Most AI support agents derive their behaviour from prompts — natural language instructions the model interprets differently each time. FortiAgent's behaviour comes from your configuration: which knowledge sources it answers from, which connector APIs it can call, which rules it applies. It cannot answer outside what you have configured. It cannot call an API you have not registered. It cannot invent a refund policy.
And before any response reaches the customer, FortiVault checks whether FortiAgent's accuracy in that query category currently meets your threshold. If it does not, the response goes to a human reviewer — not to the customer.
14-day free trial · No credit card required · 1 business day setup
The Problem
The AI support agent market has a fundamental product gap. Most tools are general-purpose language models with a support-focused prompt layer. Their behaviour is emergent, their knowledge boundaries are porous, their API access is unconstrained, and there is no enforcement layer between response generation and customer delivery.
Generic AI support agents
The wrong-answer problem: when the model fills in the blanks
When a generic AI chatbot cannot find your answer in your knowledge base, it falls back to what it knows from training data. That training data includes documentation from other companies, outdated product versions, and generic industry information. The model responds confidently — citing something that sounds like your policy but is not.
FortiAgent
FortiAgent does not fall back to training data. If the answer is not in your configured knowledge sources, FortiAgent escalates to a human agent. Customers never receive an answer FortiAgent invented.
Generic AI support agents
The API access problem: models that decide what to call
Most AI agent frameworks give the model a list of available tools and let it decide which ones to call, with what parameters. This is convenient for demos. In production, it means the model may call a cancel-order endpoint when it should have called a track-order endpoint — because both were in the tool list and the model's interpretation of the query led it there.
FortiAgent
FortiAgent can only call connectors registered in the connector registry. The registry defines which APIs exist, what parameters they accept, and under which query conditions they can be called. The model does not decide — the configuration does.
Generic AI support agents
The prompt drift problem: behaviour that changes without notice
Generic AI agents behave based on how the model interprets your prompt today. Model updates change prompt interpretation. The same instruction produces different outputs across model versions. Your billing escalation rule that worked last month may behave differently after the next model update — and you will not know until a customer reports it.
FortiAgent
FortiAgent's escalation conditions, tone rules, and compliance constraints are explicit guidance rules in configuration — not prompt instructions. They apply identically across every conversation and every model version. Configuration changes are versioned and auditable.
Generic AI support agents
The discovery problem: finding out what went wrong after it already did
Generic AI chatbots have nothing between response generation and customer delivery. If the AI tells a customer their refund was approved when it was not — or that their account is active when it has been cancelled — the customer receives that response. The support team finds out from the escalation or the complaint.
FortiAgent
FortiVault's accuracy gate runs between FortiAgent and the customer. Before any response is delivered, FortiVault checks whether FortiAgent's current Trust Score in the query category meets your threshold. If it does not, the response goes to human review — not to the customer.
How FortiAgent Works
Every aspect of FortiAgent's behaviour is defined at configuration time. Knowledge sources, connector registry, guidance rules, escalation conditions, write-back procedures — all explicit, all auditable, all bounded.
FortiAgent searches your configured FAQs, policies, and product guides for every response. If a matching answer is not in your knowledge base, the query escalates to a human agent. There is no fallback to model training knowledge — which means FortiAgent cannot state an incorrect refund policy, invent a product feature, or answer a question with data from your competitor's documentation.
FortiAgent can retrieve live Shopify orders, Stripe billing state, Zendesk ticket history, Salesforce account data, and any REST API you register. It cannot call an API you have not explicitly registered in the connector registry. No unrestricted tool use. Every call is logged with parameters sent and data returned.
Escalation conditions, tone rules, prohibited response patterns, and compliance constraints are set at configuration time as explicit guidance rules. They apply consistently across every conversation — they do not vary based on the model's interpretation of a prompt on any given day. Every response shows which rules were applied.
Order cancellations, subscription changes, refund processing, account updates — configured as multi-step procedures where each step is defined, validated, and audited. FortiAgent cannot execute a write operation that was not explicitly configured as a procedure. No free-form API writes.
FortiVault measures FortiAgent's accuracy in the query's category and compares it against your configured threshold. If accuracy is below the gate — billing queries failing at 76% when your threshold is 90% — the response enters the human review queue. FortiAgent does not send. This runs on every response.
Deploy the FortiAgent widget with one script tag. Responses stream via SSE in real time — no polling, no page reloads. Integrates with Zendesk, Freshdesk, and existing helpdesk workflows. Appearance and behaviour configured from the admin console.
Live Connector Data
When a support query requires live data — order status, subscription state, account details — FortiAgent calls the appropriate registered connector API. Connectors are configured explicitly in the registry. Only registered connectors can be called. Every call is logged.
Commerce
Shopify, WooCommerce, Magento, BigCommerce
Billing
Stripe, Chargebee, GoCardless, Xero
CRM
Salesforce, HubSpot, Zoho, Pipedrive
Helpdesk
Zendesk, Freshdesk, Intercom, HelpScout
Logistics
DHL, FedEx, Royal Mail, UPS, DPD
Identity
Auth0, Okta, internal SSO
Internal APIs
Any REST API — custom connectors supported
Why Governance Matters
FortiAgent's configuration-defined behaviour means it operates within tighter bounds than a generic AI chatbot. But bounded execution is not the same as governed execution. Governance means measuring whether the execution is accurate, enforcing policy before responses are delivered, and maintaining a full audit trail of every decision.
FortiVault provides the three components FortiAgent cannot provide for itself: an AI Trust Score per support category, automation gating that holds responses below the accuracy threshold, and a per-decision audit trail that makes every FortiAgent response traceable and disputable.
This is why FortiAgent is deployed as part of the FortiVault platform — not as a standalone product. The execution layer and the governance layer are designed to work together.
What FortiVault adds to FortiAgent
AI Trust Score
Continuously updated accuracy signal per category — tells you whether FortiAgent is ready to automate in billing, returns, login, or any other category.
Automation Gating
Enforces accuracy thresholds before responses are sent. Below threshold: human review queue. Above threshold: automated. Applied per response, in real time.
Full Audit Trail
Every FortiAgent decision logged: knowledge source, connector call, guidance rule, gate state, outcome. Immutable, queryable, exportable.
Human Review Queue
Responses held for review show the full decision context. Review actions feed the accuracy model — corrections improve the Trust Score.
FAQ
FortiAgent is bounded by configuration in ways that generic AI chatbots are not. It answers only from your explicitly configured knowledge sources — no fallback to training data. It calls only connectors registered in the connector registry — no unrestricted API access. Its behaviour is defined by explicit guidance rules, not by prompt interpretation. And every response is governed by FortiVault's AI Trust Score before it reaches the customer. A generic AI chatbot does none of these things structurally.
If FortiAgent cannot retrieve a relevant answer from your configured knowledge sources, it escalates the query to a human agent rather than attempting to answer from model training data. The escalation is logged with full context — what knowledge was searched, what was retrieved, and why it was insufficient — so support teams can identify knowledge gaps and fill them.
FortiAgent supports write-back procedures — multi-step operations that modify records in connected systems. Refund processing, subscription changes, account updates, and order cancellations can all be configured as explicit write-back procedures. Each procedure is defined step-by-step in configuration: what data to retrieve, what validation to apply, what API call to make. No free-form write operations are possible — only explicitly configured procedures can execute.
FortiAgent does not attempt to expand outside its configuration. If a query type is not covered by the configured knowledge sources and connectors, FortiAgent escalates it to a human agent with the full conversation context. This boundary is structural — it is not a prompt instruction that the model can interpret around. Configuration defines the boundary of what FortiAgent can and cannot do.
FortiVault evaluates every FortiAgent response before it is delivered to the customer. It checks the AI Trust Score for the query category against the configured threshold. If accuracy meets the threshold, the response is sent. If not, the response enters the human review queue. FortiVault also logs a full audit trail for every FortiAgent decision: knowledge source, connector call, guidance rule applied, gate state, and outcome. FortiAgent operates within this governance system — it does not operate independently of it.
Get Started
We'll show FortiAgent handling real support query types — grounded knowledge retrieval, live connector calls, guidance rule application, and FortiVault governance in action.
14-day free trial · No credit card required