How to set your escalation policy: When to bring in a person directly

Start with what’s already built in

In the case of Fin, escalations automatically occur when a customer clearly asks for a person, when it detects strong frustration or anger, and when a customer gets stuck repeating themselves in a loop. None of this needs setup so it’s worth checking what your Agent handles out of the box. As a result, the real work of designing an escalation policy is figuring out where you need more precision than the default gives you, and adding customizations on top.

For Fin, escalation policy is set via Escalation Rules and Escalation Guidance. Both override its default behavior, and can prevent an escalation just as easily as trigger one.

Escalation Rules are data-driven: they fire off attributes like sentiment, order value, or a VIP flag. Escalation Guidance is natural language, better suited to intent or behavior that isn’t captured in a data field, like a customer repeatedly visiting your cancellation page.

After building out your escalation policy, it’s important you configure where conversations are sent. This is typically handled in workflow branches and your helpdesk’s own rules. Rules and guidance answer when and why; workflows answer what happens next.

Where to add guidance on top of the default

A few categories worth writing guidance for:

Broad guidance backfires

This is where most teams get it wrong in the opposite direction. Broad or generic guidance doesn’t just add a few extra handoffs, it causes escalations to spike and resolution rate to drop, and disproportionately so. Part of the reason is that most customers accept an escalation offer when it’s presented, even when the Agent could have solved the problem on its own. Once you offer the handoff, people take it.

So write guidance for the narrowest situation where a person is genuinely required. “Escalate when the customer wants to discuss billing” may be overly broad as many billing queries can be handled by an Agent, even complex motions like negotiating pricing changes. Have confidence in your Agent’s abilities and save the offer for when it's truly needed, because making the offer changes the outcome.

Your policy should evolve

Every rule is worth revisiting as your content and confidence mature. You might hand an entire topic to your team today and take it back for the Agent later.

We saw this play out with pricing and billing at Fin during a stretch of frequent changes. We didn’t trust the Agent to be consistently accurate, so we routed all of it to our human reps, built better content in the background, and eventually handed those queries back over. Keep a running list of what's currently held back and when you plan to revisit it.

Ownership and measurement

Once the policy exists, it needs a named owner: one person who oversees the rules, checks they’re working, and changes them when they’re not.

In our case, the whole handoff process is managed by our conversation designer within our AI support team. Some of our customers rely on the individual within their organization who owns Fin and its implementation. Whatever their role, what’s important is that someone takes responsibility for your escalation rules and handoff points.

Start measurement with what’s already available in your reporting. Escalated conversations and escalation rate give you volume and percentage. Fin can break this down further by providing configuration-based escalation reasons, showing exactly why a conversation was escalated (guidance applied, rule applied, and so on), which tells you whether a spike is coming from something you wrote or from Fin’s own judgment.

Beyond that, having your team flag whether a handoff actually needed a person adds a layer volume metrics can’t give you. It’s a lightweight review process worth setting up alongside your native reporting, and tells you not just how often your Agent hands off, but how well those handoffs are timed.

What good looks like:

• You know where the line sits, and which parts are the Agent’s default versus what you've configured.
• Your guidance targets narrow, necessary situations rather than generic catch-alls, because you know broad guidance hurts resolution rate.
• Your Agent knows when a human is available and can route sensitive-topics to them immediately.
• One person owns the whole thing, watching volume metrics and sampling conversations to get a qualitative read on timing.