---
title: "Observability"
url: "https://mahouse.world/glossary/observability"
type: "glossary term"
author: "Ihor Vlasov"
reviewed: "2026-09-30"
published: "2026-09-22"
---

# Observability

Observability is the degree to which the internal state of a running system can be worked out from the data it emits — in software, its metrics, logs and traces. For AI it adds prompts, responses, token counts, cost, tool calls and quality scores, showing not only that a system is up but whether its answers are right.

**Also known as:** Software observability, AI observability, LLM observability, o11y

## Key facts

- **Origin:** a control-theory concept introduced by Rudolf Kálmán in his work on linear dynamic systems in the 1960s, borrowed by software engineering for distributed, cloud-based systems
- **Core signals:** metrics (measurements over time), logs (records of events) and traces (the path of one request across services), often called the three pillars; AI observability adds prompt and response capture, token usage and cost, latency per model call, agent and tool steps, drift, and automated evaluation of output quality
- **Standards:** OpenTelemetry, a Cloud Native Computing Foundation project formed by merging OpenTracing and OpenCensus, is the open, vendor-neutral standard for collecting telemetry; its semantic conventions for generative AI were still under development in 2026
- **Binding:** no rule requires "observability" by name, but several require what it delivers — the EU's Digital Operational Resilience Act, applying since 17 January 2025, obliges financial entities to manage and report major ICT incidents, and the EU AI Act requires high-risk systems, credit scoring among them, to be able to log events automatically from 2 December 2027
- **Market:** a large software category — Cisco's $28bn all-cash acquisition of Splunk, announced in September 2023 and completed in March 2024, was the largest deal in Cisco's history

## How observability works in practice

Monitoring answers questions set in advance: is the error rate above 1%, is a server out of memory. Observability is about answering new questions after something has gone wrong — why did payments from one bank slow down for twenty minutes last Tuesday — from telemetry already collected. A payments platform instruments each service so that one authorisation leaves a trace across the gateway, fraud engine, processor connection and ledger, with the merchant, card scheme and latency attached to every step. When approval rates dip, an engineer filters traces by issuer and finds the timeout in minutes rather than hours.

AI systems need more. A language model can return fast, error-free responses that are wrong. AI observability records each call — prompt, retrieved documents, model and version, tokens, cost, tool calls and the final answer — and scores samples for accuracy or policy breaches, often with a second model acting as judge. For an agent that takes ten steps, the trace shows which step went astray.

For a buyer, observability is diligence evidence. A target that can produce traces, incident timelines and a history of model evaluations can show how its systems behave under stress; one that cannot is asking the buyer to trust a dashboard. Telemetry is also a cost: high-cardinality data grows quickly, and what it takes to store belongs in the cloud budget a buyer reviews.

## Example

A lending platform runs a language-model assistant that answers 300,000 customer questions a month. Its infrastructure dashboards show 99.9% availability and a median response of 1.8 seconds. It adds AI observability: every call is traced, and a second model grades a 2% sample each day. Questions about early-repayment fees make up a quarter of the volume, and the grading shows that 4% of those answers quote an outdated charge — about 3,000 customers a month. The traces point to one policy document that the retrieval step still returns after a fee change. Fixing the index takes a day. The telemetry costs £3,000 a month, a tenth of the assistant's model spend.

## Common mistakes

- Taking uptime and latency as proof that an AI system is working.
- Logging without a request identifier, so one transaction cannot be followed across services.
- Keeping every prompt and response indefinitely without checking what personal data they contain.
- Instrumenting with a proprietary agent, so changing vendor means instrumenting everything again.
- Sampling so thinly that rare but costly failures never reach the data.

## Related terms

- [MLOps](https://mahouse.world/glossary/mlops.md)
- [AI Agent](https://mahouse.world/glossary/ai-agent.md)
- [Large Language Model](https://mahouse.world/glossary/large-language-model.md)
- [Hallucination](https://mahouse.world/glossary/hallucination.md)
- [Audit Trail](https://mahouse.world/glossary/audit-trail.md)
- [Failover](https://mahouse.world/glossary/failover.md)

— Ihor Vlasov, M&A FinTech Expert. Last reviewed 2026-09-30. https://mahouse.world/authors/ihor-vlasov