What to look for in a support escalation platform

Most vendor pages for this category read the same: "AI-powered," "enterprise-grade," "resolves tickets instantly." None of that tells you what the tool actually does when it is wrong, which is the question that matters once it is running against your production systems.

Here is what to actually check, in the order a security review will ask about it.

What access does it need, precisely

Not "secure" or "enterprise-grade." The literal list: which credentials, scoped to which permissions, reaching which systems. A vendor that answers with a policy paragraph instead of a list has not built the constraint into the system, they have written it into a document, and a document does not stop a bad prompt or a bad deploy from doing something the sales page said would never happen.

The answer worth hearing: a read-only database role, an explicit table allowlist, a statement timeout, and no write-capable credential anywhere in the execution path. If a vendor's answer includes "and it can also update the ticket status" or "and it can issue refunds," that is a different product category (an autonomous resolution platform), not a design flaw, but you should know which one you are buying.

Is every claim checkable, or is it a summary you have to trust

Ask what happens when the tool is confident and wrong. A system with no citation mechanism has no answer to this beyond "the model is usually accurate," which is not a number you can put in a security review. A system worth buying can tell you, per claim, which probe or which document produced it, and can tell you the rate at which it drops claims it cannot back up.

The evidence contract we enforce is one concrete version of this: every claim carries a probe id, and the renderer deletes any claim whose probe did not actually run in that session, then counts the deletion. Ask your shortlist for their equivalent number, not their equivalent adjective.

What is the actual pricing basis

Per-seat pricing on a support tool taxes exactly the behaviour you want: your whole engineering team reading escalations and correcting the ones that are wrong. Per-resolution or per-conversation pricing is the right model for tools measured on deflection volume. Neither is inherently better, they are priced for different jobs. What is worth checking is whether the pricing basis matches what you are actually buying it to do. See Tier2's own pricing for one example priced against an engineer-hour rather than a ticket count.

Autonomous action, or draft and approve

This is the single biggest fork in the category, and it is worth being deliberate about which side you want before you take a demo, because a demo optimises for looking impressive and autonomous action always looks more impressive than a draft waiting for a click.

• If your queue is high-volume and low-variance (password resets, refunds, order status), an autonomous resolution platform measured on deflection rate is probably the right purchase.
• If your queue is the technical escalations that end up on an engineer's desk anyway, regardless of what closed the ticket, a draft-only tool that shows its work is worth the extra click, because the cost of a wrong autonomous action on infrastructure is not symmetric with the cost of a wrong drafted reply someone catches before it sends.

The full, named comparison covers three vendors on the autonomous side of that fork, each claim sourced to their own marketing, if you want the specific answer for specific tools.

Questions worth asking before the demo, not after the contract

1. What is the complete list of things the write path can do, and can I see the code path, not the policy?
2. What is your hallucinated or unsupported claim rate, and how is it measured?
3. What is the actual query load pattern against my database, and what protects it from an unbounded query?
4. What happens to a ticket you cannot answer confidently: does it escalate, or does it guess?
5. Who owns the correction when you are wrong, and does that correction actually change future behaviour or just this one ticket?

A vendor that answers all five with specifics, not adjectives, is the one worth trialling.