Assist & Token - How Assist and tokens define your Now Assist licensing

[PaulSylo]

Hello All, 

By now many of you are familiar with "Now Assist" and starting to work in "Now Assist", experiencing its potential to enhance workflows and provide valuable insights. However, the underlying mechanisms that drive the "assist" calculations – that crucial license components – often feel like a black box to me.

Understanding precisely how this Assist is calculated has been a topic of much discussion and exploration within our community.
While the exact formula might remain somewhat not clear, this article aims to provide my perspective by relating this with a a key component: Tokens. By exploring, how Now Assist likely breaks down information into tokens and how these tokens might contribute to the assist score, we can begin to unravel some of the mystery.

My primary goal here is to share my current understanding and offer a framework that helps explain this concept behind, to some extent, the fascinating process behind Now Assist's intelligent suggestions.

let's start with Assist and then Tokens, then respective examples. 

What is an Assist ?

“Assist" as a unit of measurement that tracks AI interactions. “Assist” Measures the number of interactions done by your AI Engine.

 Example 1:

Q1.  A user asks AI bot - " What is the Capital of France?"

R1. AI Bot processes this question and  responds: "The capital of France is Paris"

In this case, above single Back and forth  - user's question and AI's direct answer considered as 1 Assist. Understanding is every time AI successfully handles request form a user, it increments the "Assist" count by one.

Example 2: 

Q1.  A user interacts with AI bot - "Can you give me a recipes for making chocolate cake?"

R1. AI bot processes and responds : " Chocolate, Flour, sugar, eggs, Salt.." and instructions. this consumes 1 Assist.

Q2. User : "Thanks ! how long should i bake them for a soft and fluffy cake?"

R2. Now, AI bot processes and responds : " for fluffy bake it for 20 min", this consumes 1 Assist , so totally 2 Assist are consumed.

Hope you go this basic understanding and how Assist are consumed. Kindly note, This means it's not about the internal processing (tokens) but about the user-facing action or completed task. Lets take ServiceNow related example,

Lets decode the "Assist calculation" by a scenario,

Scenario 1 : Incident Summarization in ServiceNow

Imagine a user has a long, detailed incident ticket in ServiceNow, containing:
* Description: A lengthy explanation of the problem.
* Steps to reproduce: Detailed instructions for replicating the issue.
* Logs: Technical data and error messages.
* Comments: A thread of back-and-forth communication between the user and support staff.
The user wants Now Assist to generate a concise summary of this incident.

Practical Example :

Let's say you use Now assist to

Refer:  https://www.servicenow.com/docs/bundle/vancouver-platform-administration/page/administer/subscriptio...

Skill	Complexity	Token	Assist units
Chat reply recommendation	High ( multiturn context, Predictive generation)	Many tokens, dynamic reason.	5
Change risk explanation	Low( single record, static reason)	few tokens, less logic	1
Email summarization*	Medium( single record, Status reason)	Many tokens, less login	1 (The number of tokens used would depend on the length of the summary.)

How "Assist" Works:
* User Request: The user clicks a button or triggers a command in Now Assist to summarize the incident.
* AI Processing: Now Assist's AI model takes the entire incident ticket as input includes short description, Description, Work notes ( this is customizable in Yokohama)
* Summary Generation: The AI processes the information and generates a short, coherent summary.
* Display Summary: Now Assist displays the summary to the user.
This entire process, from the user's initial request to the display of the summary, counts as ONE "Assist".

Now Lets talk about tokens 

• "Tokens" are the units of data that the AI uses to perform that task, and the number of tokens affects the computational cost

Ex: lets take same above example "What is the capital of France ?"  - this input string broke down into individual token by AI tokenizer like "What", "Is" "the" "capital" "of" "France" "?" , so roughly 7 tokens 

Response will be "The capital of France is Paris." it breaks into "the" "Capital" "Of" "France" "is" "Paris", so roughly 6 token, altogether 13 token are processed. Then imagine, a long email how may token can be used and how these computational cost are incurred, that why ServiceNow Skill kit restricts token s.

How "Tokens" Work:
* Input Tokens: The AI model breaks down the entire incident ticket (description, steps, logs, comments) into individual tokens. Each word, punctuation mark, and even parts of words become tokens.
* Processing Tokens: The AI processes these input tokens, analyzing the content to understand the context and key information.
* Output Tokens: The AI generates the summary, which is also composed of tokens (words, phrases, etc.).
The total number of tokens used in this process includes:
* The tokens from the original incident ticket (input).
* The tokens generated for the summary (output).

AI incurs a computational costs for processing both input tokens ( to understand the questions) and for generating output tokens( to form answers). the more tokens, more higher processing times, high energy..

I hope this article gives you basic understating on Assist, its calculation and tokens. in My next articles i will writer on this Assist can be related to Tokens and computational cost and why "Now assist" comes with predefined assist

Kindly feel free to edit or correct me if my understanding ! Hope this helps.!

Regards,

Paul