Agents Augment Automation
Over the last few months I’ve seen variations of posts on LinkedIn suggesting that we stop creating metadata such as validation rules, flows to update records, and simply hand everything over to Agentforce. It can consider a database operation, decide if the affected record is valid, and then apply any field changes needed based on its understanding of our business. Now I get that these are most likely clickbait, or a naked attempt to get likes/reposts from Marc Benioff, but I decided to think about them seriously anyway.
Let me start by saying I don’t have any problem with looking to apply Agentforce/Generative AI wherever it will bring benefits. I remain convinced that this is the future, though not quite as simple and flawless as some vendors would have me believe.
What I’m not convinced about is replacing rules-based automation with decision making Agents. The whole point of rules-based automation is that we want to take action on every record that satisfies a rule, forever. In technical requirements we’d describe this as deterministic and predictable - the same input will always produce the same output. Generative AI does not offer this behaviour, even with all the guardrails we can think of. Its strength is its probabilistic attributes: creativity and unpredictability, but those aren’t benefits for compliance or financial processing automation.
Rules-based automation typically involves very simple rules that are easily expressed in a declarative tool and consume very little resource to evaluate. Consider a validation rule formula to stop a Contact record being saved without a way to get in touch :
ISBLANK(Email) && ISBLANK(Phone) && ISBLANK(MobilePhone)
Replicating this with an Agent would require a prompt to explain the business rule in natural language, grounding with the Contact record fields, and a round trip to an LLM. Slower, more resource intensive, and re-interpreting the requirements and data every time, thanks to zero-retention. Then there’s the cost implication - validation rules, flows and other automation are included with your Salesforce license. Once you introduce AI into automation, you’ll find yourself burning a lot of Einstein Request tokens, to evaluate each record even if no action is required.
The final drawback to this approach is around governance. It’s far easier to explain why an action was taken based on a simple rule as opposed to justifying a choice made by an LLM.
This isn’t to say that Generative AI doesn’t have a role to play in automation, though. It really comes into it’s own when the processing takes place in the unstructured world of text and free-form customer interactions. It’s far better at distilling sentiment from a message than rules looking for specific words, and therefore likely to be far more successful at automation in this space, escalating cases based on customer responses for example. Agents won’t replace simple, structured rules-based processing. Instead they will parse and classify information from unstructured data so that it can be used by new automation. They’ll augment existing processing rather than taking over.