Sometimes. The permission comes from the customer and the offer book. An AI support agent can use purchase history to suggest a pre-approved offer if the customer allows personalised offers and the conversation is not about a problem. It must not invent discounts, sell to a customer who is complaining or struggling, or turn support into sales.
Which conditions decide it?
What the customer agreed to. Using purchase history to tailor marketing is profiling for marketing purposes. Whether that is allowed depends on what the customer consented to or was told, and on any opt-out they made. The agent reads the preference flag before it considers an offer. If the customer opted out of marketing, the answer is no, however good the offer.
Whether the offer exists. The agent can only present offers from an approved offer book, with its eligibility rules, amounts and end dates. It cannot compose a discount because the customer seems likely to buy. A made-up offer is a promise, and the business may be held to it; the risk is set out in the promises your support agent must never make.
Why the customer got in touch. A customer chasing a late delivery or a wrong charge did not come to be sold to. Offers in that moment feel like deflection and damage trust. Offers fit when the customer asks for one, asks about a product, or has finished a routine task and the agent offers something plainly relevant once.
The customer's circumstances. If the conversation contains signs of financial difficulty or distress, no offer is made, and certainly no offer of credit.
What may the agent do, and what stays with a person?
| Situation | The agent | Stays with a person |
|---|---|---|
| Customer asks what offers they qualify for | Reads eligibility from the offer book and history; presents matching offers | Offers outside the book |
| Routine contact resolved, marketing allowed | May mention one relevant offer, once | Nothing |
| Customer complaining | Resolves or escalates; makes no offer | Goodwill gestures |
| Customer says they want to leave | Applies the retention rules only | Anything beyond them |
| Customer asks for a better price | Restates available offers | Negotiation |
| Offer involves credit or a regulated product | Gives product information only | Suitability and sale |
The leaving row is its own capability with its own limits, as set out in making retention offers within set limits. The reasoning for keeping it rule-based rather than discretionary is in retention offers need limits, not judgement.
An illustrative example
Take an online pet supplies retailer. A verified customer asks the chat agent to change the delivery address on a repeat order of dog food.
The agent changes the address. The customer's preferences allow personalised offers, and the offer book has a current rule: customers who have bought the same food three times may subscribe for a fixed discount. The customer qualifies. Once the address change is confirmed, the agent mentions the subscription offer in one sentence with its terms. The customer declines. The agent does not raise it again in that conversation, and records that it was offered and declined.
A second customer, chasing a missing parcel, asks "is there anything you can do for me?" The agent resolves the parcel issue first. The offer book has no goodwill rule for late deliveries, so it does not invent one. It refers the request for a goodwill gesture to a person, because that is an exception, and exceptions to company policy are not the agent's to grant.
What controls make it safe?
- Preference check first. No offer without reading the customer's marketing and profiling preferences, every time.
- Offer book only. The agent selects from approved offers. The amount, terms and end date come from the book, not the model.
- One mention. A declined offer is not repeated in the same conversation.
- Context blocks. Complaint, distress and difficulty signals switch offers off.
- Channel rules. Offers follow the channel the customer is in, and outbound offers need their own consent; see which channels a support agent may use.
- A record. Every offer shown, accepted or declined is filed with the rule that made the customer eligible.
Where Pryme fits
In Pryme, "present an offer" is a separate capability from support, with the offer book as a sealed source and the customer's preferences in the minimum chain. If preferences or context block an offer, the verdict says so. More is in the AI support agent that knows when to say no.
Questions people ask
Can the agent use purchase history to answer support questions?
Yes. Reading order history to resolve a problem is support, not marketing, and does not need marketing consent.
Can the agent tell a customer what other customers bought?
Only in general product terms. It must not reveal anything about another identifiable customer.
Who writes the offer book?
The marketing or commercial team, with sign-off from whoever owns pricing. The agent reads it; it never edits it.