Only for what they send or approve. A customer support agent who sends a reply the AI agent drafted owns that reply, including any promise in it. A refund they approve is theirs too. Replies the AI agent sent alone within approved bounds belong to whoever set those bounds. The record shows which is which.
Who is accountable for what in a support team?
The word "agent" covers two very different things here, so it helps to be exact. The AI agent drafts replies, looks up orders, issues small refunds within limits and holds anything larger. The human support agent reviews and sends drafts, approves held refunds, and takes over conversations the AI agent escalates.
Drafting a reply that a person sends is depth 2 work. Issuing a small refund on its own is depth 3. Neither moves the accountability for a sent message away from the person who pressed send. That is the deal: the AI agent saves the typing, the person keeps the judgement.
What may the AI agent do, and what stays with the person?
- The AI agent may: look up order status, draft replies from approved knowledge, issue refunds below the team's limit, and tag conversations.
- Held for a support agent: refunds over the limit, replies that make commitments about delivery dates or compensation, and any complaint that mentions harm or legal action.
- Refused: changing account ownership or contact details without verification, and making promises outside policy.
- The support agent keeps: reading the customer's tone, deciding on goodwill beyond the script, and spotting a customer in distress.
What does the record show about who decided?
Every draft the AI agent produces is filed with its sources and the verdict. If a reply quoted the returns policy, the record shows which version of the knowledge article it used. Order data from the order system is sealed. A customer's claim that a parcel never arrived is declared. A courier's delivery scan is signed.
When a support agent sends a draft, the record shows whether they edited it and how long it was open. A promise of next-day delivery that came from the AI agent's draft and was sent unedited is the sender's promise. A promise the AI agent made alone, inside bounds, traces to the bounds.
An example: the delivery date that was not real
Take an online homeware retailer. A customer asks where a sofa is. The AI agent drafts a reply saying it will arrive on Friday, using a delivery estimate that was declared by the warehouse team a week earlier and never updated. The draft marks the estimate's age. The support agent sends it in seconds.
The sofa arrives the following Wednesday. The customer complains and asks for a 150 refund on delivery charges. The record shows the stale estimate, its label and the unedited send. The support agent is accountable for the promise. The team lead is accountable for a knowledge source nobody kept current.
What controls keep it fair?
- Show source age on every draft. A support agent should see at a glance how old a fact is.
- Hold commitments. Any reply promising a date or compensation should wait for a person.
- Train before granting send rights. A two-week practicum for customer support agents, logged in a practicum log template for customer support agents, builds the habit of reading before sending.
- Keep pace with the AI agent. Its certificate lapses after 90 days and knowledge changes weekly; see how to prepare customer support agents for 90-day recertification.
The people who own the team's knowledge and limits face a different version of this; see support team leads and AI agent mistakes. The wider topic is in AI literacy and certification for people who work with agents.
Where the Academy fits
The Foundations track at Pryme Intelligence Academy takes about four hours and gives every support agent the vocabulary of depth, held actions and the decision chain. Those who approve held refunds every day fit the Operator track, which adds human review, reading the workspace and keeping knowledge current, tested in their own industry and role.
Questions people ask
Is a support agent accountable for a reply they did not read?
Yes, if they sent it. Sending is approval. That is why teams should slow down sends on drafts that carry commitments.
What if the AI agent made the promise in a live chat on its own?
Then it acted within bounds someone set. The record names the bound, and the person who approved it owns the outcome.
Should support agents be able to raise refund limits?
No. Limits are a team lead or manager decision, held and filed like any other change to the AI agent's bounds.