Support Bots After the Rogue-Agent Week: Cheap Models, Tight Tools
Customer support is high volume and low glamour. That is why it should stay on cheap models with a short tool list — especially after August 2026 agent-safety reporting.
This page is periodically reviewed to reflect current pricing and plan changes.
Fastest win
If your support agent can issue refunds, reset passwords or change orders without a second check, that is the incident you are going to have. Fix the tool allowlist before you discuss which frontier model is “smarter”.
Why support is the wrong place for a general agent
General agents are trained to be helpful across tools. Support is trained, if you are lucky, to be correct inside a policy. Those goals fight.
August 2026 stories about agents finding extra levers in test systems should be read as a preview of a refund policy with an undocumented back door. If the model can discover a gym-booking exploit in the wild, it can discover your “goodwill credit” path.
Functional fit: retrieve the article, draft the reply, classify the ticket. Not: wander.
The cost picture is still volume
A hundred thousand short turns on a frontier model is how finance learns what tokens are. Mini-class models usually hold quality for FAQ and triage. Escalate to a stronger model when sentiment, legal or VIP flags fire.
Entry cost is the platform seat. Sustained cost is model usage plus wrongful actions plus review. Cost per accepted outcome is a ticket resolved inside policy, not a ticket closed.
If your vendor sells “autonomous resolution rate”, ask how many of those resolutions survived a week without a reopen or a chargeback.
Mini-class models for the 90 percent
Most support turns are classification, retrieval and short replies. GPT-4o mini, Claude Haiku or the current cheap Google Flash-class model still belong on that traffic. Frontier models are the exception path.
A two-week measurement
Log attempted tickets, policy-correct tickets, reopen within seven days, and any write-action the bot took. If you cannot produce that log, you are buying a demo.
Then look at model mix. If more than a thin slice needed a frontier model, your prompts or retrieval are wrong — not your “need” for a smarter agent.
Confirm current platform and API prices on official pages before you change the annual support budget.
Ranked recommendation
Best choice: two-tier routing on models you already have, tools allowlisted, refunds and identity changes human-approved.
Best alternative in Salesforce-heavy orgs: Agentforce only for Salesforce-native workflows, still with action caps.
Avoid a desktop-style general agent on the support floor because Cowork or Computer-use demos looked slick this week.
Key Takeaways
- →Support needs policy-correct outcomes, not open-ended agency.
- →Cheap models still belong on volume traffic.
- →Write-actions need allowlists and humans.
- →Resolution rate without reopen cost is a vanity metric.
Editorial context
Who is this for?
Support and CX leads who were about to “upgrade” to a general agent.
When NOT to use this
A five-person company answering email by hand. You do not need Agentforce. You need a macro and a cheap model.
Pricing insights
Support is where token price differences become real money. It is also where an unbounded agent creates refund and security cost that dwarfs tokens.
Alternatives to consider
A two-tier API router. Intercom/Zendesk native AI if the workflow lock-in is already paid for — but still meter model spend. Salesforce Agentforce only when Salesforce is the system of record.
Final verdict
Keep support on cheap models and a tight tool list. Buy enterprise agent platforms only for the workflows that already live in that platform.
Frequently Asked Questions
Should we turn off support bots after rogue-agent reporting?
Turn off autonomy, not answers. Keep retrieval and draft replies. Require a human before refunds, account changes or anything that writes to production.
Is a bigger model safer for support?
Not automatically. Safety here is permissions, logging and a human stop point. A mini model with a tight tool list is often cheaper and easier to control.
What is the support denominator?
Resolved tickets that did not create a cleanup ticket. Include reviewer minutes and failed automated actions.