AuraCrate's Incident Assistant Agent

When an incident strikes not every customer is affected the same. Our AI agent analyzes data to tailor personalized care—handling minor cases, leaving complex ones to specialized human agents.

Comment 1

Inspiration

Having many years of experience in customer service (as an agent and in tech), I think all the time how companies that strive to provide a good after care, often struggle to do so in a coordinated way. I lead a team that implemented Service Cloud a couple of years ago and, while many things were achieved, I strongly believe that Agentforce is a game changer in this industry, I can see it having an enormous impact on actual personalized care, understanding that even if an incident | situation is the same, it does not debilitate the relationship of the customer and the company in the same way, and therefore, tailoring resolution and communication in the direction of not only maintain, but strengthen those relationships is the difference between losing hard-earned customers and keeping them loyal and showing them actually, care.

What it does

The AI agent will understand from the context of a given Slack channel, which 'orders' are impacted by an unavoidable incident (Think, power outages, vehicles breaking down, etc) and swiftly navigating different data points to deduce:

  • What's the likelihood of this customer leaving the company after this incident?
  • How can we best communicate with this customer (preferred communication channel)?
  • How does this customer likes to be addressed (communication tone: formal / informal / neutral)?
  • What's a suitable compensation for a customer given the impact of this incident on their relation with the company?
  • Is it acceptable that the AI agent takes over actually communicating about it with the customer or is it warranted to pass it to a human agent to ensure an optimal resolution?

How we built it

All of this was built in Salesforce, using basically three things: Agentforce, Prompt builder and Apex (because, code). The agent acts as an orchestrator, using its NLP capabilities, making sure that all actions are executed as needed. Prompt builder acts, in this case as a contextual-decision maker based on clear instructions with defined boundaries on 'simple' required data points to optimize aftercare (e.g preferred communication channel). Apex is used for more 'operational' tasks (go to this database, find x record, then find related records, process data in this way, give it back), I understand flows can also be used for this purpose and so the decision to go for Apex is just a matter of personal preference.

Challenges we ran into

  • It took me a good deal of time to understand that in Agentforce, actions are not declaratively chained but that this can be achieved via instructions and by writing in plain english to the agent what it is expected to do.

  • I could not find (in the time I had) a smooth way to parallelize or have a single agent execute all the actions for all the impacted orders. The idea of a second agent that deals with that came in too late for me.

  • I spent a good amount of time trying to figure out why a field was not shown as an option when creating a Prompt Builder - Field generation option. I went for flex (which is less ideal here, but works) instead.

  • I didn't really get to the point of fully embedding the agent to work from Slack which was my ultimate goal, I think it can be done of course, I just focused too much on having the functional part (Agent executing all the steps) instead of that integration.

Accomplishments that we're proud of

  • Running an agent that as I believe, can have such a huge impact on how aftercare is given in basically ANY INDUSTRY!

What we learned

  • A ton about generative AI.

  • Hands on Agentforce experience.

  • As I was building, more and more ideas popped in my head on how this Agent could help even more!

What's next for AuraCrate's Incident Assistant Agent

  • Oh, I would love to see this baby in action. To see a company learning that delivering "bad news" can be done in different ways and that there are "subtle" things (like the tone, or communication channel) that can make a difference for customers and can change their perception of how a company deals with unforeseen circumstances towards their customers. I would love to pick it up somewhere and see it live someday!

Built With

Share this project:

Updates