When Does an AI-Generated Quote Become a Commitment by the Business?

Geoffrey Chen

After AI Enters the Workflow · Season Three: “When AI Starts Acting for the Organisation” · Article Three

“It should cost about $10,000.” “The current price is $9,800.” “This formal quotation remains valid for fourteen days.” “We accept your proposed price of $9,000.”

All four sentences contain a price, but they do not perform the same act. The first is an estimate. The second may provide current information. The third invites reliance for a defined period. The fourth may complete a negotiation.

People ordinarily distinguish these statements through role, document format, context and wording. Once AI enters quotation work, the language may look the same while the production path changes. A system may retrieve a price list, apply a discount from customer data, select a validity period and send the result from a sales mailbox.

The important question is no longer only whether AI calculated accurately. It is the point at which the business turned price information into a commitment on which the customer could rely.

One price can cross five levels of authority

A quotation system needs more than an on-or-off permission to generate. Outward pricing activity can be divided into five levels:

Level What AI does How the customer may understand it Required control
1 Displays a published list price General information Version and scope
2 Calculates an estimate from inputs Non-guaranteed prediction Assumptions, tax, fees and error range
3 Drafts a quotation for employee review A proposal not yet issued Human price and authority check
4 Issues a formal quotation within fixed conditions Reliance during the stated validity Limits on value, discount and period
5 Changes terms and accepts a counter-offer Negotiation and commitment Explicit authority, escalation and transaction record

Risk often arises when a service is described internally as level two but operates externally as level four or five. A page may call itself an “instant estimate”, while the downloaded document carries the company’s letterhead, a unique quotation number, a payment link and an “accept” button. Customers interpret the whole transaction design, not the product label chosen by the internal team.

Quotation authority must therefore be defined by external effect rather than whether the company calls the tool a calculator, assistant or agent.

An accurate price can still be unauthorised

Treating error only as miscalculation misses three deeper problems.

The first is source applicability. AI may read an old price list correctly without recognising that it applied only last quarter. It may apply a discount exactly while failing to see that the customer did not meet the minimum volume.

The second is the origin of discretion. A discount in historical data may have been an exception made by a sales manager for one relationship, not a general formula. Learning the pattern does not grant authority to repeat the exception.

The third is authority to commit. The number and terms may be commercially reasonable while the sender still lacks authority to bind the business. Accuracy asks whether the price was calculated correctly. Authority asks who may make that condition the company’s position.

The distinction also matters in dynamic pricing. A system may respond to inventory, time, demand or service cost, but the business must decide which variables are legitimate pricing inputs, which personal characteristics should not affect price and when a customer should know that a price can vary. A generative model cannot turn correlation into pricing policy by itself.

Automation does not place the transaction outside ordinary law

Australia’s Electronic Transactions Act 1999 provides a technology-neutral framework for electronic communications, signatures, dispatch, receipt and attribution. Federal Register of Legislation, Electronic Transactions Act 1999 Whether a quotation is an offer, a click constitutes acceptance, or a person or system has actual or apparent authority depends on the facts and applicable law. “Generated by AI” is not a universal answer.

The United Nations Commission on International Trade Law adopted a Model Law on Automated Contracting in 2024. It addresses legal recognition, attribution and error in transactions involving automated systems without treating the machine as a new contracting person. UNCITRAL, Model Law on Automated Contracting Adoption and implementation must be checked jurisdiction by jurisdiction, but the underlying problem is clear: a business that chooses to form transactions through automation cannot thereby place them in an attribution vacuum.

Technical teams should not decide alone that a system “only recommends” and therefore “cannot quote”. External documents, customer journeys and authority settings must be aligned by business, legal and technical owners.

The customer sees an overall impression, not a database field

A quotation may carry estimate_only=true in a database. The customer does not see that field. The customer sees brand, precise amount, expiry date, a salesperson’s name and a payment button. If small print says “for information only” while the primary design urges immediate acceptance, the internal label does not determine external understanding.

The Australian Competition and Consumer Commission’s price display guidance says businesses must display the minimum total price a consumer must pay to obtain a product or service, including specified charges. ACCC, “Price displays” Its guidance on false or misleading claims emphasises the overall impression created. ACCC, “False or misleading claims” Compliance in a particular AI quotation depends on context, but automation does not reduce the requirement that price communication be clear, accurate and not misleading.

A useful interface should let a customer distinguish estimate from quotation, identify included and excluded items, see the validity period, recognise inputs supplied by the customer, understand what triggers recalculation and know whether the next click applies, reserves or accepts.

Give AI conditional permissions, not a general “right to quote”

An organisation can express authority as an executable matrix rather than the sentence “the sales AI may quote”:

Standard catalogue item + published price
→ automatic display permitted

Approved discount range + complete data
→ automatic issue permitted, with a record

Discount above limit / non-standard term / conflicting information
→ designated manager approval

Liability change, exclusivity, long price lock or material value
→ no AI commitment; formal review required

The matrix should include cumulative risk. One discount may be small, but the same mistake sent to thousands of customers in a day can create material exposure. Value, volume, time window and anomalous pattern need to operate together.

Authority must also bind the relevant identity. Which legal entity, business unit, region, currency and tax regime does the system represent? AI should not infer those facts from brand similarity. A shared model across a corporate group can easily apply one entity’s conditions to another.

A formal quotation needs a reconstructable commitment record

When a customer accepts, the company should be able to reconstruct the full conditions at the time rather than preserve only the AI’s final prose. The minimum record includes:

  • quotation number, time and relevant legal entity;
  • price list, business rules and versions used;
  • material customer inputs;
  • fields proposed or changed by AI;
  • results of automated authority checks;
  • human approver and approval scope; and
  • the final version shown to and accepted by the customer.

This record separates input, calculation, permission and communication errors. It also allows the institution to withdraw a repeated error, identify other customers and correct the system.

If AI negotiates, the record must also preserve counter-proposals and each change of condition. Keeping only the final document loses evidence needed to determine whether the system crossed its authority or treated an exploratory statement as acceptance.

“Human approved” must be more than clicking a finished email

A reviewer needs to see what changed rather than reread an entire fluent document. The interface should highlight deviation from the standard price, the discount rule used, terms added by AI, missing information and the obligation created on acceptance.

Standard low-risk quotations may use automatic approval with sampling. Boundary conditions may require a second confirmation. Irreversible or material commitments should remain with a person holding the relevant delegated authority. This is not a claim that people are always more accurate. It locates commercial discretion and responsibility where they can be identified.

Review load must also be controlled. If sales staff must approve dozens of complex quotations each hour to meet response targets, the nominal human stage will not produce judgement. The organisation must narrow automated quotation or increase review capacity; it should not use an employee’s name to certify unreviewable volume.

Correcting the price is not enough after failure

If AI issues an unauthorised low price, the organisation should first stop further generation, identify issued and accepted quotations and place a person with authority in charge. Communication to customers should state which condition was wrong, what options now exist, how earlier payment or arrangements will be handled and who owns the follow-up.

Whether the business must perform, may withdraw or owes compensation is a fact-specific legal question and should not be answered automatically by the system that failed. The company should not begin by shifting responsibility to customers: “You should have known it was AI.” If formal channels and design encouraged reliance, the institution must address the reliance it created.

Nor should an incident review end with “hallucination”. The real cause may be ownerless price data, an obsolete version, uncoded discount policy, an authority check that occurred after sending or a customer interface that dressed an estimate as a formal quotation. Only repair of the chain resolves the failure.

Conclusion: quotation capacity must remain smaller than commitment capacity

AI can reduce the time needed to estimate, compare options and prepare documents. The stronger its language and calculation become, the more deliberately an organisation must design the commercial and legal status of the output.

The final principle is:

AI may calculate every price the organisation permits it to calculate, but it may commit only within explicit boundaries of entity, value, term, duration and exception. Anything a customer can rely upon must have an authority record the organisation can reconstruct and is prepared to honour.

A good quotation is more than a correct number. It identifies conditions, duration, identity and method of acceptance, and lets the company explain why it had authority to make the commitment. AI can make quotation faster. It cannot decide for the business what is worth promising.

Primary sources and further reading

Continue reading: Explore the After AI Enters the Workflow series.

Discover more from Geoffrey Chen

Subscribe to get the latest posts sent to your email.