GenerativeAI properly trained through detailed, extensive data sets

[Per-Kristian BM]

3 questions:

1:

In the document on generativeAI, under the heading "how does generative AI work?" (Generative AI - ServiceNow)

it says:

To reiterate, generative AI operates on the principle of machine learning through the application of neural networks, allowing the AI to generate new content from provided prompts. But before a prompt is ever introduced, the AI needs to be properly trained through detailed, extensive data sets.

How many data sets is enough in this case? is it a minimum number? or does it depends on how consistent the answer is before the AI can conclude on the answer?

2:

example: A customer wants to get value of NowAssist in Agent workspace. they have 40 thousand incident and incident task records. How ever, many of these records have less detailed work notes, comments and resolution notes. for example, I am generalizing a bit here, but I see support agents write "talk to the user on teams. issue fixed" in the resolution notes field in an incident, lets say the incident was short description "sound not working in teams"

My question. Will generativeAI at some point recognize the issue on "sound in teams not working", and NowAssist in agent workspace will recommend actions to the agent to: "talk to the user on teams to fix the issue"

Based on the data sets on incidents, this is the answer given my other agents on the same issue.

3:

Can the answer AI finds be based of knowledge records as well as work notes?

We work along the KCS principles and have over 1000 knowledge articles using the KCS template, meaning we have a very detailed overview on what users and agents search for and what the resolutions steps are. ODes generative AI give as much emphasis on knowledge records as it does work notes and other notes in other tasks?

[Tricia Cornish]

Hi @Per-Kristian BM  - 

It might help to identify what the Now Assist solution does comparative to the other ML solutions in the platform. 
In the platform today (pre-GenAI solutions like Now Assist), we allow for supervised (Classification) and unsupervised (Clustering and Similarity) models. For these frameworks, the output is based on on models built labeled data from your instance. 

Supervised models that require historical records to make a meaningful prediction for an output field usually need a minimum of 10k records. Clustering and Similarity require much less because the algorithm is based on grouping relationships. 
Link to Documentation: https://docs.servicenow.com/bundle/vancouver-release-notes/page/release-notes/analytics-intelligence...
NLU for Virtual Agent and AI Search are more reliant on linguistic rules and you can fine tune your NLU model for the specific use of the chatbot or search to include your corporate-specific lingo.  

Comparatively, the LLMs for Now Assist are based on deep learning. These LLMs are more human-like and can work from labeled and unlabeled data. With the new Generative AI solution, you could integrate to an external solution like Open AI or Azure Open AI or use the Now LLM, and you don't have to use your own data to train the model for Now Assist to be effective. The data set in the baseline LLM is substantial enough to take in content from your task record and then summarize information that is presented. Alternatively, I could could generate new content (for a KB article) kind of like you might do today with ChatGPT - the difference is we've incorporated that new content creation into the business process.   

For the use case you describe in step 2, the current solution is the Recommendations Framework, which takes advantage of the Similarity ML algorithm: https://docs.servicenow.com/bundle/vancouver-it-service-management/page/product/recommendation-frame...
For the use case, "sound not working in teams", the agent might be presented with other like incidents where the short description matched. We have not yet migrated recommendations to a Generative AI solution. 

For #3: yes, AI can find the answer in your KB. Have you reviewed AI Search and Genius Results to see how we improve search with ML-based relevancy models? That is the basis for Now Assist for AI Search - the output of a Genius Result is further summarized using GenAI for more personalized responses. Still, just AI Search with Genius Results provides a usability boost over the previous search solution.  https://docs.servicenow.com/bundle/vancouver-platform-administration/page/administer/ai-search/conce...

As for worknotes - the current predictive intelligence solutions work on string fields. Work notes and activity notes are journal fields. Today, you would have to convert those journal fields  and then find a way to parse / truncate those resulting converted notes into something meaningful. With the summarization component in the Generative AI Controller, I can take activity and work note journal fields, convert those to string fields then summarize that output (using a GenAI model) and populate a custom field in the task record. Now I have a string field with more intelligence and relevance in which I can use in my Similarity Model or my Clustering model. 
 
There is a ton of information available around the other ML frameworks that solve the use cases you presented. I would check out the AI Academy for more information on AI search, the conversational Interfaces Academy for information around NLU and Virtual Agent. Find and follow my colleague, Lener Pacania. He has written a few articles around the current Predictive Intelligence solutions. 

I just did a webinar on "Empowering your workforce with Generative AI" and use a couple of examples of summarization and KB Search in the portal and virtual agent: https://gateway.on24.com/wcc/eh/1237365/lp/4228569/empowering-your-workforce-with-generative-ai
The current Agent Assist and Recommendations framework would be the ML-based solutions to use that complement these user-focused experiences.

I hope that helps - 
Tricia 

[Per-Kristian BM]

Thank you @Tricia Cornish for your detailed response.

follow-up question:

1. In regards to your response, that predictive intelligence work on work notes (no, as it is a journal field, not string field), and your response that it reads knowledge articles.
   We use both custom made knowledge article templates as well as OOTB (KCS-article from the KM advanced plugin). These templates uses both string and HTML fields. Will GenAI read both types of fields and the setup of the KA-template should not make a difference? for exmaple the KCS-article? In other words, does it looks at the content in the article as a whole?, not by fields?
2. GenAI work on all languages?
   We use several languages in addition to English, for example we auto-translate and auto-publish (as child article of the english parent article) all knowledge articles in the service portal, to 8 languages (Norwegian, German, French, Spanish, Portguese and more, using dynamic translation plugin with MS Azure)

[Tricia Cornish · Accepted solution]

Thanks for pointing out the questions, @Per-Kristian BM I had missed these. 
Predictive Intelligence does not work on work notes. Confirmed. 

1. Now Assist does not read knowledge articles. To clarify, my statement is: "(AI Search) is the basis for Now Assist for AI Search - the output of a Genius Result is further summarized using GenAI for more personalized responses". We are aggregating information from your KB using AI Search, then the best of the best gets further summarized - Now Assist is not a separate search engine. Your  delivery mechanism for Now Assist for Search is AI Search. 
In my screenshots, this is regular AI Search with Genius Results (in vancouver, the EVAM card has been updated with "*AI Suggestion": 

That same query on the portal with Now Assist Q&A is a summary of that same article. 

The Now Assist Q&A is a Genius Result script so it looks at kb_knowledge. It is possible in this first iteration some kb configuration does make a difference. I had some issues with templates, and templates are stored in kb extended tables, and the script very specifically looks at the kb_knowledge root table (versus the kb_template_faq) but it could be user error. 

I'm not sure I understand content in the article v fields. I believe the Now Assist Q&A script looks at genius result result, specifically the text in the article_body for the answer. 

The benefits of Now Assist, summarizing a Genius Result surfaced by AI Search is that when questions are presented in a more conversational format, we can summarize and return an answer. 

In my first two examples, "What is..." in both instances is a snippet or summary of the best predicted kb article. In this third screenshot, Now Assist can answer the question presented with a more conversational and intelligent answer.

2. In this first store release, English is the only supported language for Now Assist for Search. 

Hope that helps! 

[Magnus Hovik]

@Tricia Cornish I have another question. The search engine ServiceNow uses is either Zing search or AI search. 
What search engine does this (Powered by Now assist, image below) use? And if that one should use GPT-engine, would that mean the search uses two engines? 

 

One for provided Now assist summary and one for showing the articles below?

Our biggest issue with AI search is the lack for Norwegian support (such as Typo handling) making the search engine less-than-optimal for searching. But, if the engine changes to GPT it's another thing.