Agentforce vs a Routed Mini Model: Buy the System of Record, Not the Demo
Enterprise support platforms will use August’s agent news to sell more autonomy. The cheaper correct answer is often still a small model on top of the CRM you already pay for — unless Salesforce truly is the workflow.
This page is periodically reviewed to reflect current pricing and plan changes.
Fastest win
If the ticket, the customer and the policy live in Salesforce, Agentforce can be the fit. If they live in Zendesk, email or a homegrown admin, do not relocate the business to justify an agent SKU.
Functional fit
Agentforce is compelling when the accepted outcome is a Salesforce workflow that already has identity, audit and data. It is a poor fit when the demo is generic “talks to customers” and your records are somewhere else.
Microsoft Copilot Studio has the same logic on the Power Platform side. The August agent-safety news does not change that architecture test. It makes the test stricter: fewer tools, clearer logs.
Cost per policy-correct ticket
Include licence, credits, implementation, and the humans who still handle the 20 percent. A cheap model with 30 percent wrongful automation is not cheap.
Ask vendors for failed-action rates, not only containment storyboards. If they cannot talk about retries and reopen, you do not have a price.
Confirm current Salesforce, Microsoft and model API rates on official pages. This article does not invent credit prices.
Routed cheap models when you own the workflow
When you already have the ticket system and a developer, a mini-class or DeepSeek-class model with retrieval is usually the cost-per-ticket winner. Agentforce is a Salesforce-shaped product, not a generic chatbot.
Questions procurement should ask
Where does the ticket live today? Which write-actions are allowed without a human? What is the unit when a run fails? Can we switch the underlying model without relocating the CRM?
If the answers require a three-week workshop, you are not ready for a company-wide agent SKU. You are ready for a ten-ticket pilot.
The August safety reporting should make those questions easier to ask, not harder.
Ranked recommendation
Best choice in Salesforce-native support: Agentforce with a tight action set.
Best choice if you own the desk: routed mini/DeepSeek with retrieval and a human on write-actions.
Avoid buying both “just to compare in production”. Parallel production agents are how you get two wrong refunds.
Key Takeaways
- →System of record decides the platform.
- →Credits and tokens are not comparable until you share a denominator.
- →Safety news argues for fewer actions, not a bigger SKU.
- →Do not run two production agents on the same ticket queue.
Editorial context
Who is this for?
Ops and procurement comparing CRM-native agents with a DIY router in 2026.
When NOT to use this
Companies with no CRM and no tickets. You are not in this market yet.
Pricing insights
Agent platforms bill in conversations, credits or users. DIY bills in tokens plus engineering. Compare both against policy-correct tickets, including failed automations.
Alternatives to consider
Copilot Studio if Microsoft is the estate. Zendesk/Intercom native AI if that is already the desk. DIY router if engineering capacity exists.
Final verdict
Match the agent to the system of record. If there isn’t one, do not buy an enterprise agent to invent it.
Frequently Asked Questions
Is Agentforce cheaper than routing mini models?
Only if it retires other tools and the accepted resolution rate holds after human review. Platform seats look tidy and then grow with unused bots.
When do mini models win?
High-volume, low-risk classification, summaries and draft replies with a human on the write path. Confirm current API rates before you budget.
What should we run as a two-week test?
The same ten ticket types on the platform bot and on a routed mini-model path. Score accepted resolutions, not chatbot volume.