AI ticket triage: all three got the labels right, but “unknown” should not mean last

Our choice is Claude’s parallel clarification

Claude left F unassigned with priority pending, then proposed sending questions alongside the urgent work. These are separate decisions: assigning severity requires facts, but gathering facts need not wait until every ordinary inquiry is resolved.

ChatGPT explained the rules clearly, including why eleven affected people do not automatically mean P0. Gemini’s concise list was easy to scan. We are not manufacturing a classification winner when all three matched the same seven expected pairs.

WekeyLab AI would combine Claude’s parallel follow-up with ChatGPT’s explanations, while recording proposed, approved and completed actions separately.

The fixed rules behind the seven tickets

The fictional software team routes suspected access to another customer’s data to Security at P0, even before confirmation. Other tickets get a team by topic and a priority independently. Ten or more people completely unable to work with no workaround means P0; one to nine in the same situation means P1; an available workaround or an inquiry means P2.

F says only “It does not work. Fix it quickly.” The feature, affected count, degree of blockage and workaround are unknown. D repeatedly says “urgent,” but asks how to change an invoice name and reports no blocked work. Urgent wording alone does not change the supplied rules.

We fixed the expected seven assignments and the no-merging/no-resolution-promise checks before submitting the same unrestricted Chinese prompt once per service. A, B and C were independent reports, not a confirmed common incident. This English edition localizes that run.

All three matched these seven assignments

A → Technical/P0: exactly ten people cannot export their daily report, their work is fully blocked and no alternative exists.

B → Technical/P1: nine people are blocked with no alternative. Do not add A’s count to cross the threshold.

C → Technical/P2: eleven people cannot export automatically, but manual download is confirmed to complete the same job.

D → Billing/P2: an invoice-name how-to request with no blocked work.

E → Security/P0: possible exposure of another customer’s address overrides the fact that it appears on an invoice page.

F → Unassigned/Pending confirmation. G → Customer Support/P2: a how-to question about adding a report column while work remains possible.

We extracted team and priority from each original table and compared them with the fixed reference: seven matches per service. This is the result for seven synthetic cases, not a general accuracy rate or evidence that a helpdesk has resolved them.

ChatGPT explains the labels, then puts F at the end

ChatGPT opened with security override, topic routing and independent priority assessment. Its handoff summary makes the counterintuitive eleven-person-versus-nine-person result easy to explain to a colleague.

Its table says to contact F first for clarification, but the ordered narrative places F after the P2 tickets under “Last: F.” It did not label F P2 or invent its impact. The operational concern is that someone following the sequence might finish ordinary inquiries before discovering whether F is urgent.

WekeyLab AI would change that instruction to “ask in parallel and reclassify on receipt.” Unknown severity is not evidence of low severity. This is an editorial workflow judgment, not a claim that the correct label was wrong.

ChatGPT: “最后：F” — “Last: F.” The actual label remains unassigned/pending.

Gemini’s account-isolation suggestion needs a separate decision

Gemini put E and A first and asked the right questions for F: which feature, how many users, complete blockage and possible alternatives. It said to initiate those questions first, so its answer does not simply bury F at the end.

For E, however, it recommended immediately isolating access for the relevant account. That is an additional proposed response measure. The supplied policy authorized Security/P0 routing; it did not specify a support agent’s authority to isolate accounts or the scope of any restriction.

WekeyLab AI would write “Security takes ownership and assesses whether access restrictions are needed and at what scope.” Immediate routing and a particular containment action are different decisions. Nothing in this test establishes that isolation happened or that it was the only valid response.

Gemini: “立即隔离相关账号访问” — “Immediately isolate access for the relevant account.”

Claude separates primary ownership, collaboration and suggestions

Claude kept Security as E’s one primary owner, with Billing and Technical as possible collaborators. An invoice page did not turn a suspected security problem into a Billing-owned ticket. It also explained that F could turn out to be P0/P1 or merely an inquiry, supporting parallel questions.

Its C→D→G ordering within P2 was explicitly its own suggestion, not an invented company rule. That distinction makes the draft easier to review and adapt.

The security questions include whether the issue can be reproduced, while its final note says not to ask the customer to repeatedly view other people’s data. A clearer handoff records what was already observed and sends it through the team’s approved channel, leaving the investigation to Security.

WekeyLab AI’s reusable handoff

Start now: E→Security/P0 and A→Technical/P0. Describe E as suspected cross-customer exposure, not confirmed leakage. Keep A’s ten-person count as its own report.

Ask in parallel: F stays unassigned/pending; the on-duty support agent follows up for the missing facts. That follow-up owner is our operational addition, not a premature choice of F’s final resolving team.

Then handle B→Technical/P1, C→Technical/P2, D→Billing/P2 and G→Customer Support/P2. C’s manual download is an already confirmed workaround, not a feature the AI has repaired.

For every ticket, record the customer’s words, known impact, missing facts, primary team, priority, rule basis, next step, follow-up owner and actual execution state. Keep A/B/C separate until Technical establishes a link. Do not fill in an unapproved resolution deadline.

Draft question for F: Which feature is unavailable, how many people are affected, is the work completely blocked, and is there a temporary workaround? If there is an error, please include its text and when it occurred. We will use those details to confirm the responsible team and priority.

A prompt you can reuse

“Apply these internal rules to the tickets. Give one primary team, priority, factual basis and next action. Leave insufficient information pending and list follow-up questions. Separate required actions, your recommendations and completed actions. Do not combine unconfirmed related reports or promise resolution times.”

After routing, the refund-email comparison covers a customer-facing reply, and the meeting-action comparison covers owners and deadlines in team handoffs. Those are subsequent tasks: correct triage is not completed resolution.