Convot Convot
Back to AI Agent
AI Agent

The Abstain Gate

Cove AI escalates to your team instead of guessing when it is not confident - the confidence threshold controls how strict that bar is.

Updated June 11, 2026

Cove AI is built around one hard rule: do not make things up. When the bot cannot ground an answer in your knowledge sources with enough confidence, it escalates the conversation to your team instead of fabricating a reply.

This is the abstain gate.

How it works

When a customer sends a message, the bot:

  1. Retrieves the most relevant chunks from your knowledge sources.
  2. Asks the AI model to generate an answer grounded strictly in those chunks.
  3. The model returns an answer along with a confidence score (0.0 to 1.0) and a decision.

If the confidence score is below the threshold, or if the model decides the question genuinely needs a human (account-specific issues, billing disputes, requests to “talk to someone”), the bot escalates instead of answering.

In those cases the conversation is marked for your team and the customer is handed off with a brief message explaining why.

The confidence threshold

You can tune how strict the bar is in Cove AI - Guardrails - Confidence threshold:

💡
Setting Threshold Use when
Low 0.5 You want the bot to try to answer as much as possible, even on thin coverage
Medium (recommended) 0.6 Balanced - good for most setups
High 0.75 You only want the bot to answer when it is very sure

The right threshold depends on your knowledge coverage. A well-trained bot with comprehensive help articles can comfortably run at High. A bot with sparse coverage is better at Low or Medium while you fill the gaps.

💡

If you are seeing too many escalations, check the Knowledge gaps report in AI agent analytics - it surfaces the recurring questions the bot is failing on so you know what to document.

The scope fence

The scope fence is a separate safety control. You can list topics that should always go to a human regardless of confidence - billing disputes, account cancellations, legal questions, and so on.

Any message that matches a fenced topic is escalated immediately without the bot attempting an answer. Configure it in Cove AI - Guardrails - Never answer these topics (one topic per line).

Max replies before handoff

You can also set a maximum number of bot replies per conversation. Once the bot has replied N times in a single thread, it hands off to a human automatically. Set it to 0 for no limit.

This is useful for preventing long bot-only threads where a customer might be looping without getting real help.

AI disclosure

Guardrails includes an AI disclosure toggle. When it is on, the bot’s replies are labeled so customers know they are talking to an AI assistant rather than a human. Turn it on to be transparent and to meet disclosure expectations in some regions; turn it off if you prefer the bot to blend in with the rest of your team. Configure it in Cove AI - Guardrails.

What the customer sees

When the bot escalates, it sends a short, warm handoff line to the customer (generated in context, not a generic template) before routing the conversation to your inbox. Customers are not left in the dark.

Was this article helpful?