Bionic is a generative AI chatbot designed to educate, qualify, and deflect leads — essentially acting as your AI SDR. Unlike rule-based bots with strict playbook paths and expected answers, Bionic Chatbots are more conversational and probabilistic. That means there’s no single “correct” response, which can make it tricky to know when to give a thumbs up or a thumbs down in your Message Inspector.
So the team thought it would be great to form up some best practices for evaluating responses to ensure you're getting the most out of Bionic Chatbots while helping it learn and improve over time.
Before we get into it though, it’s a pretty good idea to familiarize yourself with the following help docs available over on the Salesloft knowledge site since all of the retraining starts with the message inspector and this post assumes you know at least the basics of it all:
Now with that being said, let’s get some more nuanced details, as there’s a good amount of grey area in testing and training things in this wide world of AI conversations.
What Does a Thumbs Up Mean?
A thumbs up tells Bionic:
“This response was helpful, appropriate, and something I’d be comfortable sending to a prospect.”
Use thumbs up when the response is:
- Accurate: The information is factually correct, even if not exhaustive
- Helpful: The bot addresses the visitor’s question or intent
- On-Brand: Tone, clarity, and message are in line with how your team communicates
- Sendable: A human rep would reasonably have sent the same or similar response
🧠 Important: Giving a thumbs up means that response will be used again — either for that same message or for highly similar messages in the future. You're effectively teaching Bionic, "This is a good answer — use this again."
Pro Tip: Aim for useful, not perfect. The goal is to keep conversations moving and build trust — not to nitpick minor phrasing.
When Should I Give a Thumbs Down?
A thumbs down tells Bionic:
“This response missed the mark and shouldn't be used again.”
Use thumbs down when the response:
- Contains incorrect, misleading, or confusing information
- Fails to answer the visitor’s question or intent
- Uses the wrong link, or points to an outdated or irrelevant resource
- Sounds off-brand or too vague
- Suggests actions you wouldn’t want a lead to take (e.g., booking a call when they aren’t qualified)
What About In-Between Responses?
You’ll often encounter responses that are mostly right but not ideal. In those cases, ask yourself:
Would I be okay if this were sent to a prospect right now?
If yes — even if it's not perfect — we recommend a thumbs up. That helps the model reinforce helpful patterns. If you'd feel the need to stop or correct it before sending, it's better to thumbs down.
Target: 80% Acceptance Rate
We recommend targeting an 80% thumbs up rate as a benchmark for a well-performing bot. This mirrors the accuracy and helpfulness you'd expect from a human SDR — not flawless, but consistent, clear, and productive.
Don’t hold the bot to a higher bar than you’d hold a person. Generative AI is most effective when used as a tool to scale good-enough conversations, not perfect ones.
Aligning as a Team
If you have multiple people reviewing responses, it’s helpful to align on your standards. We suggest:
- Running calibration sessions where team members rate the same responses together and discuss
- Creating a lightweight scoring rubric (e.g., rate for accuracy, clarity, helpfulness)
- Reviewing sample conversations as a team to surface different perspectives
For some follow up reading, if you’d like to understand a bit more details of the retraining triggers in Bionic Chatbots, you might want to check out this supplemental documentation on the Bionic Chatbot Retraining Triggers.
If you have any questions about anything Bionic, please feel free to drop them below and I’ll be happy to answer whatever I can, (or at least try to find someone who knows!)
Cheers friends!
-nb