What is Salesforce AI Cloud: Should you subscribe?

Salesforce is stressing trust and security features as it unveils its new AI Cloud, but it's not very different from existing offerings, and analysts feel there could be few takers for the pricey AI cloud starter pack.

Salesforce’s new AI Cloud has left many baffled over what it’s all about — how it is different from the competition, what’s new in the offering, and whether one should consider subscribing. Analysts predict there could only be a few takers for the pricey new offering.

The Salesforce AI Cloud combines the company’s previously announced Slack GPT, Tableau GPT, Apex GPT, MuleSoft GPT, Flow GPT, Service GPT, Marketing GPT, and Commerce GPT along with the new Einstein Trust layer and a prompt engineering tool for training large language models (LLMs).

The new, bundled offering, which focuses on generative AI applications, according to CEO Marc Benioff, is based on the unification of Salesforce’s existing technology stack already offered via products including Einstein GPT, Data Cloud, Tableau, Flow and MuleSoft.  

Although Salesforce is promoting the AI Cloud as a bundled offering via a Cloud Starter Pack, which costs $360,000 annually including a free AI-readiness exam from Salesforce Professional Services, it does not have a single, unified interface, the company said.

“A lot of the AI cloud will actually manifest as assistants across different cloud offerings that we have, for example, Sales GPT in Sales Cloud, Marketing GPT in Marketing Cloud, and Service GPT in the Service Cloud,” said Jayesh Govindarajan, senior vice president of data science and engineering at Salesforce.

“However, if you are an independent service vendor and you want to create your own prompts, the AI Cloud will offer a separate local console with a low-code experience for training large language models,” Govindarajan said, adding that Salesforce had already signed up with service partners, including Accenture, Deloitte, and PwC.

The AI Cloud’s Starter Pack, according to Forrester principal analyst Liz Herbert, is just one product pricing model that the company is promoting, and “not many Salesforce customers will use that model or will be forced to buy it” as most enterprises already have CRM and other software suites installed.

“If you look at how a typical customer already subscribed to Salesforce, then the bundling is very unlikely to be a common way that someone would buy,” Herbert said, adding that Salesforce was also not very specific on pricing elements for enterprises that want to use just one or more offerings under the AI Cloud. The pricing structure is expected to be aligned, over time, to the usage of the different offerings under the AI Cloud, Herbert said.

Salesforce’s AI Cloud architecture is built in such a way that it supports multiple LLMs and their training, according to the company.

The AI Cloud architecture is supported by the company’s public cloud infrastructure architecture, Hyperforce, which in turn supports the Data Cloud layer, followed by a layer that hosts multiple LLMs (proprietary or third-party), on top of which resides the company’s new Einstein Trust Layer, the company said.

The Salesforce Trusted AI Cloud architecture consists of multiple layers.

Salesforce

The trust layer separates the Einstein GPT-based generative AI engines — such as Slack GPT, Tableau GPT, Marketing GPT and Commerce GPT — a prompt-engineering tool, and prediction engines, Salesforce said.

The LLMs that the AI Cloud currently offers include third-party models from Amazon Web Services (AWS), Anthropic, and Cohere, among others.

The company has also partnered with OpenAI to use its API for accessing models such as GPT-4, which drive ChatGPT.

Salesforce will also provide its own LLMs, such as CodeGen, COdeT5+, and CodeTF for generative AI implementation in applications, particularly those that aim to increase productivity, bridge the talent gap, and reduce the cost of tech implementation.

A new division, dubbed Salesforce AI Research, has been set up to try to develop new applications for AI by advancing new LLMs or further develop existing ones.

Enterprises, according to Salesforce, can also get their own LLMs trained on their own domain-specific data.

“These models, whether running through Amazon SageMaker or Google’s Vertex AI, will connect directly to AI Cloud through the Einstein GPT Trust Layer. In this scenario, customer data can remain within the customers’ trust boundaries,” the company said.

This isn’t the first time someone has offered these capabilities. Several other prominent technology vendors including AWS and IBM already provide similar offerings in the form of Amazon Bedrock and IBM Watsonx.

Contrary to practices from some large software vendors that claim to mix different LLMs together to offer generative AI applications, Salesforce’s AI Cloud “chooses the right LLM for the right task,” according to the company.

“The determination of the right LLM is based on outcomes that the system has seen previously,” Govindarajan said, adding that the LLM recommendation engine is learning from the data across all Salesforce deployments.

Giving the example of Marketing GPT creating a landing page, Govindarajan said the system tracks how that landing page has performed, as performance data is also stored inside Salesforce’s Marketing Cloud.

“If the landing page has really worked well, then it will suggest the same LLM it has used to build the landing page to another user who also wants to create a similar landing page,” Govindarajan said.

This same logic is also applied to the multiple flavors of Einstein GPT across different Salesforce Cloud offerings that continuously train the assistants, Salesforce said.

“These generative AI assistants such as Slack GPT or Tableau GPT track if the user ends up following a suggestion. It also tracks the edits or changes to the generated result. This helps it to learn continuously and become more personalized for users,” Govindarajan said.

Salesforce’s Einstein GPT Trust layer is neither new nor differentiated, according to Forrester’s Herbert, who says it follows the same principles that most other vendors are offering in terms of providing security. Like most other vendors, Salesforce uses identity and access management (IAM) rights in the enterprise to safeguard data privacy and security, Salesforce’s Govindarajan said.

Salesforce highlighted security features of its Einstein GPT Trust architecture.

Salesforce

“Once a user runs a query on their system, the technology stack behind these generative AI assistants searches the databases about the attribute that it needs and it has access to, followed by the stack creating a semantic search model based on the knowledge graph of the entire enterprise and the employee requesting the query,” Govindarajan said.

“The retrieved vectorized data is kept within the enterprise’s server, masked, and then fed to the large language model to generate a result or response. So, the stack is essentially copying or imbibing the identity and access management policies that are already existing for that employee,” Govindarajan added.

The Einstein Trust layer’s architecture is also built in such a way that once the query is processed, the information contained in the prompt, or query, is not retained, the company said, adding that the generated result is checked for toxicity and an audit trail log for all prompts is maintained.

The offerings under the AI Cloud, according to the company, will be made available in different phases, with the Einstein GPT and Service GPT being generally available in June.

The company’s Commerce GPT will be made generally available in July. Other offerings inside the bundle are expected to reach general availability next year after being in pilot for the majority of 2023.

It will be interesting to see how customers react to Salesforce’s new offering, which is more of a bundle of existing offerings than a completely new platform. However, analysts believe that Salesforce AI Cloud will help to enforce more trust in the use of generative AI by enterprises.

“Salesforce recognizes that there are barriers to AI adoption — especially trust. So, they are doubling down on this aspect. The other challenge is that we will have a plethora of LLMs soon — open source, public and private, leading to confusion on which one to deploy. So, Salesforce’s AI Cloud is letting the users plug in the LLM of their choice,” said Sanjeev Mohan, principal analyst at SanjMo.

Salesforce’s Marketing GPT and Commerce GPT to help build personalized services

Hush makes a noise about its data integration push to support a single source of truth

Retailer Hush grew quickly during the coronavirus pandemic and the retailer’s IT systems struggled to cope with new demand.

Like so many others, the firm, which sells its products online in the UK and Europe and in John Lewis department stores, needed to integrate disparate data sources in order to tap into accurate insight on sales trends and finance figures to people across the business. 

A decision was made to build more in-house technology capability in order to boost data access and increase business agility. According to  Brad Woodward, Head of Data at the women's lifestyle firm:

We wanted to make sure we kept the technology simple. And when we started to think about integration, which is obviously a key point for any business these days, we focused on how systems talk to each other and how you get data out of systems.

Hush submitted an RFP in June 2021. The initial shortlist included five vendors. After hands-on trials and demonstrations, they whittled that shortlist down to two – and SnapLogic came out as the winner in September of that year. Woodward recalls: 

SnapLogic trod the line nicely between being easy to do the quick things, but also having a lot of customizations and the ability to do more complex work. Some of the other platforms we evaluated seemed more like a blackbox in that regard and didn't give us the flexibility we were looking for.

Eighteen months on and SnapLogic sits at the heart of an integrated data stack. Hush uses the technology to move data between core enterprise systems, including order management, Microsoft Dynamics 365 ERP, Emarsys CRM, Google Analytics, Centric PLM, and Salesforce Commerce Cloud, says Woodward:

We use SnapLogic to hook those sources into our data warehouse, which is Google BigQuery. We then use the data to build reporting models, where we use dbt. And we have a reporting front end based on Domo, which gives users access to our curated data models that we build in dbt from all the data we've integrated via SnapLogic...It's opened up access to all of the data for people in our organization. And it's allowed us to connect up to the various data-generating systems in our business with greater ease and flexibility.

Hush uses its joined-up data approach to spot spikes in product demand. Before introducing SnapLogic, the only way to find out how many products had been sold on the website was to log in to the ERP system and run a report, which would take up to 15 minutes to produce results. Now, SnapLogic pulls data from Salesforce and produces sales results in a dashboard. Woodward explains:

People can get an accurate answer within three minutes. That's incredibly important and useful. Like yesterday, we went live with a promotion and we could immediately begin to measure the impact of that activity. SnapLogic has enabled us to do things like that, which weren't possible before.

Woodward’s team are running a variety of integrations and pipelines. For example, the engineering team recently introduced the ability to click and collect items through the website – and SnapLogic made it easy to surface the right data in the right format:

We need to keep the list of available click-and-collect locations up to date from the supplier. And it was used on a daily basis to pass the latest list of updated link locations to our website, which ensure data is constantly connected and reflects the options in real life that our customers see on the website.

Using SnapLogic to integrate data sources means users have quick access to accurate data on sales, finance and supply chain. Re-use is another boon. Woodward points out that both the data and engineering teams benefit from the integration work that others complete:

We can combine all of those pieces of data that exist across the business in one place...We can really inform what is going on in our business and think about the actions the consumer is taking that we need to understand and get ahead of.

The retailer now wants to use its single source of data to support personalization strategies that ensure Hush offers relevant and interesting products to customers. With a strong, centralized data foundation in place, the aim is to start exploring data science, explains Woodward: 

Here's our understanding of data, we've benchmarked it through reporting, we know that it's correct, and people are using it. Now we can throw data science at it and begin to extract value. Over the next 12 to 18 months, we want to look at tailored personalization and recommendation systems, things like forecasting and optimization of certain operational procedures.

Looking back over the past 18 months, Woodward reflects that, in many ways, technology has been the easy part of the transformation process. It’s much tougher to ensure non-technical people understand the power of the tools at their disposal. His team ensures the benefits are understood and communicated regularly:

We've done show-and-tell and lunch-and-learn activities where we show things off to the business. It can be challenging because if you show SnapLogic to someone working in fashion retail, you will scare them. So, it's been about sharing what we're working on and communicating it but doing it in a way that you don't lose people.

Woodward advises other business and digital leaders who are looking at SnapLogic to make sure they have the right capabilities in-house to build a single source of truth that’s verifiable and useful:

Although it’s a low code/no code platform, I do think you need a data engineer or an application engineer to be able to work with it. As with anything where you are creating a load of assets within a central tool, you need to think about boring things like naming conventions and governance structures because otherwise you can still end up creating a centralized mess.