Auditability Just Became an AI Support Buying Criterion

For most of this category’s history, the quality question was whether the AI gave a good answer. As agents move from answering to acting, the question becomes whether you can reconstruct why it did what it did.

Zendesk shipped a Context Graph alongside its Resolution Platform, capturing an audit trail of operational memory including agentic reasoning and performance context, explicitly framed as a trust layer for delegated outcomes. That framing is the notable part. Auditability is being positioned as a buying criterion rather than a feature.

Why this is happening now

Three pressures arriving together.

Agents take actions. Answering a question is low stakes and reversible. Issuing a refund, changing a plan, or cancelling an order is not. Once software acts on customer accounts, “it seemed right at the time” stops being an acceptable explanation.

Billing depends on it. Outcome-based pricing means a vendor counts the thing you pay for. Without an audit trail there is no way to verify the count, which is an odd position to accept in any other category of procurement.

Disputes need reconstruction. When a customer says the AI told them something, you need to know whether it did. A transcript answers that for answers. Actions need more.

What to ask for

For answering: source citation. Every factual claim traceable to the document that produced it. This is the minimum viable version and it catches most of the hallucination risk, because it converts an unverifiable statement into a checkable one.

For acting: a full trail. What action was taken, with what parameters, on whose authority, under which rule, with what confidence, at what time. Exportable, and retained at least as long as your dispute window.

For billing: the right to audit a sample of billed resolutions and confirm each met the contractual definition. If a vendor bills on a unit it counts, letting you check the count is a reasonable ask. A refusal is informative.

For quality: enough detail that your weekly review can determine whether a wrong answer was a content problem, a retrieval problem, or a guardrail problem. Those have different fixes and you cannot tell them apart without the trail.

The procurement question this becomes

Expect these to appear alongside security and data residency in questionnaires:

• What is logged for each AI interaction and each AI action?
• How long is it retained, and can we export it?
• Can we audit billed resolutions against the contractual definition?
• Can we reconstruct why a specific decision was made, six months later?
• Who at your company can see our audit data?

Teams that have been through a customer dispute involving an AI answer ask these unprompted. Teams that have not tend to discover them the hard way.

The cheaper version

Not every team needs a full agentic audit infrastructure, and most do not yet take actions at all.

If your deployment only answers, the practical requirements are smaller: answers grounded in your own content, sources cited so any claim can be checked, a log of questions the system could not answer, and a weekly sample reviewed against source documentation.

That combination gives you most of the benefit at almost no cost, and it is the discipline we recommend regardless of vendor. We wrote it up in how to measure AI support quality.

Where we sit

Fidiora answers from your own content and traces answers back to their sources, so a wrong answer points at the article that needs correcting rather than at an opaque model decision. Billing is on genuine resolutions with the definition published in full, and abandonment, timeouts and handoffs are never billed.

The broader point stands independent of vendor: once you delegate decisions to software, the ability to reconstruct them stops being a nice-to-have. Ask the questions above of whoever you are evaluating, including us.

Source: diginomica coverage of Zendesk Relate 2026.