We've been working on a use case where we’re trying to keep everything inside Salesforce instead of pushing data out to an external ERP.

Think:

• inventory movements
• purchase orders
• Advance MRP logic
• shipment tracking
• finance events (GL, Bills, invoices, payments)

The traditional approach I’ve seen is:

Salesforce (CRM) → middleware → ERP → warehouse system → accounting

Which creates:

• async delays
• reconciliation issues
• a lot of integration maintenance

We’re experimenting with a different approach:

👉 keep everything in one data model
👉 use Apex + Platform Events for async flows
👉 rely on object relationships instead of sync jobs

But a few challenges are coming up:

1. Governor limits under load
   • especially when handling inventory updates + related records
2. Transaction design
   • where to split logic (Queueables vs synchronous)
3. Data modelling
   • custom objects vs extending standard objects
4. Performance at scale
   • especially with stock movements and order flows

I’m curious:

• Has anyone here actually pushed Salesforce this far operationally?
• Did you hit limits where it stopped making sense?
• Any patterns you’d recommend (or avoid)?

Not looking for “don’t do ERP on Salesforce” 😄 — more interested in real-world architectural trade-offs.

been experimenting with Agentforce beyond the usual “chat assistant” use cases and trying to push it into actual operational workflows inside Salesforce.

Not talking about:

• answering questions
• summarising records

More like:

• triggering inventory actions
• handling multi-step processes (e.g. pick → pack → ship)
• orchestrating updates across related objects
• guiding users through workflows dynamically

The idea is:

Instead of building everything as:

• Flows
• Apex triggers
• UI logic

…you let an agent interpret intent and execute actions across the data model.

What I’m trying to understand:

1. Where does Agentforce actually sit architecturally?
Is it:

• a layer on top of Flow?
• replacing some Apex logic?
• or just orchestration + decisioning?

2. How are people handling multi-step transactions?
For example:

User says:
“Ship order 123”

Behind the scenes you need:

• validate stock
• reserve inventory
• generate shipment
• update order status

Are people chaining:

• Apex actions?
• Flow actions?
• external services?

3. How do you deal with consistency / rollback?
Since this isn’t a traditional transaction model, I’m not sure:

• where atomicity is enforced
• how failures are handled mid-execution

4. Limits & performance
If Agentforce starts triggering:

• multiple object updates
• async jobs
• integrations

Does it hit the same governor constraints indirectly?

5. When does it stop making sense?
At what point would you say:

>

Current thinking

Agentforce feels powerful for:

• decisioning
• orchestration
• user interaction layer

But I’m not convinced yet where the boundary is between:

Agent → Flow → Apex

vs just building deterministic logic directly.

Would be great to hear if anyone has:

• pushed this into real workflows
• hit limitations
• or found patterns that actually work in production