Somewhere in your queue right now, a ticket is asking how to reset an MFA token. It is the fourteenth one today. Next to it sits a license request, a VPN reconnect, and a question about how to request a new laptop that has been answered in the knowledge base since 2021.
A support org I worked with last quarter had 11,400 tickets a month. Sixty-two percent of them fell into nine recurring questions. They did not buy Aisera to “transform the service desk.” They bought it to delete those nine questions, and six weeks later deflection on those intents was running at 41%, with CSAT on contained tickets slightly higher than the human baseline. That gap between a boring, scoped rollout and a failed big-bang launch is almost entirely about sequencing.
Here is the sequence I would hand to a team starting Monday.
Week 1: Audit Your Tickets Before You Touch the Platform
Skip the demo. Export 90 days of tickets from your ITSM or helpdesk, and cluster them into intents. You do not need a data science team for this. Two analysts with a spreadsheet and a shared vocabulary can do it in three days.
For each intent you identify, record five things:
- Monthly volume (anything under 150 tickets a month is not worth automating first)
- Whether the answer already exists in a system of record, or has to be assembled from three places
- Whether the resolution requires an action (unlocking an account, provisioning software) or just an explanation
- Current handle time, so you have a baseline to compare against later
- Blast radius if the AI gets it wrong (a wrong VPN instruction is annoying; a wrong payroll answer is a lawsuit)
Teams that skip this step usually end up automating whatever looks impressive in a vendor demo. Teams that do it end up automating the boring stuff that actually moves the deflection number. If you want the background on how the platform turns these signals into autonomous resolution, this breakdown of how Aisera handles enterprise support tickets end to end is a useful primer before you configure anything.
Pick Three Intents. Not Ten. Three.
Take the highest-volume, lowest-risk intents from your audit and stop there. For most IT service desks, that means something close to this:
Example set: MFA token reset (2,300 tickets/month), VPN connectivity troubleshooting (900/month), and software license requests for pre-approved tools (610/month).
Each of those has three properties that make them good first candidates: a finite set of causes, a deterministic resolution path, and a system that can confirm whether the fix worked. A password reset either unlocks the account or it doesn’t. You get a clean signal on accuracy within hours.
Leave anything touching money, contracts, security incidents, or employment status for month three. Those intents carry an evaluation cost you cannot afford while you are still calibrating.
Prepare the Knowledge Base Like Your Deflection Rate Depends On It
It does. Aisera’s answers are only as good as the corpus it retrieves from, and most enterprise knowledge bases are a graveyard. Before you connect anything, run a pruning pass:
- Archive articles with no views in 12 months and no owner
- Rewrite titles as the questions users actually type, not internal jargon
- Split multi-answer articles so each one resolves exactly one problem
- Add a named owner and a review date to everything that survives
Then wire up the connectors, and switch on change notifications so a modified article refreshes the retrieval index instead of quietly going stale. This is the unglamorous work that separates a 40% containment rate from a 12% one. The service desk teams getting real results tend to treat knowledge hygiene as a permanent job, not a launch task, which is a theme worth reading up on in how AI agents are changing the service desk.
Write Escalation Rules Before Automation Rules
Most rollouts fail here. Teams spend two weeks tuning answers and ten minutes deciding when the agent should give up. Flip that.
Define hard handoff triggers first:
- Low retrieval confidence on the top matching article
- The user repeats the same question in different words within one conversation
- Detected entities like “contract,” “refund,” “legal,” or a named executive
- Frustration signals across two consecutive messages
- Any intent not in your approved three
Then set a precision target for escalations, not just a recall target. If 30% of your handoffs are things the agent could have handled, you have built an expensive router. If you are running Zendesk alongside Aisera during the transition, the cost and behavior differences between the two are worth understanding, and this teardown of where Zendesk’s AI agents break down covers the gaps you will need to bridge.
Train on Your Own Labels, Not the Default Ontology
Aisera learns intent classification from historical tickets. If your historical tickets are labeled badly, you are teaching it badly. Spend two days relabeling roughly 2,000 past tickets into your three launch intents plus an “everything else” bucket. Two people, one afternoon shift each, done.
Then look for confusion pairs. On one rollout, “VPN won’t connect” and “VPN is slow” were being merged, and the agent kept suggesting a reinstall for a problem that needed a bandwidth check. Renaming the intent and adding twelve examples fixed it in a day.
Run a Shadow Week Before Anyone Sees It
For five working days, let incoming tickets flow through the agent’s routing and suggested-answer pipeline while humans still write the replies. Nobody on the outside knows.
Score every interaction on one question: would this answer have resolved the ticket as written? Tally the failures into three buckets, wrong answer, right answer at the wrong moment, and no answer offered. Do not go live until 80% of suggested responses clear that bar on your launch intents. One team I know went live at 55% because an executive had a launch date in mind. Reopen rates doubled and they spent the next month rebuilding agent trust, which is far harder than delaying a week.
Launch to 20% of Traffic, With a Kill Switch
Pick one channel, one region, and about a fifth of eligible volume. Keep the switch that turns containment off in a place a support lead can reach without a change request.
For the first week, review the ten worst interactions every morning as a group. You will find that most are not model failures but missing knowledge or a bad escalation rule. Week-one numbers worth tracking:
- Containment rate on launch intents (expect 30 to 45%, not 80%)
- CSAT on contained tickets versus human-handled, measured separately
- Escalation precision, meaning handoffs the human agreed were necessary
- Reopen rate within 72 hours, your best early warning signal
Then add intents three at a time, running the same shadow loop for each. Employee IT first, HR second, customer-facing last, because the tolerance for a wrong answer rises sharply once an external customer is reading it. If you are heading toward customer support, how Sierra AI structures enterprise customer support agents is a fair comparison point for the evaluation bar you should hold yourself to.
Know When to Build the Workflow Yourself
Aisera is strong at conversational resolution. It is not the right tool for a provisioning request that needs twelve API calls across four systems with rollback on failure. For those, orchestrate the agent yourself and let Aisera handle the front door. Lightweight Python frameworks have gotten good enough that this is a two-week job rather than a quarter, as shown in this look at Agno, a framework built for agents you actually ship. Draw the line clearly: conversation and retrieval in the platform, multi-step system choreography in code you own and can test.
The Failure Modes That Sink These Projects
Nearly every stalled rollout I have seen traces back to one of five decisions:
- Launching with 30 or 40 intents because the demo handled them all
- Nobody owning the knowledge base after launch week
- Reporting deflection alone, with no CSAT or reopen data beside it
- Ignoring front-line agents, who will quietly tell users to skip the bot
- No documented path back to human-only handling if accuracy slides
That fourth one is underrated. Your tier-one staff are the fastest feedback channel you have, and the teams that win treat them as reviewers rather than as people whose jobs are being automated. Give them a way to flag a bad answer in one click, and review those flags weekly with visible follow-through. A rollout where agents feel heard improves on its own; one where they don’t gets routed around within a fortnight.
What Day 90 Should Look Like
By the end of the quarter you want a boring dashboard. Containment above 45% on a set of intents that cover a third of your total volume. Escalation precision above 80%. CSAT on contained tickets within a point of human-handled tickets, or ahead of it. Reopen rate under 5%. A knowledge base where the top 100 articles all have owners and review dates.
The number that actually predicts the next two quarters, though, is how many new intents you added in weeks five through twelve. Teams that ship three more intents a month keep compounding. Teams that stall at their launch set spend the rest of the year defending a pilot that never became infrastructure. Pick your three, shadow them properly, and keep the calendar moving.