A tiered agent system that resolves routine support tickets end-to-end, escalates the rest with full context, and keeps humans on the hard 20%.
A SaaS support org at scale handles six- or seven-figure annual ticket volumes, and the uncomfortable truth is that most of those tickets are variations on a few hundred questions that have already been answered. The default response — hire linearly with volume — is the wrong answer and an expensive one. A tiered multi-agent system fixes the economics: a front-line agent resolves the routine, well-understood tickets end-to-end (password and provisioning issues, known errors, how-do-I questions answered from docs), a triage layer classifies and enriches everything else, and the genuinely hard 20% escalate to humans with full context already attached. The design principle is conservative on purpose — deflect only what the system can resolve confidently and safely, escalate the rest rather than guess, and instrument every decision so you can measure deflection without eroding CSAT. Humans stay on the complex, high-empathy, account-sensitive work where they are actually needed.
A B2B SaaS company with ~100k tickets a year deployed a tiered system that auto-resolves provisioning and known-error tickets from a governed answer base and escalates the rest with a pre-written summary; sustained deflection settled around 45% with CSAT flat-to-up because escalations arrived fully contexted. A mid-market SaaS support team uses a confidence threshold — below it, the agent never sends a customer-facing answer, it only drafts for a human — which kept hallucinated answers out of the customer’s inbox entirely. A scale-up redeployed the tier-1 capacity it freed into a proactive retention and onboarding desk rather than cutting heads.
Cluster 12-24 months of closed tickets to find the repeatable intents. The top 150-300 intents typically cover 70%+ of volume — that set defines what the agents can safely own.
A triage agent classifies intent and urgency; a retrieval agent answers from a governed knowledge base. Ground every answer in citations so reviewers can audit it.
Give the action agent narrowly-scoped tools (refund under a cap, plan change, address update). Every write is permissioned, logged, and reversible.
Anything low-confidence, out-of-policy, or emotionally charged escalates to a human — with a structured summary, attempted steps, and suggested resolution attached.
Run the system in suggest-only mode against live tickets, measure deflection and CSAT against a holdout, then ramp auto-resolution intent by intent.
Tuned for SaaS & Tech Companies. Use as-is or adapt to your voice.
Classify every inbound ticket into: AUTO-RESOLVE (matches a known, low-risk intent with a verified answer and the agent’s confidence ≥ threshold), ASSIST (agent drafts a reply for human review — used for moderate confidence or moderate risk), or ESCALATE (account-sensitive, billing disputes, churn signals, security/legal, or low confidence). Never AUTO-RESOLVE a ticket that mentions cancellation, a security concern, or legal/contract terms — those always route to a human. Log the chosen tier, the confidence, and the matched intent for every ticket.
ESCALATION — [ticket id] Customer: [name, plan tier, ARR, tenure] Sentiment / risk flags: [frustrated / churn-risk / security / billing] What they want (one line): [...] What the agent already tried: [steps + outcomes] Relevant account state: [recent errors, plan limits, open incidents] Known-good answer if any, and why the agent did not send it: [...] Suggested next action for the human: [...]
Set a per-intent confidence threshold; the system may send a customer-facing answer ONLY above it. Maintain a deny-list of intents that are never auto-resolved (cancellation, refunds, security disclosure, data deletion, anything legal). Require that every auto-sent answer be grounded in a cited, approved knowledge source — answers with no source are downgraded to ASSIST. Sample 2–5% of auto-resolved tickets weekly for human QA and feed misses back into the intent base. Track deflection %, CSAT on deflected vs human-handled, and reopen rate as the guardrail trio.
## Escalation
Intent: {classified_intent}
Confidence: {score}
Customer sentiment: {sentiment}
Attempted: {steps_taken}
Knowledge gaps hit: {missing_docs}
Suggested resolution: {recommendation}
Account flags: {risk_flags}Score each auto-resolved ticket 0-2 on: (1) factual correctness vs. docs, (2) policy compliance, (3) tone. Ship an intent to auto-resolve only when it holds >=95% at 2/2 across a 100-ticket sample.
Get one new AI workflow per week, tuned for SaaS & Tech Companies teams. Real templates, real ROI.
Do not use this when your knowledge base is stale or contradictory — agents amplify bad documentation. Fix the source of truth first. Also skip full auto-resolution for regulated or safety-critical support until human review data proves the intent.
One practical AI tip per week. No fluff.
Get the full guide with niche templates and workflow imports.