An AI that whispers answers to your operators, while they keep talking to the customer
Real-time co-pilot for support agents. Drafts responses, surfaces relevant knowledge, suggests next actions. Operators stay in control; AI handles the lookup work.
Discuss a pilot →Quick facts
- Business size
- Support operations with more than 5 agents and complex conversations
- Timeline
- 3-5 weeks test, 2-3 months pilot, 3-5 months production
- Budget range
- Pilot from €25k.
- Hardware
- Cloud-based.
- Data needed
- Knowledge base, past chat / call transcripts, agent workflow definitions.
- Evolution
-
- Genesis
- Custom-built
- Product
- Commodity
No product gives you this. We assemble and train it around you.
What this scale means
Further right means more proven and cheaper. Further left means newer and riskier. Here is the test for each step.
- Commodity
- You could get the result yourself from a ready service, with almost no work. We rarely take these on.
- Product
- A vendor already sells this result turnkey, like shelf recognition from Trax or document reading from ABBYY. If one of them fits you, use it. You come to us when it does not: when it has to run on your own servers, cost less, or fit systems the product cannot reach.
- Custom-built
- Vendors sell only the parts. A tool like Tableau hands you charts, but the dashboards and metrics for your business still have to be built. That build is the work, and that is us.
- Genesis
- The approach exists but does not work reliably yet. You are betting on it maturing, so it costs more and carries more risk.
Expected outcomes
The Problem
Support agents retype answers they have written many times before: the FAQ reply, the password-reset steps, the “let me check that for you” while they hop across three internal systems. The customer waits. The agent works hard but slowly. Throughput caps at human typing speed plus the time spent navigating tools.
One approach replaces the agent with an AI chatbot (see customer-support-chatbot). This page covers a second approach: keep the agent, and speed them up with AI in the background. The customer talks to a human, because they are. The agent gets an AI that drafts replies, looks up information, and surfaces related context in real time.
This is the “suflyor” pattern, an Italian-theater word for the prompter who whispers the next lines to actors. The actor performs. The prompter helps. For support work, it sits between full automation (cheaper, less trusted) and full manual (slower, costly), and often it is the right place to sit.
What the Solution Does
A real-time AI assist layer in the agent workspace.
- Live conversation watch, the AI reads the customer’s messages as they arrive.
- Draft suggestions, the AI proposes a response; the agent edits or sends as-is.
- Knowledge surfacing, relevant FAQ / docs / policy excerpts appear alongside the conversation.
- Action suggestions, “this customer is asking about X, open their order page” or “this looks like a refund request, start the refund workflow”.
- Continuous learning, agent edits to suggestions feed back; the AI gets better at matching agent tone.
Where It Fits
This makes sense if you…
- Operate support with > 5 agents and meaningful conversation complexity
- Have a knowledge base / docs / policies the AI can ground in
- Want to keep human agents in the conversation
- Value the right balance of automation and human touch
This is probably not the right time if you…
- Have a small team that would touch AI features now and then, with no core workflow around them
- Conduct conversations where written drafts don’t help (voice-only)
- Have agents who’ll reject AI assistance (change-management challenge)
Business Value
Throughput. Typically 25-50% more conversations per agent per hour. The lift comes from cutting the lookup-and-retype overhead. The conversation itself is unchanged.
First-response speed. AI-drafted opening responses ship within seconds. Customers wait less.
Onboarding speed. With AI assistance, new agents perform closer to experienced ones. Onboarding time typically drops 40-70%.
Consistent tone and policy. Agents working with AI suggestions are less likely to wing it on edge cases. Suggestions encode policy.
How It Works
1. Live message watch
The agent workspace pipes incoming customer messages to the AI. The AI handles each message as it arrives, one at a time.
2. Draft generation
For each incoming message, the AI generates a draft response based on:
- Customer message
- Conversation history
- Customer account context (orders, history, segment)
- Knowledge base
- Agent-specific tone preferences (learned over time)
3. Knowledge surfacing
Relevant FAQ articles / policy excerpts / product docs appear alongside the conversation. One-click insert into the agent’s response.
4. Action suggestions
For specific patterns (refund mention, order status question, escalation language), the AI suggests next actions and pre-fills relevant fields.
5. Agent editing
The agent reviews, edits, sends. The edit deltas (the changes they made to the AI’s draft) feed retraining, so the AI learns the agent’s preferred tone.
6. Continuous improvement
Datapipe-orchestrated retraining. New conversation patterns get added; new knowledge gets ingested; new agent personas tracked.
Stack
OpenAI / Anthropic / self-hosted LLMs, RAG over knowledge base, integration with operator workplace (Chatwoot / Intercom / Zendesk / Front), event-driven backend, Datapipe for retraining.
What You Need to Make This Work
Data. Knowledge base, past chat transcripts (for tuning), agent workflow definitions.
Integrations. Operator workplace integration. CRM read access for customer context.
Hardware. Cloud-based.
Team. Support-ops lead. Pilot agent group willing to engage with AI suggestions.
Implementation Roadmap
1. Test (3-5 weeks)
Connect to operator workplace. Build first version of suggestion engine. Pilot with 2-3 agents. Output: working suggestion layer with measured time-savings.
2. Pilot (2-3 months)
Roll out to broader agent team. Tune suggestion logic based on agent feedback. Build dashboards for throughput / CSAT. Output: working production deployment.
3. Production (3-5 months)
Full rollout. Continuous improvement.
Keep in Mind
- Agent buy-in is critical. Agents who feel monitored by AI resent it. Position the AI as their own tool. It is there to help them do the job, framed that way from day one.
- Bad suggestions destroy trust. Tune conservatively at first: fewer suggestions, higher quality. High suggestion volume early on backfires.
- Edit-deltas are the training signal. The AI can only learn if you measure what agents change in its suggestions. Build this in from day one.
- CSAT can dip during adoption. Agents new to AI suggestions can over-rely on them or fight them. Stable CSAT usually returns post-onboarding.
- Knowledge base quality is the ceiling. Wrong KB content = wrong suggestions.
FAQ
Will this make agents lazy / less skilled?
Risk exists if AI suggestions are accepted blindly. A good design pushes agents to review and edit each suggestion. We measure edit-vs-accept rates to keep them engaged.
How does this differ from a chatbot replacement?
Chatbot replaces the agent. Co-pilot keeps the agent and adds an AI layer. Different value, different design.
Can the AI be visible to the customer?
Configurable. Some deployments are fully invisible (customer doesn’t know AI is involved); others are explicit (“AI suggested this response, your agent confirmed”).
What about voice support?
Voice co-pilot is harder, transcription latency, agent attention split. Possible but less mature than text co-pilot.
Ready to Discuss?
If you operate support with multiple agents and per-conversation throughput is a real cost, this is a focused pilot. We’ll walk through your operator workplace and your knowledge base, and tell you what to expect.