Agent Assist — Real-Time AI Suggestions for Human Agents

What Is Agent Assist?

Agent Assist shows a live AI suggestion panel alongside the agent's chat window during every conversation. As the customer types, the AI reads the conversation and prepares a suggested reply — ranked with a confidence score. The agent can send it verbatim, edit it first, or dismiss it. The agent is always in control.

Over time, as the AI learns from approved suggestions, it can optionally send replies automatically when confidence is high and the agent has consistently approved recent suggestions.

---

How It Works

The Suggestion Panel

While an agent is in a live conversation, a panel appears beside the chat window showing the AI's current suggestion and its confidence score (0–100%).

• Below suggestion threshold (default 40%): The panel stays silent. The AI doesn't surface a suggestion when it's not confident enough — low-quality suggestions add noise, not help.
• Above suggestion threshold: A suggested reply appears. The agent sees the text, the confidence score, and three buttons: Send, Edit & Send, and Dismiss.
• Send — sends the suggestion verbatim without editing.
• Edit & Send — opens the suggestion in the message composer for the agent to modify before sending.
• Dismiss — rejects the suggestion. The AI records this signal and adjusts over time.

Confidence Scoring and Learning

Every time an agent approves (Send or Edit & Send) or dismisses a suggestion, the AI records the outcome. The confidence model updates based on:

• How often similar suggestions were approved vs dismissed
• The CSAT score on conversations where suggestions were used
• The source quality of the knowledge backing the suggestion (KB article vs web training vs conversation training)

Confidence scores improve as more data accumulates. A freshly enabled Agent Assist installation starts conservative and becomes more precise over the first 30–90 days.

---

Auto-Send Mode

When Agent Assist has built up sufficient confidence on a topic, it can send replies automatically — without waiting for agent approval.

Two conditions must both be true before auto-send triggers:

1. The suggestion confidence score is at or above the auto-send threshold (default: 85%)

2. The agent has consecutively approved the last N suggestions without a dismissal (default: 5 consecutive approvals)

When both conditions are met, the bot sends the reply and shows the agent a 3-second undo window — a banner appears so the agent can cancel the send if they spot a problem. After 3 seconds, the message is delivered.

Auto-send only activates on a per-topic, per-agent basis. If confidence drops or the agent starts dismissing suggestions, auto-send turns off until the thresholds are met again.

---

Conversation Training: How Agent Assist Gets Smarter

Agent Assist draws its suggestions from all of the bot's knowledge sources simultaneously — KB articles, web training, and file uploads. But the most powerful source over time is Conversation Training: real Q&A pairs extracted from your agents' past conversations.

How Conversation Training Works

When a conversation resolves, Velaro automatically:

1. Checks that it meets the quality bar (minimum 4 messages, at least one agent reply)

2. If CSAT was collected, only indexes conversations rated 4 or 5 stars

3. Extracts visitor question → agent answer pairs

4. Converts them to vector embeddings and adds them to the bot's knowledge index

Agent Assist then searches these indexed conversations — along with KB articles and other sources — when generating suggestions. Because the training data comes from real agents answering real customers, the suggestions often match your team's preferred phrasing and policies exactly.

How Agent Assist Picks What to Search

When a customer message arrives, Agent Assist takes the last customer message as a search query and performs a vector search against the site's knowledge index before calling the AI. The AI then uses both the conversation transcript and the matching knowledge chunks to write its suggestion.

One unified index per site. All knowledge sources (scraped pages, KB articles, uploaded files, conversation training) are stored in a single shared Azure Search index tagged by site_id and index_name. Agent Assist always searches across all sources enabled for that site — it does not try to route to a team-specific or bot-specific index. This is intentional: for a human agent getting a suggestion, the widest possible relevant context produces the best answer.

How this differs from workflow bots. Automated workflow bots can be configured with a specific virtual index (e.g., one bot only knows about products, another only about support docs). Agent Assist deliberately ignores this — it always searches everything the site has published. When you have an agent answering questions, you want them backed by all your knowledge, not a subset.

The loop: agents approve good suggestions → good conversations get indexed → better suggestions → higher confidence → eventually auto-send. Each resolved conversation with a high CSAT makes the system more accurate for the next similar question.

CSAT Is the Quality Gate

Without CSAT ratings, all conversations with 4+ messages are indexed regardless of quality. With CSAT enabled, only conversations rated 4 or 5 stars get indexed. This is the single most impactful configuration choice for training quality — it ensures only high-performing agent conversations teach the AI.

---

Enabling Agent Assist

Agent Assist and Conversation Training are controlled by subscription flags. A SuperAdmin must enable them per site.

Flag	Location	Effect
EnableAgentAssist	Admin → Subscription → Agent Assist	Shows the suggestion panel to agents
EnableConversationTraining	Admin → Subscription → Conversation Training	Indexes resolved conversations into the AI knowledge index

Both flags can be toggled independently. You can run Agent Assist without Conversation Training (suggestions come from KB + web training only), but enabling both produces the strongest results.

---

Configuration

Go to Admin → Subscription (SuperAdmin) to adjust Agent Assist thresholds:

Setting	Default	Description
Suggestion Threshold	40%	Minimum confidence before a suggestion appears in the panel
Auto-Send Threshold	85%	Confidence required for the bot to send a reply automatically
Consecutive Approvals Required	5	How many consecutive approvals before auto-send becomes eligible

Raising the suggestion threshold reduces the number of suggestions shown — useful if agents find low-confidence suggestions distracting. Setting it to 70%+ shows only high-confidence suggestions.

Raising the auto-send threshold makes auto-send harder to trigger — recommended for sensitive industries (healthcare, financial services) where every reply should be reviewed. Setting it to 95%+ effectively disables auto-send for most topics.

Raising the consecutive approvals required adds more proof before the bot earns auto-send privileges on a topic.

---

Reports

Agent Assist reports appear in Reports → Agent Assist. Key metrics:

Report	What It Shows
Acceptance Rate	% of suggestions the agent approved (Send or Edit & Send) vs dismissed
Estimated Time Saved	Approved suggestions × 1.5 minutes per suggestion (average time saved per accepted reply)
CSAT Comparison	Average CSAT for conversations where Agent Assist was used vs conversations without it
Weekly Trend	Acceptance rate and suggestion volume over the past 8 weeks
Patterns Learned	Topics where confidence has crossed key thresholds; topics with high dismiss rates

Time Saved formula: Each accepted suggestion saves approximately 1.5 minutes of compose time. For a team accepting 200 suggestions/month, that's ~5 hours of agent time saved monthly — before factoring in auto-send.

---

Plan Availability

Feature	Starter	Standard	Enterprise
Agent Assist (suggestion panel)	No	Yes	Yes
Agent Assist auto-send	No	Yes	Yes
Conversation Training	No	No	Yes

Conversation Training is an Enterprise-only feature. Standard plans get Agent Assist with suggestions sourced from KB articles and web training. Enterprise plans get the full self-improving loop: agent conversations index automatically, confidence improves over time, and auto-send thresholds can be reached on most high-volume topics within 60–90 days.

---

FAQ

Does Agent Assist work on SMS and email, or only web chat?

Agent Assist shows in the agent inbox for all channel types — web chat, SMS, email, WhatsApp, and Facebook. Suggestions are generated regardless of channel. Conversation Training also indexes from all channels.

What happens if the agent dismisses all suggestions?

Dismissals are recorded as negative signals. The AI adjusts confidence scores downward for that topic. Suggestion frequency may decrease until the model retrains on better examples. If an agent consistently dismisses suggestions, check whether the KB articles covering that topic are accurate and up to date.

Can I see which knowledge source the suggestion came from?

Yes. Each suggestion in the panel shows a source tag (KB article title, web page URL, or "Conversation Training") so agents know where the AI's answer originated. This helps agents spot when a suggestion is based on outdated KB content vs a recent agent conversation.

Can I remove a bad conversation from training?

Yes. Go to Bots → [Your Bot] → AI Settings → Training → Conversation Training, find the conversation, and click Remove. The Q&A pairs from that conversation are immediately deleted from the index.