Discovering Agent Use Cases - Automations as Primitives

Hello agent builders!

This bonus newsletter talks through how we discover highly valuable AI agent use cases with the startup and enterprise clients we work with.

What we tend to find is that the templated AI agents provided by the builder tools (like Agentforce) are a great place for teams to start, but they sometimes require customisation, or do not cover some of the tasks that are great targets for AI agents for that organisation.

The first of the methods that we use we call ‘Automations as Primitives’. The automations refer to things like robotic process automation built on tools like UiPath and automation workflows built with tools like Power Automate or Zapier. The ‘primitives’ part of the name simply suggests that the automations are a basis for an agent use case that can be built upon with the added capability that gen AI (agents) possess.

Why do this?

Automations are highly valuable but they have some notable limitations. Firstly they require structured inputs and so the system setup required to provide the data in the right way is often limiting. Secondly, they are ‘brittle’ and can break if input sources or output locations change even slightly. So as you might guess we believe that migrating these processes to AI agents is more valuable.

From here we can take single automations and create AI agent use cases from them, or we map several automations together into a larger cluster or workflow from which define use cases.

To keep this brief for now, the following steps involve writing a description of the overall functions or cluster or workflow and then formatting it in a way consistent with AI agent implementation.

This structure is:

1. Name and high level description of the agent.

2. Goal of the agent. For example to synthesise brand health metrics.

3. Instructions. For example to take daily updates of brand tracker, NPS and customer feedback data > synthesise it into a raw findings doc > write a report based on the raw data with an additional hypothesis regarding any large changes > a final draft report delivered to all key stakeholders.

4. Data sources & tools required. For example as above, NPS, customer feedback S3 bucket. Or a Python tool that runs statistical analysis.

5. Depending on the tool it can require ‘few shot examples’. For example 5 example reports that were particularly insightful with regards to changes in brand health metrics.

6. Evaluation framework. For example metrics or criteria that the agent outputs can be evaluated against.

Once we have the above the team is ready for prototyping. There is an additional analytical step that we almost always go through, but the above is a sure fired way of discovering AI agent use cases in any organisation. Assuming they currently use automations, however for most mid to large organisations that is a safe assumption.

In the next instalment of this we will walk through the next AI agent use case discovery method we call ‘Digital Employees’. In the meantime though it is worth either looking at automations in your team directly or speaking with IT teams to find some clues into great AI agent use cases using Automations as Primitives!

1  https://www.youtube.com/watch?v=vVb366mGtXo Hard and Soft edge problems.