Two meanings
The term is used in two ways:
- For AI builders, observability means monitoring AI applications: model latency, errors, token usage, answer quality and traces of each step an agent took. Many tools exist for this.
- For organisations using AI, observability means understanding how AI is used across the business: who uses which tools, which systems AI reaches, what it does there and whether that is safe and useful.
This article is mainly about the second meaning, though the two overlap.
What organisations want to see
- Connected agents. Which AI tools are connected to company systems, and for whom.
- Systems reached. Which data sources and applications AI is using.
- Actions taken. What AI read, created, changed or sent.
- Approvals. What waited for a person, and what they decided.
- Refusals. What was denied, and which rule denied it.
- Workflows. Which recurring jobs run through AI, and where they stall.
- Adoption. Which teams are using AI productively, and which are not.
Observability versus audit
They share data but serve different purposes. Audit is a complete, reliable record for accountability: who did what, when, under which rule. Observability is the view built on top: patterns, trends, anomalies and questions like “is AI adoption growing in the tax team?” You need an audit trail to have trustworthy observability.
Observability is not surveillance
It is possible to build AI monitoring that reads every employee’s conversations. That tends to damage trust, raises data protection questions under UK GDPR, and is rarely necessary. Most organisational questions can be answered from the requests AI makes to company systems: what was accessed and what was done, rather than what someone typed privately. If you do monitor more closely, be transparent with staff and understand your obligations. The ICO publishes guidance on monitoring workers.
Where the data comes from
- Business systems’ own logs, which often show only the connector’s account.
- AI providers’ administrative dashboards, which show usage within that one provider.
- Connector and MCP server logs, which vary widely.
- A central gateway, which can see every request across every AI in one format.
Useful questions to answer
- Which AI tools are connected to our systems today?
- Which agent accessed client data last week, and for whom?
- How often are AI-drafted emails edited before approval?
- Which rules refuse the most requests, and why?
- Which systems do people most want AI to reach but cannot?
How Kroy approaches it
Kroy Observe shows which agents are connected, which Skills run, what was approved or denied, which artifacts were produced and which systems AI reaches through Kroy. It is built from the requests that pass through Kroy. Kroy does not read people’s private AI chats. See how AI is actually being used.