretation, legal risk, security obligations, financial trade-offs, delivery capacity, and authorization.
Keep the source next to the claim. Keep the customer request separate from company acceptance. Keep scenarios separate from forecasts. Keep exceptions visible. Keep approvals explicit and human.
The machine can make a complex deal easier to inspect. It cannot make the deal true, safe, profitable, or approved.
For a broader field guide to staying useful while AI makes knowledge work faster and stranger, read Don’t Replace Me by Dmitry Kargaev. The durable advantage is not typing the cleverest prompt. It is knowing which decisions must remain attached to evidence, taste, and a human name.
