# From sandbox to a governed AI agent in production.

> We took it from a test to a governed agent in production, with the controls a consumer brand needs.

- Source: https://synconai.com/case-studies/agentforce-sandbox-to-production
- Publisher: SynconAI (https://synconai.com)
- Type: Case study
- Evidence: Delivered work
- Industry: Consumer Goods & Retail
- Products: Agentforce, Einstein, Salesforce Platform
- Service: Salesforce Agentforce Consulting (https://synconai.com/salesforce-agentforce-consulting)
- Reading time: 1 minutes

## In short

We took it from a test to a governed agent in production, with the controls a consumer brand needs.

## How we help

We took it from a test to a governed agent in production, with the controls a consumer brand needs.

## Before and after

| Area | Before | After |
| --- | --- | --- |
| AI agent | Stuck in testing | Live and governed in production |
| Controls | Unclear what it could touch | Access limited to its job |
| Confidence | Nobody ready to sign it off | Tested against a defined job |

## Results

- AI in front of customers, safely
- A repeatable path for the next agent
- Clear ownership after go-live

## How it works

An Agentforce agent is given a defined job, the topics it may handle, and the actions it may take: Flows, Apex, prompt templates and API calls that run with the agent’s own permissions. The reasoning engine decides which action fits each request; the actions do the work that must not vary.

Grounding decides answer quality. Agents read trusted data from Salesforce records, knowledge and Data 360, and guardrails, testing and a clear hand-off to a person decide whether it is safe to put in front of customers.

## Our approach

1. **Pick one job.** Choose a high-volume, well-understood task and define what success and failure look like.
2. **Design topics and actions.** Map each request type to a topic, and each step to a Flow, Apex or API action with least-privilege access.
3. **Ground and guard.** Connect the right records and knowledge, then set instructions, limits and escalation rules.
4. **Test before customers do.** Run scripted and adversarial tests in a sandbox until the agent passes a defined bar.
5. **Release and measure.** Roll out in stages, watch resolution and escalation rates, and tune from real conversations.

## What to get right

- Give the agent its own user with only the access its job needs.
- Keep deterministic steps in Flow or Apex, not in the prompt.
- Make escalation to a person carry the full context.
- Treat testing as a release gate, not a demo.

## Questions

### How long does it take to get an Agentforce agent into production?

A focused agent with one clear job typically reaches production in a few weeks once data and permissions are ready. The time goes into grounding, testing and access design, not into the agent itself.

### Do we need Data 360 to use Agentforce?

Not always. Many agents work from Salesforce records and knowledge alone. Data 360 helps when the agent needs data from outside Salesforce or a unified customer profile.

### How do you stop an agent doing something it should not?

Least-privilege permissions on the agent user, actions that validate their inputs, clear instructions and topic boundaries, and tested escalation to a person.

---

SynconAI. https://synconai.com/case-studies/agentforce-sandbox-to-production