Illustrative photograph. Photo: Royal Society library books.jpg by Mike Peel, via Wikimedia Commons — CC0 / public domain. Downloaded Wikimedia version; no custom edits.
An assistant answers a customer confidently: yes, the service includes weekend support. The team delivering the service says it does not.
The problem began before the conversation. Somewhere, the agent received ambiguous, stale, or incomplete information, or failed to respect a boundary.
Whatever platform you choose for a business agent, prepare the answer system before polishing its tone. The workflow below is platform-independent; it does not assume access to the reported ChatGPT Business Hub.
Start with questions that affect a purchase
Take the questions your team repeatedly answers before a sale or booking. Group them into offer, eligibility, price, delivery, policies, and escalation.
Choose an initial set small enough that a named person can review every answer. Include difficult questions, not just convenient ones.
Examples include: “Can you guarantee this outcome?”, “Will this work with our current system?”, and “Can you make an exception to the cancellation policy?”
Store more than an answer paragraph
Each entry should contain the question, approved answer, source, owner, verification date, relevant conditions, and escalation rule.
For a fictional service business:
Field | Example |
|---|---|
Question | Can I book outside normal opening hours? |
Approved answer | Requests can be submitted; availability requires team confirmation |
Source | Current operating policy |
Boundary | Do not confirm a slot without verified availability |
Escalation | Send the request through the approved booking route |
Owner | Operations manager |
The boundary may be more important than the wording. “We can check” and “you are booked” have different consequences.
Separate explanation from authority to act
An agent may be allowed to explain a refund policy without being allowed to grant a refund. It may collect requirements without committing to a quote.
Write down which actions are supported, which need a human, and which require a customer confirmation. Connect commitments to the relevant system of record rather than a static paragraph.
When live data is unavailable, the response should acknowledge the limitation and provide the approved next step. Do not let an outdated availability statement become a promise.
Test contradictions and edge cases
Evaluate each answer against realistic variations. Try an expired promotion, conflicting pages, an unsupported region, an unusual product combination, and a request to ignore the policy.
Score correctness, source support, handling of conditions, escalation behavior, and whether the reply creates an unauthorized commitment.
Include a customer who asks the same question in unfamiliar language. A library that works only when someone repeats its exact headings is not ready for normal use.
Give knowledge an expiration process
Assign review triggers to prices, offers, compatibility, and policies. When one changes, update the source and the agent’s approved knowledge through the platform’s supported workflow.
Keep a test set for consequential answers and rerun it after updates. Review customer conversations only through authorized access and appropriate privacy handling.
A business agent becomes useful when the team can explain what it knows, where that information came from, what it may do, and when it must hand over. Those decisions remain yours even when the interface changes.
Next read: Build the underlying business information register.
