How Aeion Helpdesk Actually Works
A closer look at the AI, reporting, routing, and self-service engine behind the pitch — for the technical evaluators on your team who want the details before they sign off.
Built for Speed, at Every Step
AI Ticket Analysis
Under 200ms per ticket
Auto-Resolution
Only fires above an 85% confidence bar
Knowledge Base Search
Under 100ms, understands intent not just keywords
Smart Routing
Under 50ms per assignment decision
Live Chat
Under 150ms message delivery
Reporting
A 90-day analytics report in under 5 seconds
1
A Ticket Comes In — From Anywhere
Email, your self-service portal, live chat, SMS, WhatsApp, an Instagram or Facebook Messenger or X direct message, or a phone call: they all become the same kind of ticket, in the same queue, on the same SLA clock. The conversation that started on WhatsApp keeps its channel recorded on the ticket — including which social platform, so an agent knows they're about to write an Instagram DM and not an email.
Every ticket is then read by AI: what's it about, how urgent, how frustrated is the customer. That analysis drives everything that follows — no agent has to triage it first. If the customer wrote in a language your team doesn't speak, it's detected and translated at this point, with the original always kept alongside.
Because channel is recorded rather than inferred, you can also set different SLA targets per channel. A WhatsApp message and an email don't carry the same expectation of speed, and the timer reflects that.
2
AI Decides — Resolve It, Act on It, or Route It
If the AI finds a knowledge base article that confidently answers the question, it replies and closes the ticket automatically — genuinely resolved, no agent involved.
Where it goes beyond answering: the agent can hold a multi-turn conversation and call tools you define — an order lookup, a subscription check, one of your own webhooks, or a connected system — so it can carry out the request instead of describing how. Any tool with side effects stops and waits for a human, showing the exact arguments it intends to use, which a supervisor can edit before approving or reject outright. The conversation stays parked until someone decides; nothing runs against your production systems unsupervised.
Everything below the confidence bar is routed to the right person using whichever strategy fits your team: round-robin, skill-based matching, current workload, a blend of both, or manual assignment. A support lead can also pin specific categories to specific teams — billing questions always go to the billing team, no matter what the general strategy says.
3
The Reply Goes Back Out — and Tells You If It Arrived
An agent replies once, in the ticket. It leaves on the channel the customer started: the WhatsApp thread, the SMS conversation, the Instagram DM, or their email. They never learn your internal tooling exists.
Then the part most tools skip. The reply carries a delivery receipt, and the states are kept distinct rather than blurred into one checkmark:
- Sent — a provider accepted it. That is not the same as arrival, and it isn't reported as such.
- Delivered — the transport confirmed it reached the recipient.
- Not delivered — with the reason: unreachable number, a channel that can't carry outbound replies, outbound email not switched on yet.
- Unknown — nothing ever confirmed. Said plainly, rather than shown as success.
Two channels are honest exceptions: live chat and voice tickets arrive with full transcripts and SLA tracking, but a typed reply doesn't push back into a live chat session or a phone call. The composer says so before the agent writes, rather than accepting a reply that goes nowhere — chat is answered in the chat console, voice is followed up with a callback from the in-ticket softphone.
4
Agents Get AI Backup, Not Busywork
When a ticket does need a human, the agent isn't starting from scratch. They see the AI's summary, a suggested reply already drafted, and the customer's full history — CRM record, active subscriptions, past orders, prior tickets — in the same screen. Automatic reply-guarding keeps "thank you!" and out-of-office auto-replies from reopening closed tickets and corrupting your SLA numbers.
5
SLAs That Respect Business Hours
Response and resolution timers run against your actual business hours, pause automatically while you're waiting on the customer, and account for holidays. Tickets at risk of breaching escalate before they breach, not after.
6
Customers Can Help Themselves
Your self-service portal doesn't just list articles — it understands what customers are actually asking, so "how do I get my money back" surfaces your refund policy even if the article is titled "Cancellations & Refunds." Customers can also track their own open tickets without emailing to ask for a status update.
7
Reporting That Answers Real Questions
Ticket volume, resolution time, CSAT, agent performance, SLA compliance, and topic trends are all one click away — pick a report type, a date range, and get tables and charts instantly. Schedule any report to land in your inbox daily, weekly, or monthly, no one has to remember to run it.
Screen Share, When Words Aren't Enough
For issues that are easier to show than describe, agents can start a remote screen-share session directly from a ticket. Customers join with a single click — no software to install. Every session is recorded for compliance, and remote-control access is off by default: the customer has to explicitly grant it before an agent can take over their screen.
An AI diagnostic assistant can optionally join the session too, helping the agent spot the issue faster.
Ticket Handling Built for the Real World
Tickets don't always arrive clean. Two customers reporting the same bug should probably be one ticket, not two — so agents can merge them, combining the full conversation history. A single ticket covering three unrelated issues can be split apart, each tracked separately. Related tickets can be linked together, and internal notes let agents coordinate without the customer seeing the back-and-forth.
Quality Review That Isn't Spot-Checking
Sampling rules decide what gets reviewed, so nobody picks conversations by hand. A rule can take a random percentage of resolved tickets, or target the ones that matter — low CSAT, escalated, reopened, VIP customers, long handle times, specific agents or queues. Each rule carries a daily quota so a bad day doesn't flood the review queue, and reviewers can be assigned round-robin from a pool.
Reviews are scored against a scorecard you define: named dimensions, each weighted, each with its own maximum. Mark a dimension critical and scoring it zero fails the whole review regardless of the weighted total — the compliance step that either happened or didn't. AI drafts the per-dimension scores with its reasoning; a human accepts or edits them before publishing. The final score is always recomputed server-side at publish, so the number on the record is the one the rubric produces, not one typed in.
Two things make it a development tool rather than a surveillance one. Calibration sessions have several reviewers score the same conversation independently and report how closely they agreed — if your reviewers disagree, that's a rubric problem, not an agent problem. And agents can dispute a published score, which routes to a lead who upholds, adjusts, or overturns it, with the adjustment recorded.
Reaching Customers Before They Write In
Proactive messages fire on an event (a ticket resolved, an SLA breached, a survey submitted), on a recurring schedule, or as a one-off broadcast to a segment. They go out by email, SMS or WhatsApp, or render in your portal and chat widget as a banner, modal, tooltip or inline note.
The event list is a closed set of triggers the platform genuinely publishes. That sounds like a small detail; it isn't. A free-text event field lets someone save an active message on an event nothing emits, which then never fires and never explains itself. Choosing from the wired list makes that failure impossible to configure.
Guardrails are on by default rather than opt-in: a frequency cap per recipient, quiet hours, a de-duplication window so the same message can't reach someone twice, and automatic skipping of anyone who has opted out. The authoring screen restates the whole policy in a sentence so you can check it at a glance, and it lists anything that would stop the message sending — an unset event, an audience with no usable conditions, empty content for the channel you chose — before you can switch it on.
Before a broadcast, you see how many people match and a sample of who they are. If part of your filter couldn't be applied, it says so and warns that the audience is wider than you asked for. That's the difference between a message to forty VIPs and a message to your whole customer base.
Knowing What Customers Keep Contacting You About
Contact-reason topics are discovered from the text of real tickets rather than from a taxonomy someone maintained once. Each topic carries a volume series, a trend direction, statistical spike detection, and a short forecast, so a problem shows up while it's still small. Every topic links back to the tickets behind it, and the drivers view breaks volume down by your existing categories for the questions that are better answered that way.
A Knowledge Base That Improves Itself
Every edit to a help article is versioned, so nothing is ever lost and any change can be rolled back. Behind the scenes, the system also watches for patterns: tickets everyone's asking that don't have a good article yet, and agent replies that work well enough to become the next article — turning your support history into a knowledge base that keeps getting better without extra effort from your team.
Frequently Asked Questions
Auto-resolution only fires above an 85% confidence bar. The engine reads customer sentiment first — anyone who reads as genuinely upset always goes to a human, never a bot — then searches your knowledge base by meaning, not keywords, and only replies and closes when it's confident the matched article actually answers the question. Everything below that bar is routed to an agent instead of guessed at.
Both. Agents can merge duplicate tickets into one, preserving the full combined conversation history, or split a single ticket covering unrelated issues so each is tracked separately. Related tickets can also be linked, and internal notes let agents coordinate on a ticket without the customer seeing the back-and-forth.
No. Response and resolution timers run against your actual business hours, account for holidays, and pause automatically the moment the ball is in the customer's court — resuming when they reply. Tickets at risk of breaching escalate before they breach, not after, so your SLA numbers reflect real agent responsiveness rather than customer delays.
No — remote-control access is off by default. An agent can share screen and audio and the customer joins with one click and no install, but the customer has to explicitly grant control before an agent can take over their screen. Every session is recorded for compliance, and an AI diagnostic assistant can optionally join to help spot the issue faster.
Every edit is versioned automatically, so nothing is ever lost and any change can be rolled back to a prior version. Beyond that, the system watches for patterns — recurring questions that don't have a good article yet, and agent replies that work well enough to become one — and drafts new articles from your resolved-ticket history so the KB keeps improving without extra effort.
The messaging layer receives the inbound message on your connected account, and the helpdesk creates or updates a ticket linked to that conversation — so a follow-up message lands on the existing ticket rather than opening a second one. The ticket records the channel, the platform, and which connected account owns the thread, which is what lets a reply leave from the right identity when you run several accounts.
It means the transport confirmed delivery. We keep that distinct from "sent", which only means a provider accepted the message — collapsing the two is how a broken channel goes unnoticed for weeks. If no confirmation ever arrives, the state is reported as unknown rather than shown as success, and if a channel can't carry an outbound reply at all, that's said before the agent writes rather than after they send.
Only if you decide they can. Every tool is declared by you, and anything with side effects requires human approval by default: the conversation parks, the exact arguments are shown, and a supervisor edits, approves or rejects. You can relax that per tool once you trust it. Read-only lookups can be allowed to run freely from the start.
The final score is recomputed server-side from the rubric when a review is published, so it always reflects the weights and maximums you configured rather than a number entered by hand. AI drafts per-dimension scores, a human accepts or edits each one, and any dimension you've marked critical fails the review outright when scored zero regardless of the weighted total.
A frequency cap limits sends per recipient over a rolling window, and a separate de-duplication window prevents the same message reaching the same person twice. Quiet hours suppress sends overnight, and anyone who has opted out is skipped. All four are on by default, and the authoring screen states the active policy in plain language so you can verify it without decoding individual settings.
See the Engine Behind the Pitch
Walk your technical evaluators through AI triage, business-hours SLAs, screen share, and the self-populating knowledge base — live.