Every new Salesforce feature arrives on a wave of marketing. Agentforce was no different — which is probably why it generated as much skepticism as excitement.
After months working with the platform, here's my honest take.
What works well
Case triage agents with real CRM context
This is the most solid use case I've seen so far. An agent configured for triage can access customer history, product type, configured SLAs — and make routing decisions with real context, not just keywords. The results for high-volume simple cases are genuinely impressive.
Response automation without custom code
For first-level responses, Agent Builder can deliver useful flows without a single line of Apex. For teams that always needed a developer for any automation, this is a real shift.
Flow Builder integration
The Agentforce + Flow combination for orchestrating complex actions works better than I expected. You define what the agent can do, Flow executes it — and the division of responsibilities stays clear.
Where you'll hit walls
Grounding with dirty data
This is the biggest practical problem. The agent is only as good as the data grounding it. If your CRM has inconsistent fields, stale values, or broken relationships — and every real CRM does — the agent will produce responses that sound confident but are wrong.
The finding that stuck with me most: the problem wasn't the agent, it was the data. Before any serious Agentforce implementation, audit your data.
Einstein Trust Layer has real volume limits
On the basic tier, you'll hit call limits that aren't clearly documented. Capacity planning is essential — especially for high-volume use cases.
Sandbox behavior ≠ production behavior
I learned this the hard way. The agent that worked perfectly in sandbox started behaving differently in production — higher volume, data variations that didn't exist in sandbox, users asking questions no test script would have covered.
What isn't ready for mission-critical use
Financial decisions without human review
Any flow where the agent makes a decision with direct financial impact — credit approval, pricing, refunds — still needs a human review layer. Today. The guardrails exist, but they're not sufficient for demanding regulatory environments.
Environments with strict compliance requirements
The Einstein Trust Layer addresses most PII concerns, but if you work in healthcare, financial services, or any sector with strict decision auditing, the compliance conversation needs to happen before the pilot — not after.
What this means for architects
Agentforce is genuinely powerful. But like any Salesforce tool, implementation is where the project succeeds or fails.
What I recommend before any implementation:
- Data audit before any agent configuration
- Explicit mapping of required security guardrails
- Pilot at real scale (not just sandbox) before committing to a roadmap
- Team training on what the agent can and cannot decide on its own
The promise is real. The execution needs to be careful.
Building with Agentforce? What's been your biggest obstacle? Tell me on LinkedIn.
