AI پیشنهاد می‌دهد؛ چه کسی معامله را متعهد می‌کند؟

AI can generate recommendations for pricing, discounts, credit, product substitution, or delivery timing; however, a recommendation is not a commercial commitment. When the two are treated as the same thing in a sales process, a polished message can become a promise that inventory, operations, finance, or management cannot fulfill.

The issue is not simply that “a human must approve AI output.” Vague approval creates vague accountability. The healthy boundary is to establish, before any proposal is sent, who may commit each part of the deal, within what limits, on the basis of which data, and under which conditions the matter must be escalated.

This distinction matters more in industrial sales. The price of an order may depend on volume, grade, specification, payment timing, allocated capacity, shipping route, and buyer credit. AI can identify patterns across these variables, but it cannot implicitly own the risk created by a commercial promise.

The best starting point is to break a “proposal” into its commitment components. In every quotation or sales response, view at least five elements separately: price and discount, credit and payment terms, product or technical substitute, delivery date and method, and contractual or operational exceptions. Each element may have a different authority owner.

To control these elements, a four-level authority matrix is sufficient and practical. Level one is open recommendation within pre-approved limits. AI or a salesperson may prepare the response, but only within valid price, terms, and inventory information recorded in the system. This level does not permit creating an exception or inventing a new interpretation of data.

Level two is operational commitment within the deal framework. A named owner may send and record a standard proposal to the customer, provided all required data is valid and no unusual condition has been added. Sending the proposal is permitted at this level, but changing the underlying deal rules is not.

Level three is a limited exception with approval from the responsible function. A hypothetical example is a discount outside the usual range but within an approved cap, or a product substitution whose specifications have been approved by the technical owner. AI may highlight such an option, but it must not present it as a firm commitment.

Level four is a high-risk or cross-functional commitment. Credit outside policy, a delivery promise dependent on unconfirmed supply, a specification change with contractual impact, reservation of scarce capacity, or acceptance of special penalties and liabilities belong here. The decision needs a named owner, and its approval path must be recorded in the deal file.

These levels are not only for price. Every promise needs a “commitment owner.” The commitment owner is the person with final authority to say, “this term is valid on our behalf”; not necessarily the person who wrote the message, operated the AI, or spoke with the customer.

Within this structure, AI has three useful jobs: suggesting options, showing relevant data, and warning about conflicts. In a hypothetical case, it may detect that a proposed discount conflicts with the target margin, or that the delivery date in a sales response does not align with the actual supply status. A warning, however, does not replace the commitment owner’s decision.

An authority matrix is incomplete without escalation triggers. Define five clear triggers so the system and the salesperson know when to leave the standard path. The first trigger is any departure from the price floor, discount ceiling, or approved margin logic, even when AI recommends it to preserve a sales opportunity.

The second trigger is any change to credit, payment term, debt ceiling, or guarantee method. This must not be hidden in the friendly tone of a sales response. If payment terms change, the request should be escalated to the credit owner or financial-policy owner, and the result recorded before committing it to the customer.

The third trigger is any substitution of product, grade, source, packaging, or specification that could alter performance, technical acceptance, or traceability. Phrases such as “roughly equivalent” are not enough for a commitment. A substitution may be sent only when the scope of equivalence and its approval owner are clear.

The fourth trigger is any delivery promise that does not rely on confirmed inventory, reserved capacity, or an approved supplier. AI may suggest a plausible date from past lead times or order patterns, but plausibility is not authorization to commit. A customer response must operationally distinguish between an estimated date and a confirmed date.

The fifth trigger is any exception involving several functions or shifting the cost of an incorrect decision to another function. A hypothetical example is an arrangement that sales finds attractive but that disrupts production planning for operations or creates a new collection risk for finance. Such a decision requires an accountable functional owner and explicit coordination.

To bring AI into a controlled path, every recommendation should have a short decision packet. It includes the material input data, the model recommendation, the authority level, the commitment owner, the status of escalation triggers, and the final decision. The record need not be lengthy, but it should make it possible to understand later why a promise was made.

Recording the “reason for accepting or rejecting” an AI recommendation is also essential. Write the reason in operational terms: inventory data was not current, the customer had no approved credit, the margin would fall below the permitted floor, the technical substitute was not approved, or the delivery date had no reliable source. A vague record such as “manager’s view” has little value for process learning.

These records are not intended to slow the sales team. If recurring decisions are defined at levels one and two, most standard responses become faster. Real delay occurs when each person must guess from scratch whether they have authority, or when the team later enters an exhausting negotiation to correct a promise made too early.

Customer messaging must also align with the authority level. When something is not confirmed, the wording must not make it appear final. Rather than promising delivery, state the review status and who owns the timing confirmation. Rather than announcing a final discount, state the range or approval condition. Linguistic clarity is part of commercial control.

Consider a hypothetical example: AI recommends a larger discount and faster delivery to retain an order. The salesperson should not be trapped between “send immediately” and “ignore AI.” They should split the recommendation into two separate commitments: review the discount through the pricing authority path and the delivery through the inventory and operations path. One may be approved while the other is declined.

This separation prevents silent margin erosion. Many apparently successful deals are harmed not by one large discount, but by the accumulation of small exceptions: a few extra days of credit, absorbed shipping cost, an unapproved substitution, or a date that requires hidden expense to meet. The authority matrix must make that accumulation visible.

For initial implementation, start with one product, one customer type, or one sales route. List the five commitment components, define four authority levels for each, activate the five escalation triggers, and name the owner of every decision. Then test a sample of AI responses only against these rules, not against the polish of the wording.

After implementation, review should focus on decisions, not merely the quality of AI text. Examine which triggers generated the most escalations, where the commitment owner was unclear, which data was not current, and where the model recommendation was correctly rejected or needed revision. The goal is to improve data, authority, and customer messaging together.

AI becomes more useful to sales when its role is precise: recommender and risk signaler, not owner of the promise. An organization that defines the commitment owner, authority boundary, exception rationale, and escalation path before sending a proposal does not have to choose between speed and control. Speed comes from standardizing permitted decisions, while control comes from visible accountability for exceptional ones.