How to Prepare Your Data and Teams for Agentforce

Long story short: Everyone wants to flip the "AI switch" and watch their operational bottlenecks vanish. But if your data is a mess, if you don’t have a clear vision of what problem AI should solve, then turning on an autonomous agent won't fix your problems; it will just execute your broken processes at a million times the speed.

In this article, I want to provide some clarity on what things to consider when starting with Agentforce in Salesforce. Let’s look at a classic struggle businesses face today. A company has a massive volume of incoming customer service cases and wants an Agentforce Service Agent to respond automatically with a resolution based on previous interactions. It sounds like a breathtaking use case! Except, if your existing cases don’t always capture the actual resolution (or if the notes are vague), your AI-generated outcome will not be what you expect. If you don't map your business needs to the concrete reality of what data you actually hold, you are setting yourself up for failure.

The Secure Innovation Approach

Across recent Agentforce workshops and use case design sessions, we have brought business and IT teams together with our clients to define valuable, AI-driven business use cases based on their current Salesforce implementation, data maturity, and licensing model. Our objective is always to narrow the discussion down to a small number of concrete, business-oriented Agentforce use cases that can move towards implementation without turning the first step into a long transformation programme. And yes, that usually means plenty of future joy for our deployment teams. 🤤

To achieve secure innovation, our workshop established a few non-negotiable rules for Agentforce deployment:

1. Analyse standard built-in features first: Before writing custom code, look at what Agentforce can do out-of-the-box. There are massive quick-wins available just by enabling standard features in your org.

2. Map consumption to business value: Not all use cases are created equally. Some use cases consume much more credits, while the output is not as valuable. The return must justify the credits spent.

3. Start small & iterate: Start small, deploy, and plan to build upon it. Once human users start analysing and interacting with the agentic use cases, they will be the ones to concretely define the most valuable next steps.

Architecting the Use Case: What to Consider

For every Agentforce use case you define, you must rigorously evaluate two core matrices:

The Data Input Matrix:

• Core Objects: What specific Salesforce objects and fields does the agent need access to?

• Permissions: What kind of strict access/permissions should the agent operate under? Any assumptions? 

• Historical Depth: How far back does the agent need to look to find behavioural or resolution patterns?

• External Data: Do you need to bring in external data? Data Cloud connections are mandatory and come with the Agentforce license, but keep in mind that not all use cases automatically trigger Data Cloud consumption.

• Data Format: Are you analysing clean, structured data or messy, unstructured text? How is the agent supposed to treat it?

The Output & Action Matrix:

• Expected Actions: What exactly is the agent supposed to do? Are there scenarios in which the agent should opt-out?

• Human in the Loop: Will the agent act autonomously, or will a human need to approve recommended actions before execution?

• Latency: How fast does the response need to be?

• ROI: The ultimate balance of business value vs implementation effort vs consumption.

Technical Guardrails for the Real World

As a Salesforce Architect, I must emphasise a few technical realities. First, check your language support. While many languages are supported natively by Agentforce, if you are a global company, you must verify that every local language you need is fully functional and not beta. It is important to specify to the agent what the default language is (as a fallback) and, if needed, to define a list of allowed languages.

Second, testing is everything. To properly test your agent and its guardrails, a Full Copy Sandbox is highly recommended. ✅ You need to test the agent's results based on real, historical data, not just dummy records.

Not Every Technically Feasible Idea is Worth Building

Last but not least, a useful Agentforce use case needs to stand on its own as a business decision. It should be clear what success looks like, why it matters, and how quickly you can validate whether it works. In practice, that means choosing scenarios that are specific enough to measure, valuable enough to justify the effort, and realistic enough to test without turning the first iteration into a long delivery cycle.

A simple way to frame it:

What tends to happen in early AI discussions is that teams gravitate towards the most ambitious version of the idea. Fully autonomous agents, end-to-end automation, minimal human involvement. These are valid long-term goals, but they are rarely the right place to begin. They require mature data, stable processes, and a high level of trust in the system.

A better starting point is a use case where the outcome is visible, the impact is measurable, and the risk is manageable. Something that improves an existing process rather than trying to replace it entirely. Once that first use case is in place and delivering value, it becomes much easier to expand, automate further, and build towards more advanced scenarios.

The goal is not to build the most advanced agent. The goal is to build the one that delivers value first.

Don't Burn Teams for AI Dreams

Implementing AI is a massive structural shift, not a short-term marketing opportunity. I firmly believe that companies should use technology to empower their people. By focusing on clean data, standard features, and clear business value, you can build an ecosystem that is truly powered by data and inspired by humans.