Dreamforce 2023: Einstein Won

in AI on September 27, 2023

By Charles McGuinness

Dreamforce 2023 is behind us, and it certainly had a consistent focus on a new product area, AI.

Still, at times the presentations felt a bit rushed or the wrong presenters were covering topics¹. I found the topics to be non-linear; I had to jump back and forth in my notes as the discussion bounced from one topic to another and then back to an earlier one again.

In previous years Dreamforces have been more of a smorgasbord of topics, while this year’s was more focused on one (AI), which made the disjointed sequencing more obvious. That’s not very relevant, other than confusing me a bit at times.

Speaking of my notes, they aren’t that helpful. But feel free to browse them if you like. Who else uses Godfather references to discuss Dreamforce?

What I Heard

As I heard it, when it comes to AI, Salesforce is arguing the following:

• Demand for AI is already large and quickly growing.
• There’s lots of reasons to be wary about Generative AI² and you need to insulate users from the risks.
• You need to bring most of your data together in one place and in a consistent format to do good AI.
• It’s not just bringing all the data together that is goodness, but having that also enables you to have a common metadata that makes it possible for everyone to understand what data exists and in what format.
• That makes it easy to build tools in a consistent way.

To meet those challenges, this is what I heard Salesforce say they are offering / doing:

• Salesforce is putting the tools in place to trust AI³.
• Data Cloud is the tool to bring together all your data and make it globally accessible as well as provide the metadata needed.
• Salesforce is moving all its applications onto a common, core infrastructure so maximize interoperability and reuse. It also makes it much easier for them to have common data and metadata repositories.
• The Einstein 1 Platform brings all the tools you need onto one common platform which makes it easier to build new AI applications.
• Copilots are being developed for a variety of applications. Copilots are:
  • Conversational: you chat with them.
  • Proactive: they will warn you about things you should know unsolicited.
  • Extensible: you can expand their abilities using Copilot Studio using flow and other tools.
• There’s a lot of (Sales, Service, etc.) GPT products mentioned as having Q4 deliverables.

High Level Commentary

• Trust Layer
  • The trust layer performs two tasks, masking data so the AI engine cannot see sensitive data, and filtering out toxic answers.
  • For the first, I think that it’s not a bad idea, even if it is not always necessary. I don’t feel the same level of threat that Salesforce is representing, but at the same time I understand that Salesforce does not want to be a contributor to a data leak and this is a decent precaution.
  • For the second, that’s going to be a big challenge; OpenAI has struggled with it at the cost of holding back or delaying useful features.
  • Regardless, I suspect that this will be an ongoing area of evolving work, akin to threat detection.
• Data Cloud
  • It continues to be a bit of an enigma to me⁴.
  • Building a common Enterprise data repository has been the holy grail for decades, and it seems hard to believe that it will ever be solved. Personally, I think of the ultimate model as being something more like this:
     
    
    The Skee-Ball model!
     
  • This is where a small amount of data from elsewhere in the enterprise is copied into Salesforce itself, a slightly larger amount gets moved into Data Cloud, but the bulk of it resides outside of either Salesforce or Data Cloud and lives in the systems of record. I don’t think this is a hinderance in the end to achieving the things Salesforce wants to achieve.

• Einstein 1 Platform
  • A lot of things, some new and some not, have been grouped together, and I’m not sure what all they are or how they fit yet. It will take time to tease them apart.
• Co-Pilots
  • This is really, really clever, especially the concept of skills. Whoever came up with this, kudos. It’s all the things people like about langchain without the hard bits.
  • Skills appear to be things that allow you to build what OpenAI calls “functions” and have them available in your Co-Pilots.
  • For my readers, we will have to review of how Functions work so you can see a bit behind the curtain on how this can be done. It looks magic until you know the trick, and they it just looks really handy.
• Future Deliverables
  • Nothing promised seems unobtainable in the timeframes mentioned with normal caveats, generally. No worries here.

Final Thoughts

Dreamforce was a lot better than I feared it might be in terms of AI content. It evidences a seriousness of commitment. Did it over generalize? Sure. Did it kind gloss over the differences between things? At times. It’s a marketing event, not a technical conference. Still, were I a Salesforce customer, I would be intrigued.

In my pre-Dreamforce article, I showed an evaluation sheet like this:

If I were to place what I saw, a lot of it goes into the Tool / Platform category, and much of it straddles between Standalone and Full Solution. I think most of it was based upon OpenAI, so I’d say most stuff is part external. There’s no right or wrong placement, to be clear. For example, having something that is fully internal at the cost of not using a much better external tool is not a good thing. Salesforce gets a faster time to market and customers faster time to ROI by not reinventing every wheel.

What’s next?

I want to take a look at Co-Pilots in detail and get a sense of what is behind them. As I mentioned above, delving into Functions will help explain a bit of how these (might) work.

Beyond that, I’m interested in seeing a bit more of some of the other features (GPT for ___) work and will be rolled out.

Final, Final Thought

An article like this is naturally prone to give Salesforce the benefit of the doubt even if the reader might not think it’s warranted. That’s because I’m trying to make sense of what was said and shown, not review it or validate it. The question is not whether this is a good start, it is. It will be interesting to see where the journey takes Salesforce.

————————————-

1. My favorite presenter was Patrick Stokes, and if Dreamforce had been just him, it would have been a lot shorter and better for me. Probably not for anyone else, though. Especially not Patrick.
2. I really, really want to use GenAI as the moniker for the newest generation, which they seem to be calling Generation α, a most unserious name.
3. Should Salesforce ever need a theme song for their trust layer, I have to say that this one came to me during all the presentations.
4. For example, they say that if you have your own Snowflake that’s cool. I’m not sure what that means precisely.