Category comparison

AI monitoring vs AI assurance.

Monitoring describes. Evaluation predicts. Assurance proves. Three layers, three different readers, and one distinction that decides which artifact a reviewer will accept.

The short answer

AI monitoring, sometimes called LLM observability, tells you how a system behaved. Traces, spans, latency, token spend, output quality scores, drift over time. The record is written by the system about itself, which is exactly right for the job it does.

AI assurance tells a third party which controls ran on a given action, in a form they can check without trusting the operator. The record is signed, content-free, and chained to the record before it.

The difference is not accuracy. A good observability trace is usually accurate. The difference is who has to be trusted for the record to count as evidence, and that only starts to matter when the reader is somebody outside the company.

Three layers, three questions

Before deployment

Evaluation

Benchmarks, red-team suites, and offline scoring. Answers whether a model is good enough to ship, and what it was designed to do.

During operation

Monitoring

Traces, dashboards, alerts, and quality scores. Answers how the system is behaving today, and which change explains a regression.

At the boundary

Assurance

Signed records of control execution. Answers what a control decided on one action, in a form a reader outside the company can check.

These are complementary and the order matters. Evaluation cannot show what a system keeps doing in production against real inputs six months later, and it is not meant to. Monitoring cannot make a self-written record independently checkable, and it is not meant to. Assurance answers neither of the first two questions and would be a poor tool for either.

Side by side

Evaluation, monitoring and assurance, compared by what each one produces
Dimension Model evaluation Monitoring and observability Runtime assurance
Question Is this model good enough to ship? How is the system behaving right now? Did the control run on this action?
Moment Before deployment, and on each release Continuously, after deployment At the instant a governed action occurs
Output artifact A score, a benchmark table, a red-team report A trace, a dashboard, an alert A signed, hash-chained receipt
Primary reader The team choosing or tuning the model The team operating the service Somebody outside the company: customer, auditor, regulator, underwriter
Trust model The reader trusts the test harness and the dataset The reader trusts the operator and the logging pipeline The reader checks Ed25519 signatures and SHA-256 commitments
Contents Prompts, outputs, and scores from a test set Payloads, metadata, and timing from production Hashes and control outcomes, never the payload
Tamper posture Reproducible if the harness and dataset are shared Mutable by design, because operators need to correct and redact Tamper-evident: an edit breaks a signature or a chain link
Handing it to a buyer Usually shared as a summary or a report Rarely shared directly; often contains sensitive payloads Designed to be handed over; carries commitments, not content

Note the last row, because it is the practical one. An observability trace is often the richest record a team has and the one it can least easily send to a customer, precisely because it contains the prompt and the output.

What monitoring platforms do well

The observability category is mature, competitive, and genuinely good. Datadog and the general-purpose observability vendors brought distributed tracing to a standard most teams could never have built in-house. The AI-native platforms — Arize, Fiddler AI, WhyLabs, Arthur AI, LangSmith and their peers — extended it into territory that classical monitoring never had to handle: prompt versioning, embedding drift, output quality scoring, evaluation harnesses wired to production traffic, and per-span cost attribution across model calls.

None of that is diminished by anything below. A team without observability on a production AI system has a more urgent problem than evidence, and should fix that first.

The one thing a self-written record cannot do

Observability records behaviour in logs the system wrote about itself. For the team operating the service that is fine, because they are inside the trust boundary already. For a reader outside it, the record has an awkward property: the party under examination produced it, and could in principle have amended it.

This is not an accusation against any vendor or any operator. It is a structural fact about self-attestation, and it applies equally to a receipt an operator signs with a key only they hold. It matters in exactly one situation: when a reader who does not trust the operator needs the record to count.

The same logic applies to governance features built into an AI platform. A system that generates both the AI output and its own governance record is self-attesting by definition. Those features are worth turning on, and they cannot resolve the independence question by themselves, because independence is not a feature you can add from inside.

A runtime receipt takes a different route. Signatures and hash commitments are checked by the reader, in their own browser, against an open standard anybody can implement. Nothing has to be trusted for the check to mean something — not the AI vendor, not the deploying organisation, not the evidence layer. You can verify a receipt yourself without an account to see what that involves.

What assurance adds, and what it deliberately omits

What it adds

A record that survives the handoff

Each governed action leaves a receipt carrying the outcome of every control that ran, a signed timestamp, signatures, and a position in a hash chain. Because it holds commitments rather than content, it is safe to hand to a party who should never see the payload.

What it omits

Everything you need to run the service

No latency histogram, no spend attribution, no prompt diff, no trace viewer, no drift chart. A receipt is a poor operating tool and was never meant to be one. Keep the dashboard.

The claim is deliberately narrow. A receipt proves faithful recording under an enrolled identity. It does not prove that every action was captured, it does not prove a system is safe, and it is never a compliance certificate. Not every receipt is witnessed either: receipts minted through the portal are countersigned by an independent witness with an inclusion proof, while receipts minted through the self-hosted kit are self-signed, and the verifier says which is which.

When monitoring is enough

The signal that the calculus has changed is usually a security questionnaire with rows about control execution, an incident where the record was the disputed thing, or an insurer asking how controls are known to have operated through the policy period.

Questions we get asked

What is the difference between AI monitoring and AI assurance?

Monitoring describes how a system behaved, in records the system wrote about itself. Assurance produces a record a third party can check without trusting the operator. The difference is not accuracy, because a good observability trace is usually accurate. The difference is who has to be trusted for the record to count as evidence.

Is an observability trace not already evidence?

It is evidence for the team operating the system, and excellent evidence at that. It is weaker for a reader outside the company, because the party under examination produced the record and could in principle have amended it. That is a structural property of self-written logs rather than a criticism of any product, and it only becomes a problem when the reader is a customer, an auditor, a regulator, or an underwriter.

Does runtime assurance replace an observability platform?

No, and swapping one for the other would be a bad trade. Observability platforms answer the questions you have every day: latency, cost, error rates, token spend, drift, which prompt version regressed, and which trace explains a complaint. Runtime assurance answers a question that arrives occasionally and matters enormously when it does. Most teams run both, and the receipt sits underneath the dashboard.

Where does model evaluation fit?

Evaluation sits before deployment and answers whether a model is good enough to ship. It tells you what a system was designed to do. It cannot show what the system keeps doing in production against real inputs six months later, and it is not meant to. Evaluation, monitoring, and assurance answer three different questions at three different moments.

Can we get assurance from the AI vendor built-in governance features?

Those features are useful and they are worth turning on. The limit is structural: a platform that generates both the AI output and its own governance record is self-attesting by definition. Independent evidence requires independence, which is why the verification path has to work without trusting the AI vendor, the deploying organisation, or the evidence layer itself.

What does a runtime receipt contain that a trace does not?

Less content and more proof. A receipt carries a hash of the input and a hash of the output rather than the text of either, the outcome of each control that ran, a signed timestamp, one or more Ed25519 signatures, and its position in a hash chain plus the hash of the record before it. A trace usually carries the payload and no signature. The receipt is smaller, safer to hand out, and checkable by someone who has never met you.

Related reading

Keep the dashboard. Add the receipt.

Start free and sign your first governed action today, or talk to us about the workflow your customers keep asking about.