AI Chatbots That Promise Delivery Dates They Cannot Know
A shopping assistant that answers when will it arrive with a confident date, drawn from nowhere, creates a promise the store then has to keep. Where that answer should come from, and when the bot should refuse.
Ask a store's AI assistant when an order will arrive and there is a fair chance it will give you a date. Not a range from the shipping policy, not a calculation from stock and carrier, a date, delivered in the same confident tone it uses for everything else. If the underlying model was never connected to inventory, fulfilment location, or carrier cutoffs, that date is invented. The customer will still treat it as a promise, screenshot it, and quote it back to support when the parcel is late.
Delivery estimates are a data problem wearing a language problem's clothes. A model that is good at sounding helpful will fill a gap rather than admit one, unless it is built and instructed to do the opposite. Stores deploying assistants on product and checkout pages need to decide which questions the assistant is allowed to answer from connected data, which it answers from policy text, and which it hands to a human or declines.
Three kinds of delivery answer
The first kind comes from a system of record: this item, in stock at this location, with this carrier service, ordered before this cutoff, is estimated for this date range. That is a calculation, and an assistant can present it if it is genuinely wired to those inputs and shows a range rather than a single day. The second kind is policy: standard delivery is three to five business days. That is publishable text, and the assistant can quote it as policy, attributed as policy.
The third kind is the dangerous one, where the assistant blends policy, optimism, and a holiday it half remembers into a specific date for a specific order. Ban that pattern explicitly in the assistant's instructions, test it with prompts designed to elicit it, and log the conversations where it happens anyway. Language models drift back toward being agreeable, so a rule written once in a system prompt is a starting point, not a guarantee.
Holidays, cutoffs, and the questions shoppers really ask
Delivery questions spike around cutoffs. Will it arrive before Christmas, before a birthday, before Friday. Those questions have a correct answer that depends on the current date, the destination, and stock, and they are the worst questions to improvise. Give the assistant the actual cutoff table for the season, in structured form, and make it answer from the table or say it cannot confirm and point to the checkout estimate.
Also decide what happens after purchase. An assistant that can look up an order should use carrier tracking data and say what the carrier says, including when the carrier says nothing new for three days. An assistant that cannot look orders up should not speculate about where a parcel is. I do not have access to your order status is a better answer than a plausible journey invented for a tracking number it never read.
Testing before the peak, not during it
Build a test set of delivery prompts: in stock and out of stock items, a preorder, a split shipment, an international address, a PO box, a holiday cutoff asked the day after it passed. Run it after every change to the assistant's instructions, model, or connected tools, and score the answers against the shipping policy and the stock file. Delivery accuracy is measurable in a way that general helpfulness is not, so measure it.
Keep a human escalation that is easy to find from the chat. A shopper arguing about a date the bot invented is already frustrated. Making them repeat the story to an agent who cannot see the chat transcript adds a second failure to the first. Stores that hand over the transcript, and honour a date the assistant gave in error where they reasonably can, spend less on these cases than stores that argue the bot was not authorised to promise anything.
Related reading: AI Commerce Watch, Answers, Glossary.
Frequently asked questions
Why does an AI assistant invent delivery dates?
Because it is trained to produce a helpful sounding answer, and a specific date sounds more helpful than a policy range or an admission that it does not know. Unless it is connected to inventory and carrier data and instructed to refuse when that data is missing, it will estimate in prose.
Should a shopping assistant give a date range or a single date?
A range that reflects the carrier estimate and fulfilment time, attributed to its source. A single date implies a precision most supply chains do not have, and it converts a routine one day delay into a broken promise in the customer's mind.
How often should assistant answers be audited?
Sample conversations weekly, and run a fixed test set after any change to instructions, model, or tools. Look specifically for answers that state facts the assistant has no source for: dates, stock levels, prices, and policy terms are the four categories that cause real disputes.
Is the store responsible for what its chatbot says?
Customers will treat it as the store speaking, and regulators in several jurisdictions take a similar view of claims made by a company's automated tools. Treat assistant output as published copy: sourced, tested, and corrected when it is wrong, rather than as somebody else's improvisation.