A prompt that reads every new ticket against your own help content, drafts an answer that actually resolves it, and hands you only the ones that genuinely need a human — with the reasoning attached so you can see why each was graded the way it was.
Not because they are hard, but because deciding which are hard happens after the customer has already waited the longest.
Most tickets are not novel — they are a known question asked by someone who did not find the article. The knowledge is usually there, three clicks away in help content or an old thread. The cost is not knowing in advance which tickets are known questions, which means the cost is paid per ticket by searching for answers that usually turn out to be in the same place.
An agent opening ticket forty opens the conversation, reads the history, works out the product state, and writes a reply from scratch. Fine once, ruinous at volume — and the part most likely to go wrong is the slowest part, because the customer is reading it. Quality is hardest to maintain exactly where the queue is longest.
When something does reach a human, it frequently arrives as the customer's original message and nothing else: no history, no what was already tried, no reproduction attempt. The engineer or support lead then spends the first half of the conversation reconstructing what the customer has already been told, which is the slowest possible way to begin the part that needs their judgement.
Copy it once, paste it into Zaira, and it works the queue: read each new ticket, find the answer in your own help content, draft a reply that cites it, and escalate the rest with a summary of what the customer has already been told.
Three stages, one goal. The queue is worked in order of waiting time rather than arrival of whoever is free, and first response stops depending on the rota.
Because the queue is worked oldest-first on a schedule, the ticket that has waited eleven hours gets answered at eleven hours rather than whenever the next person logs in. Response time becomes a property of the system instead of whoever is free.
Every drafted reply cites the article it came from, and a ticket with no authoritative source is reclassified rather than answered. That is the difference between support that is fast and support that is fast and confidently wrong at scale.
By the time something reaches a person it has a summary, the account history, what was already tried, and the open question. The hardest cases get the most senior attention precisely because the slow part has already been done.
The practical questions people ask before letting an agent answer customers.
Not in this prompt — Step 3 writes drafts and Step 4 prepares escalations, and nothing is sent to a customer. Sending is a separate decision you should make deliberately, once the drafts have held up: run it for a fortnight reading every draft first, then consider letting the RESOLVE verdicts go out automatically while keeping escalations human. The grounding rules in Step 2 are what make that safe when you are ready.
That is what Step 2 is for. A RESOLVE verdict requires finding authoritative help content that actually answers the question, and the source is attached to the draft so anyone can check it in one click. A ticket with no source is reclassified rather than answered, and Step 3 explicitly forbids stating a price, limit, or policy that is not in your content or the customer's own record. Reading the first twenty drafts against your docs is the test.
Oldest-first, and the prompt says so deliberately. It is the order that reduces maximum waiting time and it surfaces the tickets most likely to be angry, which are the ones that generate the expensive replies. Newest-first is tempting because fresh tickets feel urgent, but it starves the old ones and produces exactly the escalation backlog you were trying to avoid.
Anything involving money changing hands, anything contractual, anything touching security, and anything where the customer has already been told something and now disagrees. Put those in your context section as hard exclusions and the agent will escalate them regardless of how well it thinks it could answer. Getting this list written down is the highest-value part of the setup — most support automation failures trace back to a missing exclusion.
A widget only ever sees the tickets that arrive through it, starts with no account history, and cannot act on anything already in the queue. This works the existing backlog in HubSpot with full thread and account context, which is where the awkward tickets live — the ones that started last week and have gone quiet. It also produces the gap report, so the help centre improves instead of absorbing the same unanswered questions indefinitely.
Zaira works through HubSpot's own official MCP server — the permissioned connection offered to approved AI partners, not a scraper or a browser rig. There is no shadow portal; it operates inside the access you authorized, and you can revoke it from your HubSpot settings at any time. Because Step 3 drafts rather than sends, nothing reaches a customer from outside your own inbox until you decide it should.