Nadella's 7 principles, Dynatrace's $915M Arize deal, and a debate between CIOs, founders, and investors on who should check AI agents. Updated October 2026.
An AI agent can explain a refund perfectly and still send the money to the wrong account. The transcript reads well, and the transaction is wrong.
Three groups of products exist to catch problems like this: observability, evaluation, and assurance. Vendors have started merging them. Whether buyers will follow remains open, and so does a second question: who should do the checking, and what counts as independent?
AI is progressing from merely answering questions to carrying out actions. Gartner predicts that by 2028, one-third of enterprise software applications will feature agentic AI, up from less than 1% in 2024. These agents can write and submit code, transfer money, and communicate with customers, usually without any human having to monitor them in real time.
When software only produced answers, the main question was whether they were good. Agents raise more questions: what did it do, was it right, can someone stand behind it, and can it be stopped? For CIOs, these are as much AI governance and risk management questions as they are engineering ones.
Today three markets cover parts of this:
- AI observability records what the AI did.
- AI evaluation measures how well it did it.
- AI assurance brings together the evidence, the controls, and their limits so a person can make an accountable decision. The UK government describes it as the process of measuring, evaluating, and communicating trustworthiness.
AI observability was a $1.25 billion market in 2025, and the same research expects evaluation and quality assurance to grow fastest within it. Dynatrace closed its Arize acquisition on October 1, 2026, a $915 million deal to combine evaluation and observability. Cisco completed its Galileo deal in May. On October 10, Satya Nadella published seven principles for trusting AI that touch all three markets.
The rest of this piece covers how AI got here, what Nadella proposed, and a debate between five composite voices about where the market is going.
From rules to AI agents: why trust got harder
A simplified history in four stages. Each made the machine more capable and harder to explain.
- Rules (roughly 1950s to 2000s). Most software applications were written manually, step by step. It had plenty of bugs, but an engineer could open the code and find the line that caused one.
- Patterns (2010s). With deep learning, machines learned from examples instead of rules for tasks such as image and speech recognition. Nobody wrote the logic. These systems mostly labeled, ranked, and recommended, and people or ordinary software decided what to do with the output.
- Chat (2022 to 2024). Chatbots could write, summarise, and argue. They were often wrong, but a person usually read the answer before acting on it.
- Agents (2025 onwards). Agents click, type, buy, code, and deploy. Labs train them by having them practice tasks in simulated offices, browsers, and codebases. Some run for long periods without a person reviewing each step.
In the chat stage, a bad answer usually costs a re-read. In the agent stage, a bad action can cost money, data, or a customer.
Software made costly automated decisions long before agents. The difference now is that a single general-purpose model chooses and executes a chain of actions across several tools. When something goes wrong, the cause could be the model, the prompt, the data it retrieved, an API, stale memory, its permissions, or an exchange between two agents. The conclusive answer is only part of the evidence.
Nadella describes the old situation this way: with traditional software, "we had the tools and capability to trace behaviors to a specific code path." With current models, nobody can point to the training data or weights behind a particular decision. These three markets are trying to close that gap.
Satya Nadella's AI governance framework: 7 principles for trusting AI agents
On October 10, 2026, Nadella published "Models as Insider Risks in the Super Intelligence Era" on his blog. His central proposal is that "we need to separate the supply of intelligence from the authority over it."
He compares a capable model to an employee with wide access. The concern is not that the model is malicious. Anyone with access to important systems can make mistakes or be compromised, so you limit what they can reach, record what they do, and keep the ability to stop them. In his words, "If it can't be observed, it can't be trusted!"
He calls them seven "observability principles." Read together, they work as an AI governance framework for agents. Each principle is paraphrased below, followed by my own mapping to the markets that cover it today.
- Model diversity: no single AI checks its own work or makes an important decision on its own. Market today: evaluation.
- Observe everything: every meaningful action leaves tamper-proof evidence a human can read. Market today: observability.
- Verifiability: the entire system is continuously tested, including attacks, failures, and edge cases. Market today: evaluation and assurance.
- Independent controls: the rules on what the AI can access sit outside the AI, where it can't change them. Market today: mostly security and access tools, with guardrails starting to appear in AI platforms.
- Independent auditability: evidence about the AI's behavior is kept separate from the AI being checked. Market today: assurance.
- Containment: an authorized person can pause or shut down the model mid-task. Market today: mostly runtime and security tools.
- Incident disclosure: when it fails, those affected are informed, and lessons are shared. Market today: a company practice, supported by observability records.
No single market covers more than two of the seven. Independent controls and containment sit mostly outside all three, with security and runtime tools. Taken together, the principles describe a system larger than the three markets combined.
His concluding line is: "The most trustworthy Super Intelligence system will not be the one with the model we trust most. It will be the one that enables us to trust the model the least."
Trusting the model as little as possible requires a record of what it did (observability), a way to judge it (evaluation), and evidence a person can stand behind (assurance). It also requires two things these markets have mostly left to security and runtime tools: controls the model cannot change, and a way for a person to stop it.
AI observability vs evaluation vs assurance: the market today
Think of an AI agent as a newly hired driver.
- Observability is the dashcam. It records each turn, stop, and near-miss, and answers "what did the AI do?" It is sold mainly to the teams that keep systems running.
- Evaluation is the driving test. It scores how the driver handles set situations and answers "how good is it?" It is bought mainly by engineers building AI products, and increasingly by QA teams.
- Assurance is the inspection report. It collects the evidence needed to decide whether the car and driver are fit for the road, before the first trip and throughout the driving. It answers "is there enough evidence to stand behind this?" No inspection proves a car will never crash; a good report states what was checked, what was found, and what remains unknown. This is the least developed of the three, and the work is currently spread across quality, risk, and compliance teams.
Separate tools for separate buyers worked reasonably well while AI only produced answers. With agents, the gaps between them matter more. Each tool is useful on its own, but footage nobody reviews proves little, a test goes stale when the driver changes every week, and a report depends on the footage and test results behind it.
Recent acquisitions of AI observability and evaluation platforms:
- In October 2026, Dynatrace acquired Arize for $915M, combining evaluation with production monitoring.
- In May 2026, Cisco acquired Galileo for Splunk, covering evaluation, monitoring, and guardrails for agents.
- In February 2026, Snowflake acquired Observe, bringing observability into a data platform.
- In February 2026, ClickHouse acquired Langfuse, an LLM tracing tool.
Dynatrace says the Arize deal brings evaluation and observability together so teams can check AI before launch and keep it reliable afterward. Cisco describes Galileo as covering early evaluation through production monitoring and guardrails.
These deals show vendors combining products. They do not yet show buyers merging budgets, dropping specialist tools, or buying assurance from the company that runs their AI. The rest of this piece is about those questions.
