Give Fin real-time instructions within an active conversation

​@Gustav Riederer I agree with ​@Conor but will also add one more comment, you can just add those things realtime as well and they will impact the next part of the interaction -- but if you are seeing things happening realtime while monitoring, then hop over to batch testing and add the question make the iterations and it will solve for future issues. 

Simulations in the Procedures will allow you to handle a lot of these things before you make things live but backing up a bit you may want to make sure Fin is connected to your systems so those special use cases and delivery delays are things Fin can see and respond to in a way the agent would as well --- asking yourself why is there something the agent sitting over Fin’s shoulder knows that Fin doesn’t know and how can we make sure that’s not the case in the future.   

That feels like a much better use of the agent’s time but at the same time I kind of like your idea of realtime iteration only from a launching and monitoring perspective within the first few weeks but you have the ability to do all that after the Fin interaction with the customer using the Improve Answer workflow. 

Hi ​@Conor , ​@Nathan Sudds , thanks for both of your perspectives.
I’d like to clarify our concrete use case a bit further, because this is less about theory and more about operational reality, product safety, and responsibility.

We are using Fin live in our shop.
We are a manufacturer of products for eyelash extensions — a domain where incorrect information can cause real damage (e.g. wrong usage leading to burned lashes, skin irritation, or chemical reactions).

Our setup is not basic:

• all relevant connectors are connected

• content, data sources, and logic are maintained

• overall, Fin works well

Despite this, we encounter highly complex edge cases that are not sufficiently covered.

Concrete example (from yesterday):

• Fin states via Instagram that a product is out of stock

• At the same time, Fin tells a customer via Messenger that the same product is available

In this situation, we currently have no way to immediately stop or correct Fin without actively sending a message as an agent.

Additional, even more critical example:

We have different types of adhesives:

• some work only with UV light

• others must explicitly not be used with UV

In addition, we sell lifting products where incorrect advice or usage can become dangerous.

If Fin:

• confuses products

• suggests incorrect combinations

• or provides imprecise guidance

we must be able to intervene immediately.
This is not optional — it is essential to prevent harm.

The core problem

Fin is intended to provide 24/7 support for us.
But when Fin starts delivering confusing, contradictory, or technically incorrect responses, we need a way to intervene instantly, even if we cannot reply right away (e.g. while on the go).

At the moment, “intervening” always means:

an agent must send a message

That is not always possible or appropriate in real-world operations.

Important context: rollout & optimization phase

We have been using Intercom since 01/01/2026.
During this rollout phase, it is critical for us to actively monitor, fine-tune, and optimize Fin.

This monitoring is not intended to be permanent, but rather:

• a necessary part of onboarding

• required to ensure accuracy, safety, and consistency

Why this is not a request for “live monitoring”

We do not want to supervise every conversation continuously.
What we need is a fail-safe / emergency brake for situations where:

• an agent is already viewing the conversation

• Fin is clearly providing incorrect, contradictory, or potentially harmful information

With our products, this is especially important because:

Errors are not just a CX issue — they are a safety issue.

What we expect from a solution

A way to immediately pause or stop Fin, for example:

• when a conversation is assigned to an agent

• via a manual stop / pause button

• without requiring an agent reply

• ideally also available on mobile, as the current app is very limited

For comparison:
With tools like Superchat, simply assigning the conversation to yourself automatically stops the AI agent.
This type of immediate control is currently missing with Fin.

Long term, we fully agree:
data quality, integrations, and testing should reduce these cases significantly.
But operationally, when something goes wrong in the moment, we need a safe, immediate control mechanism, without having to respond first.

In safety-critical product environments, anything else is simply too risky.