All documentation
Human in the loop means Fidiora handles the questions your docs answer while your team keeps control of everything else. Handoff triggers, forced routing by category, live takeover, weekly review of AI-resolved conversations, and fixing source docs keep people in charge. Handoffs are always free, so routing to a human never costs you.
The model: AI as level zero
The safest and most effective way to use AI in support is to treat it as level zero of your escalation matrix. It is the first responder for questions your documentation already answers. Your people remain levels one and up: judgement calls, exceptions, emotional conversations, and anything with real money or risk attached.
This is not a compromise. It is how AI support produces good outcomes. An AI that tries to handle everything will eventually answer something it should not. An AI that knows its lane, and hands off cleanly when a conversation leaves it, earns trust from customers and agents alike. Fidiora is built for this model, and its pricing reflects it: handoffs to humans are always free, so there is never a financial reason to keep a customer away from your team.
| Level | Who | Handles |
|---|---|---|
| 0 | Fidiora AI | Documented, repetitive questions: how-tos, policies, account basics, order and shipping questions your docs cover |
| 1 | Frontline agents | Handoffs for low confidence, explicit requests, undocumented questions, frustrated customers |
| 2 | Specialists or senior team | Billing disputes, refunds outside policy, VIP and enterprise accounts, technical investigations |
| 3 | Leads, legal, security, engineering | Legal questions, security reports, bugs, policy exceptions |
Adapt the levels to your team. The escalation matrix template walks through building your own.
Handoff triggers
A handoff passes a live conversation from Fidiora to a person, with the full transcript, customer context, what Fidiora already tried, and why it handed off. Write your triggers down as explicit rules. These cover most cases:
- Explicit request for a human. If a customer asks for a person, they get one. Immediately, without "let me try one more thing". This is non-negotiable.
- Low confidence. If your content does not clearly answer the question, Fidiora hands off instead of guessing. The check happens before the answer, not after a customer pushes back.
- Sensitive topics. Billing disputes, refunds outside policy, legal questions, security reports, account deletion, and complaints go to a person by rule, however confident the AI is. A confident wrong answer on a refund is worse than no answer.
- Frustration. Negative sentiment, very short replies, repeated rephrasing, or phrases like "this is useless" signal that the customer needs a person. Hand off and flag for priority.
- VIP and account rules. Enterprise customers, high-value accounts, or anyone on a named list can route straight to a senior team. See Automation rules.
- Repeated failure. If the customer rejects two answers or asks the same thing twice, hand off after the second attempt, not the fourth.
Start with a conservative confidence threshold
At launch, set the confidence threshold conservatively, so Fidiora hands off whenever it is not clearly grounded. You will see more handoffs in the first week than you eventually want. That is fine: handoffs are free, they are exactly what happens today without AI, and each one teaches you something about your content.
Loosen the threshold only after review shows that Fidiora's answers on the borderline questions were correct. Loosening later is easy. Rebuilding customer trust after a bad launch is not.
Set up human in the loop
-
List what must always reach a person
Write down the categories that should never be answered by AI alone. Start with billing disputes, refunds outside policy, legal, security, account deletion, and complaints. Add anything specific to your business.
-
Turn the list into rules
In the rule engine, write a plain-language rule for each category that routes it to the right team. Add VIP rules for your most important accounts. Rules run before the AI answers.
-
Write the handoff message
Tell the customer a person is taking over, roughly when they will hear back, where the reply will arrive, and that they do not need to repeat themselves. Be honest about timing, especially outside business hours.
-
Test the path end to end
Trigger each kind of handoff in testing: ask for a human, ask an undocumented question, raise a billing dispute. Confirm an agent can reply to each without asking anything the customer already said.
-
Schedule the review
Put a recurring weekly review in the calendar before you go live, with a named owner. Review is what keeps humans in control after the first week.
Agents taking over a live conversation
Agents can step into a live conversation and take it over. Use this when an agent watching the pipeline sees a conversation heading the wrong way, when a known customer appears, or when a sales opportunity is worth a personal reply. The customer keeps the same thread and the agent sees everything said so far.
Agree team norms for takeover: who watches live conversations, at what times, and what signals justify stepping in. A small team might only watch the needs-human stage. A larger team might have one person per shift scanning AI-handled conversations for VIPs and frustration.
Reviewing AI conversations
Review is a practice, not a toggle. The teams that get the best results from AI support read its work regularly. A workable rhythm:
- Daily in week one. Read every handoff and a sample of resolved conversations.
- Weekly after that. Review a sample of AI-resolved conversations every week, plus all handoffs flagged for frustration or repeated failure. Use a quality assurance scorecard: was the answer correct, complete, grounded in the right source, and in your tone?
- After every product or policy change. Spot-check questions about the thing that changed, since that is where old content causes wrong answers.
- Monthly with the wider team. Share what Fidiora resolved, what it handed off, and which content gaps were closed. It keeps agents invested in the system rather than suspicious of it.
Fix the source, and the AI improves
When review finds a wrong or incomplete answer, do not patch the AI. Fix the source. Find the article Fidiora cited, correct it or remove it, and re-sync the knowledge source in the portal. The next customer who asks gets the right answer, and so does every human reading your help center.
The same applies to gaps. When a handoff happened because no article existed, write the article. Over a few weeks, this loop of review, fix, re-sync is what moves your resolution rate up safely. See Connect knowledge sources.
Measure resolution, not deflection
A closed chat is not a helped customer. Deflection counts conversations that never reached a human, including customers who gave up. Resolution counts customers who actually got a correct answer or had their task completed. Fidiora measures and bills on resolution, and abandoned chats and silent timeouts are never counted.
Alongside resolution rate, watch handoff quality: how often agents have to ask customers to repeat information, time from handoff to first human reply, and satisfaction on handed-off conversations. A high handoff rate on undocumented or sensitive questions is the system working. A high rate on documented questions is a content problem. See Analytics and quality and deflection vs. resolution.
Anti-patterns to avoid
- Hiding the human option, or making customers type "agent" three times.
- Looping a customer back into the AI after a handoff.
- Letting sensitive topics depend on a confidence score instead of a rule.
- Loosening the threshold before anyone has reviewed a week of conversations.
- Treating every handoff as a failure. Many are exactly right.
- Fixing wrong answers with one-off rules instead of fixing the source article.