Quick answerWhen an AI agent generates a live, individualized price quote for a specific customer rather than placing them into a published plan or tier, you take on algorithmic price-discrimination risk that a fixed pricing page does not carry. Govern it by fixing the inputs the agent is allowed to price on in advance (never protected-class proxies like location-derived demographics), logging the exact inputs and output for every quote generated, and running a regular disparate-impact check across customer segments on the quotes actually issued, not just the pricing logic on paper.
Why a live quote is a different problem than a pricing tier
A pricing-tier decision, enterprise versus self-serve, a seasonal plan, a usage-metering formula, is a published, static structure the customer opts into. Every customer who meets the same criteria lands on the same tier at the same price. A live, individualized quote is different in kind: the agent computes a number for this specific customer, in this specific moment, based on signals available to it at conversation time. Two customers asking the identical question can walk away with two different numbers, and neither ever sees the other's. That opacity is exactly what makes it a governance problem distinct from tier design, which is already covered in the pricing-tier posts in this series.
The concrete risk: proxy discrimination without intent
Algorithmic price discrimination rarely happens because anyone deliberately coded a rule like "charge this ZIP code more." It happens because a model trained or prompted to optimize for willingness-to-pay latches onto a signal that correlates with a protected characteristic, browser and device type correlating with income, referral source correlating with age, stated company size correlating with region, without anyone choosing that correlation on purpose. The output looks like ordinary price optimization until someone runs the numbers by segment and finds a pattern that would be indefensible if a human pricing manager had drawn it by hand.
Fix the input set before you fix the algorithm
The most reliable control is upstream of the pricing model: enumerate exactly which inputs the agent is allowed to condition a quote on, and treat that list as a governed artifact that legal and finance sign off on, not an implementation detail engineering owns alone. Usage volume, contract term length, and explicitly disclosed company attributes the customer volunteers are typically defensible. Anything derived passively, IP-based geography, device fingerprinting, browsing history, behavioral inference, needs an explicit justification before it enters the quote calculation, not a post-hoc defense after a customer complains.
Log every quote, not just the logic that generates them
A pricing algorithm can be fair on paper and still produce a discriminatory pattern in practice, which is why the input list alone is not enough. Every individualized quote the agent issues should be logged with the full input snapshot and the resulting number, kept in a form that can be joined against customer demographic and firmographic data after the fact. Without that log, a discrimination complaint becomes a he-said-she-said between the customer and the model's stated logic, with no way to check what the agent actually did across the broader customer base.
Run the disparate-impact check on a cadence, not just when someone complains
Treat this the same way you would treat a fair-lending compliance program: pull a representative sample of issued quotes on a recurring cadence, segment by protected characteristics you can reasonably infer or that customers have voluntarily disclosed, and check whether the distribution of quotes shows a pattern that a human reviewer would flag. This is a proactive control, not a reactive one, since by the time an individual customer notices they got a worse quote than a similar customer, the exposure already exists across everyone who asked the same question in the same window.
FAQ
Is individualized pricing illegal?
Not inherently. Individualized or dynamic pricing is legal in most jurisdictions and common in travel, ride-sharing, and B2B software. The legal exposure comes specifically from pricing that correlates with a protected characteristic, not from personalization itself.
Should the agent ever explain how a quote was calculated?
At minimum, keep an internal, auditable record of the inputs even if the customer-facing explanation stays high-level. Regulatory and legal risk depends on being able to reconstruct why a quote was what it was, not on making the full formula public.
Does this apply to discount codes and promotions?
Rule-based discounts applied uniformly to everyone who meets a stated, published condition are closer to a pricing tier than a live individualized quote, and carry much lower discrimination risk since the rule itself is auditable and applied consistently.