[tawk.to] Apollo AI tips - Chat transfers, Escalations and more

Great to see, that the AI bot gets smarter! I like the idea to give the customer a contact form, if the AI Bot does not find an answer. Two problems here:

  1. the ticket form is not completely in the language of the property: it says “Subject”, “Message” and “Your request has been sent”. I guess these words are hardcoded or can I edit this content of the ticket form somewhere in the dashboard?

  2. It is great, that the bot shows the ticket form, if he knows no answer. But is it also possible to only show the form, if no human supporter is online and ELSE try to switch over to a human supporter?

2 Likes

I was wondering the same about the ticket form language. I do not use the tickets because the emails are not in the property language as well. Been waiting to change the language for too long to no avail :frowning:

thanks… for the workaround. I’ll try it.

Yup I wish we could change this to spanish.

2
mensaje

1 Like

Ralph,

How can I set AI to looser “given knowledge” restriction? Do you have an base prompt example?

Thank you!

Hi,
You can find an article on this here:

For now, I would recommend disabling the ‘‘When no information is found’’ Escalate setting when the pre-chat form is enabled as it counts user’s name/email as something not found in given knoweldge.

If you have an existing shortcut that introduces a pre-chat form or transfer action, you can simply assign relevant keywords that should launch that Shortcut using the AI Question field found on bottom of the Shortcut editing window.

Hope this helps.

1 Like

Hello
I tried to follow these steps to activate the escalation but no matter what i do, the bot gets stuck during the escalation. It will mention that it is transferring to an agent but it does nothing.
Can you please confirm that it is working?

Hi weelow,

Could it be that the shortcut that is used for escalation is missing the transfer ‘‘Action’’ field?

If you need help with your personal AI setup, I am happy to help you sort out the configuration if you email me the specific cases at ralfs@tawk.to

Hello @Ralph
Thanks for your reply. My configuration matches exactly what is in the screenshot.
In addition to that, i have enabled escalation and selected the shortcut from the dropdown.

I have a prechat form setup that collects the user’s name and email before the chat starts. Could that be the problem?

The transfer action should be part of the /transfer-to-agent shortcut itself.

Additionally, you might add base prompt instruction to tell Apollo that users start chatting by filling out a form, and in response to this AI should offer assistance.

tawk.to does not send a notification when the chat is escalated. The bot just stays there in the chat doing nothing. As soon as i click the monitoring button and open the chat, i get the notifications.
This means, if an agent is not online or not paying attention, they wont know that a chat is active with the bot.

Hi,

Transfer notifications are only sent to agents who are currently online on the dashboard or tawk.to app and set as transfer assignees.

Hi @Ralph,

I’m trying to work out how best to configure my chatbot using the AI Assistant with human handover/takeover only when absolutely needed.

I can share many of the concerns/queries raised here. I’m not attempting to use the “no response” escalation but I do have significant concerns over the ability to configure/use agent escalation and lead capture.

I’ve been reading your various guides and responses on threads like this one and also https://community.tawk.to/t/ai-assist-collection-of-the-most-useful-base-prompt-instructions-that-you-can-copy/2319, so have had a pretty good play around now with the various options for both lead capture and agent escalation, e.g.

• Pre-Chat form
• Chat Rescuer
• Base Prompt with Shortcuts/Triggers

The main goals are as follows:

  1. Allow the AI Agent to do all the hard work using our training data and base prompt, essentially configuring them as a sales rep with the role of gathering info, qualifying leads, and unlimitedly capturing their details into the CRM at the right (a natural) point in time. We don’t want to scare visitors off by coming across as hard-selling and asking for details upfront.

  2. Allow human agents to proactively monitor AI chats and take over if needed, or allow the visitor to request a human agent if they prefer.

The issues are:

1. LEAD CAPTURE -

I’ve tested the various ways to do this but none of them work in a natural way:

a. BASE PROMPT ONLY -

The issue with using the base prompt to collect details is that this is free text, the CRM (Monitor section and subsequently Contacts/Inbox) doesn’t recognise or store things like “name” or “email” within the contact details/attributes. Therefore, as an admin/agent, I can’t easily tell that a lead has been qualified by the AI and contact details taken. All I see is an unhelpful string as the name e.g. “V1723294151136215”

b. BASE PROMPT + TRIGGERS AND/OR SHORTCUTS

I’ve tried using Chat Rescue and shortcuts to trigger the lead capture form. This does seem to log their name and email correctly in the CRM but there are issues:

i. It doesn’t do this when at the “Monitoring” stage, so I just see that horrible string as the name. It only seems to do this when the chat ends and gets logged in the Inbox/Contacts sections. This means, there’s no way to easily tell the contact has been qualified and details have been taken – we need to see this to determine whether it’s a time-waster or genuine lead and worth taking over the chat.

ii. I can’t trigger the form using basic prompt text e.g. “… Then collect the visitor’s details using the /lead-capture shortcut” or any variation on this.

iii. If I use option buttons/shortcuts or the Chat Rescue trigger to generate the form, it kind of breaks up the flow of the conversation and feels unnatural. More importantly, the conversation dies when the user completes the form. It’s meant to happen as part of a natural conversation (sales qualification process), with the ability for the AI to continue the dialogue, e.g. “Thank you for supplying your contact info [name], is there anything else you need?”.

iv. There’s no way that I can find to configure the lead capture form, e.g. choose/add fields, add follow-up instructions (revert back to chat), etc.

v. Even if the visitor manually types a follow-up question after submitting the form, the AI Assistant still has no knowledge of their name, etc.

c. PRE-CHAT FORM

This is the only way that seems to work as the form is configurable and passes the data to the AI Assistant so it can address them by name, etc. Using this method, the “Monitoring” section seems to instantly assign the visitor’s correct name, which is great! But this would discourage most users from even engaging in a chat for the following possible reasons:

i. Feels intrusive to ask for personal details without any niceties and like a hard sell is coming
ii. The greeting isn’t personal and feels like you’ll end up not reaching a real person (or good AI assistant)

I hope I’m missing something as the pre-chat form capabilities work exactly as needed but we need to be able to control when this happens during the AI assisted dialogue, i.e. so it happens at a natural stage and allows continued conversation with AI Assistant understanding who the visitor is.

2. AI ASSIST, ALERTS/NOTIFICATIONS, AND HUMAN AGENT TRANSFER

I really like how this works and the various options available but the two main issues are:

a. NEW CHAT ALERTS/NOTIFICATIONS

Human agents don’t get an obvious alert or notification to pro-actively monitor and take over AI chats if needed. Alerts only happen when transferring to an agent, then the app makes a ringing sound (sadly, this is less noticeable on the app and I think most agents will miss the alert).

b. HUMAN TRANSFERS -

When transferring to an agent, if they are busy or don’t notice the chat (e.g. on silent), the user is left hanging with no proactive follow up from the AI. This creates a really poor experience for the user and will most likely deter them from using us.

The Chat Rescue would work if we could configure something like, “If no agent response in 30 seconds, return the chat to the AI Assistant”, then ideally, the AI Assistant would determine when to capture the details naturally. Instead, the user completes the form and is left hanging with no AI intervention or ability to add follow-up instructions. It just gives an unhelpful “Your request has been sent”.

Final thoughts

I know there are a lot of alternatives out there, but I felt it was worth the time to document my experience here. I’ve now tested a massive range of CRM/chatbot type solutions that claim to combined AI Assisted Chatbots + Human Live Chat support, e.g. LiveChat/ChatBot, FreshChat, Socialintents, Tidio (Lyro AI), Wonderchat, Landbot, etc. All have their pros and cons, but none of them work as well as Tawk in terms of having a high configurable AI chatbot (prompts, various data sources for training, etc.) and human transfer/Live Chat capability across various channels (dedicated desktop/mobile apps, browser, etc.). Plus, it’s nowhere near as expensive as the others.

I even prefer how we can add sequences using shortcuts/messages within the AI rather than fixed sequence drag and drop workflow style bot builders. The area where I feel it’s weakest is notifications and lead capture. Again, hopefully I’ve missed something?

Thanks,
David

Hi David,

  1. Lead Capture
    To get customer details into the CRM they would have to fill in a form with fields that link to CRM. This can be done through a pre-chat form or a ticket/lead form introduced by a Shortcut. The Base prompt cannot launch Shortcuts directly, but it can introduce a button that links to a Shortcut with the form. The instructions here would involve giving the AI a scenario for when it should send the button and the quoted text that introduces the button. For example:
  • If the user is interested in our services, say, "For this, you would need to get in touch with our team. You can click the button below to be transferred to an agent, or if it is outside of business hours, you can leave us a note.
    [option] Transfer me to an agent
    [option] I am happy to leave a note ‘’

  • After answering the visitor’s inquiry, add to the end of your message: “[option] How do I get started?”

From there, you would connect the button text with the Shortcut’s AI Question field.

  1. Alerts for Currently Active Chats
    By default, AI joins the chat instantly before the incoming chat notification can ring. If the AI is to work alongside human agents, we typically add a ~20-second join delay, allowing the notification to ring so human agents can react to the chat. From there, they can decide if it needs immediate attention and join, or proceed to wait and let the AI join the chat that they can monitor. You can find the join delay setting under the Behavior section in AI settings:

  2. Transfers Outside of Business Hours
    For this, we would add a button to help users get unstuck if they request a transfer in the middle of the night, something along the lines of “Transfer is taking too long,” and link that button to a follow-up Shortcut with a resolution. This can inform the user about business hours, provide an email/phone number to reach out to, or a form to fill out.


Additional notes:

  • The Rescue trigger type cannot be launched if AI has already joined the conversation, as it is designed to be sent if no one joins the chat within a set amount of time.
  • ‘No info found’ escalation setting will only launch if AI has no matches in context selection. If AI gets a partial context match it will refer to the base prompt for what to do if it does not know the answer.

Additionally, I recommend looking into creating scripted button sequences by linking Shortcuts together. People enjoy clicking buttons, so this typically boosts engagement. By using this method, you can hardwire buttons introduced by the Welcome Trigger with Shortcuts that send forms or action transfers. I have a guide that discusses this here: Help Center | tawk.to | Using Suggested Message and AI Assist to create scripted chat sequences

I appreciate the kind words and feedback. Hopefully this helps.
If anything you can always reach out to me directly at ralfs@tawk.to

Best regards,

Thanks @Ralph,

I appreciate the response but it doesn’t address some of the most important points I’ve raised:

If you read my message “BASE PROMPT + TRIGGERS AND/OR SHORTCUTS” you’ll see that I have already tried this. Can you please respond to the other points raised here as they explain why this process doesn’t work effectively.

Thanks, this does seem to work and I now get notified on the desktop app as soon as the chat is initiated.

I have already tried this, please see below:

As you can see, there are a couple of issues:

a) Not a major issue, but the AI reply actually includes the ‘[option]…’ in the reply as well as showing the buttons.

b) This relates to my lead capture queries, but if we follow this process, the visitor is left hanging mid confersation after completing the form and there is no proactive follow up from the AI Assistant. A good visitor experience here would be something like “Thank you for supplying your contact info [name], is there anything else you need?”. However, even if the user manually types a message after completing the form (which I don’t think most would), the AI Assistant doesn’t even know their name despite the Name field being sent in the form response.

If you could respond to points 1. a to c fully, I think this would help.

Regards,
David

@Ralph, here is the other screenshot for context:

Hello again,

  1. This is a formatting issue, there needs to be an empty space after the options bracket.
    image
    Base Prompt instructions can help explain to AI how to use the [option].

  2. Submitting a form does not prompt AI for a response, only messages sent in chat do. This typically is the logical conclusion for the conversation. It’s not possible to follow up a form submission with a trigger.

  3. AI Does not see details from the submitted form or Contact attributes from CRM. It can only see text that is part of the actual chat transcript.

  4. For lead capture, an alternative to forms would be asking qualifying questions one by one and sending the resulting transcript to email via Mail Notifications. This would not update the CRM, but it would make for a more organic conversational flow.

If there are any other points that I have not addressed, please send a list of the remaining questions you have.

Best regards,

I’ve not manually added this, it’s coming from the suggested message setting:

Currently I don’t have any “[option] …” shortcuts added within my prompt.

I would disagree. In any implementation of a submission form, the user can normally configure what happens next (e.g. a redirect to a confirmation page, a thank you message, etc). In this case, as it’s happening during a chat (at a point to be decided by the user), surely we should have control over what happens next. Ideally, this would be a continuation of the chat, shortcuts, triggers, etc. or at least more control over the message so it’s not just “Your request has been sent”.

This seems like an oversight given this is a live chat and it should be aware of what’s happening in the conversation (including what gets logged to the CRM) as part of the dialogue?

Even the pre-chat provides context to the AI chat as it adds something like this after the form has been submitted:

Name: John Smith

This way, it at least knows how to address the visitor.

I appreciate this is an alternative but this doesn’t seem like a good solution and wouldn’t work for us.

Surely, the idea here is to have a fully integrated CRM and Live Chat similar so that contact data can get logged/used during the chat, similar to how it works with other solution providers?

Regards,
David

Hello again,


In the screenshot you sent, it appears that there are four buttons from Shortcut and two additional buttons for transferring too long and leaving a message, which comes from Base Prompt.
This is likely caused by conflicting instructions and the behavior toggle that allows the AI to revise shortcuts.

I am here to explain what is possible with the current version of our tool.

The points you raise are valid, and we will consider them for future updates.

Best regards,