Most AI automation tools expand as broadly as possible and let you discover the errors later. FortiVault inverts that. Automation is earned category by category, based on measured AI accuracy — not enabled by default and corrected after incidents.
FortiVault's automation gating enforces per-category accuracy thresholds. Billing queries, return requests, account changes — each has its own gate. When FortiAgent's accuracy in a category meets the threshold, automation is enabled. When it drops below, every response requires human review before it reaches the customer.
14-day free trial · No credit card required · 1 business day setup
The Problem
AI accuracy is not uniform across support categories. A system that correctly answers 93% of shipping status queries may answer only 74% of billing queries correctly. The model is different, the knowledge is harder, and the connector data is more complex.
Most AI customer service tools apply the same automation policy to both. Or they apply no explicit automation policy at all — they just automate everything they can and let you discover the category-level accuracy differences after customers report errors.
The structural gap in standard AI automation
FortiVault's automation gating is the enforcement layer that standard AI customer service tools are missing. It sits between FortiAgent's response generation and customer delivery, and it enforces your accuracy policy before any response is sent.
How It Works
Every support category in FortiVault is in one of three automation states. The state is determined by comparing FortiAgent's current AI Trust Score in that category against the threshold you configure. States change automatically as accuracy improves or degrades — no manual policy updates required.
FortiAgent's Trust Score in this category meets or exceeds the configured threshold. Responses are sent automatically without human review.
Example: Billing questions at 91% accuracy, threshold set to 88%
FortiAgent drafts a response but it is held in the review queue. A support agent reviews, edits if needed, and sends. The correction is logged and feeds the accuracy model.
Example: Login issues at 84% accuracy, threshold set to 87%
FortiAgent does not attempt to resolve queries in this category. The query routes directly to a human agent with full context and transcript.
Example: Technical bugs at 72% accuracy — accuracy data still collected
How the gate evaluates each response
category = "billing_query"
trust_score = 91.4 // current rolling accuracy
threshold = 88.0 // your configured gate
Category Policy
FortiVault does not enforce a single platform-wide accuracy threshold. Each category has its own gate, calibrated to the operational risk of getting a query type wrong. Billing errors are not equivalent to FAQ errors.
| Category | Risk Level | Typical Threshold | Rationale |
|---|---|---|---|
| Billing and payment queries | High | 90%+ | Direct financial impact. Errors in billing responses damage customer trust and create chargeback risk. Requires the highest accuracy threshold before automation is enabled. |
| Returns and refunds | High | 90%+ | Policy-sensitive. Wrong refund decisions create operational overhead and customer disputes. Threshold should match your refund error tolerance, not a platform default. |
| Account changes | High | 88%+ | Irreversible operations — password resets, email changes, subscription cancellations. Any AI error requires manual intervention to correct. |
| Order status and tracking | Low | 82%+ | Informational. Errors are recoverable and easy to detect. FortiAgent typically reaches the automation threshold quickly in this category. |
| Login and access issues | Medium | 85%+ | Mix of recoverable and sensitive operations. Threshold calibration depends on what percentage of login queries involve account recovery vs. simple troubleshooting. |
| Product and FAQ queries | Low | 80%+ | Low-stakes informational content. Usually the first category to reach the automation threshold and the best category to validate deployment quality. |
These are typical starting points. FortiVault lets you configure thresholds based on your specific risk tolerance, volume, and compliance requirements.
The Improvement Loop
FortiVault treats the human review queue as a structured improvement mechanism, not just a safety fallback. When a support agent corrects a FortiAgent response, the correction is logged with full context: what FortiAgent said, what was changed, and why. That data updates the AI Trust Score for the relevant category.
The practical result is an automation rate that improves predictably over time. Categories that start in human review mode graduate to enabled automation as FortiAgent's accuracy in that category crosses the threshold. Categories that degrade — because a knowledge source goes out of date or a new query pattern emerges — drop back to review mode automatically.
This is the structural difference between a governed AI automation system and a deployed model left to drift. Accuracy is always current. Policy is always enforced against current accuracy.
FortiAgent drafts a response
Grounded in configured knowledge sources and live connector data.
FortiVault evaluates against the gate
Current Trust Score for this category checked against the configured threshold.
Below threshold: enters human review
Response held. Support agent reviews, edits if needed, sends.
Correction logged and scored
The override is logged — what changed and why. Trust Score updates.
Threshold rechecked automatically
When accuracy improves above the gate, automation enables without manual intervention.
FAQ
FortiVault compares FortiAgent's current AI Trust Score in the relevant support category against the threshold you have configured for that category. If the Trust Score is at or above the threshold, the response is sent automatically. If it is below, the response enters the human review queue before the customer sees it. The decision is made per response, per category, in real time.
Yes. Each support category — billing, returns, login, technical support, product FAQs — has its own independent threshold. A high-accuracy threshold for billing queries does not affect the threshold for informational FAQ queries. Teams configure thresholds in the FortiVault admin console and can adjust them at any time without redeploying FortiAgent.
FortiAgent does not attempt to draft a response for disabled categories. The query routes directly to a human agent with full conversation context. The interaction is still logged, which means FortiAgent continues to build accuracy data for that category even while automation is disabled — so you can see when it is approaching a useful threshold.
The improvement timeline depends on conversation volume, the quality of your knowledge sources, and how consistently support agents correct FortiAgent when it is wrong. In high-volume environments with well-configured knowledge, teams typically see meaningful Trust Score improvements within the first two to four weeks. The system collects accuracy data continuously — improvement is not manual and does not require retraining.
Automation gating is currently configured at the category level, not the customer segment level. However, guidance rules allow FortiAgent's behaviour to vary based on customer attributes — for example, enterprise customers or high-value accounts can be configured to always route to human review regardless of category Trust Score.
Get Started
We'll show you how automation thresholds work in practice — and how the human review queue and improvement loop reduce oversight overhead over time.
14-day free trial · No credit card required