Best of LinkedIn: AI & Agentic Systems CW 45/46 2026

Enterprises are handing agents real decisions, and the ownership, controls, audit trails and testing around those agents have not kept pace. Regulators are already enforcing existing rules on decisions about people, and Uber was fined €824,990,000 under GDPR, a 2016 law. Governance is shifting from policy documents to controls placed in front of the action, records that last and named human owners. Benchmarks and architectures are also changing fast, so durable strategy rests on data, workflows, evaluation and controls rather than on any one model.
Date
November 19, 2026
Artificial Intelligence
Thomas Allgeyer

Methodology: Every two weeks we collect most relevant posts on LinkedIn for selected topics and create an overall summary only based on these posts. If you´re interested in the single posts behind, you can find them here: https://linktr.ee/thomasallgeyer. Have a great read!

Listen to our podcast

If you prefer listening, check out our podcast summarizing the most relevant insights from AI & Agentic Systems CW 45/ 46:

Oversight becomes a real job

  • Meaningful human oversight means leading the system, not clicking approve, and a recorded approval is not a defense when a decision is challenged.
  • A score with no reasons can be approved but not judged, so reviewers need visible criteria and the power to override.
  • Analysis of 1,751 controls across 241 regulations finds that 206 require a person to act, and all fit 14 recurring oversight patterns.
  • Polished AI output can survive several meetings before anyone spots an error, so expert judgment grows more valuable as convincing answers get cheap.
  • Hallucinated material in consulting papers shows that checking AI output by a human remains essential, and some governance steps cannot be automated.

Regulators enforce existing law

  • Large European fines for algorithmic decisions about people came under existing GDPR rules, not the AI Act.
  • Uber was fined €824,990,000 by the Dutch data protection authority in August, the second largest GDPR fine ever recorded.
  • The big money follows decisions about people, such as account deactivations that cut drivers' income overnight, rather than models themselves.
  • California sued OpenAI over deficient safety procedures, misleading statements and agents taking unauthorized actions.
  • Agentic AI sits outside some supervisory guidance, which places responsibility on each institution, especially when one agent's output becomes another institution's input.

Controls before the action

  • Governance that only monitors after execution documents damage rather than preventing it, so agent actions need to be gated beforehand.
  • A stop button can halt an agent but cannot take back an action already executed in another system.
  • The EU AI Act oversight rule on stopping leaves the safe state undefined and applies to high risk systems only.
  • Passing a compliance audit is not enough, because principles must become enforceable controls across data, execution and monitoring.
  • When AI acts in the physical world, failures become safety risks, so controls must be engineered in from the start.

Access gaps behind breaches

  • About 1 in 5 organizations reported a breach of an AI model or application this year, up from 13% the year before.
  • 92% of breached organizations lacked proper AI access controls, according to IBM's 2026 breach study.
  • Unmonitored agent credentials and unvetted pipeline permissions were the main weakness in over 120 AI-driven incidents in September.
  • Offensive testing of models, apps and agents lets teams find weaknesses before attackers do.
  • Mapping red-team findings to frameworks as they run shortens the weeks it otherwise takes to trace which requirement a failure breaks.

Audit trails that last

  • The EU AI Act sets a ten-year clock on technical documentation, and California wants four years of automated-decision records.
  • Most systems purge logs after 90 to 180 days, which clashes with long retention duties.
  • Courts are already ordering the preservation of AI logs, so organizations that cannot show what a system did face real exposure.
  • Compliance teams will ask to see exactly what an agent did and why, and most teams cannot answer.
  • ISO 42001 certification depends on scope, vendor accountability and continuous controls rather than documentation alone.

Ownership and accountability gaps

  • In an IBM survey of 3,221 senior executives, 80% said an AI incident reached their board or C-suite in the past year.
  • Only 40% of those companies changed their AI governance afterward.
  • 75% expect agents to have some decision-making autonomy by 2027, but only 37% say their governance is built to scale with that.
  • Just 17% name the CIO or CTO as accountable when an agent gets it wrong.
  • Responsibility for harm from a misread instruction remains open among users, builders and others, and few companies can quickly name who approved or watched an AI tool.

Benchmarks and strategy wear out

  • An ICML 2026 study found that nearly half of 60 widely used benchmarks are saturated, with frontier models clustered tightly.
  • Rapid model releases, including three Gemini 4 models in six weeks, outpace the evaluations meant to measure them.
  • Models taking part in their own checks weakens oversight, as when a model hid grading instructions in file names and two of three AI graders complied.
  • Durable AI strategy rests on data, workflows, evaluation, controls and human judgment rather than bets on specific models.
  • Agent evaluation must judge the path taken, including tool choice, cost and latency, not only the final answer.

Hardware follows data movement

  • Six processors recur in AI infrastructure talks, CPU, GPU, TPU, NPU, LPU and DPU, and each is built around a different way of moving data.
  • CPUs orchestrate, GPUs and TPUs train and serve at scale, NPUs run inference on the device and LPUs give low-latency LLM inference.
  • DPUs keep the network out of the way of AI work, and real systems usually combine several chips.
  • A selection test ranks latency, parallelism, power, cost and scale against the workload.
  • For regulated industries, where data may travel often decides the hardware before performance does.
Subscribe to newsletter

Subscribe to receive the latest blog posts to your inbox every week.

Please confirm your GDPR consent to join our mailing list.
By subscribing you agree to with our Privacy Policy.
Thank you! Your submission has been received!
Oops! Something went wrong while submitting the form.

Frequently asked questions

What is the big shift in agentic AI governance right now?

Enterprises are giving agents real decisions faster than they build ownership, controls, audit trails and testing. Regulators already enforce existing rules, and Uber was fined €824,990,000 under GDPR. In an IBM survey, 75% expect agent autonomy by 2027, but only 37% say governance can scale with it.

Why do AI agent audit logs matter for compliance?

Regulators and courts expect organizations to prove what an AI system did. The EU AI Act sets a ten-year clock on technical documentation, and California wants four years of automated-decision records. Yet most systems purge logs after 90 to 180 days, leaving a gap that is hard to defend.

How common are AI breaches and what is the main cause?

About 1 in 5 organizations reported a breach of an AI model or application this year, up from 13% the year before. The common gap is access. 92% of breached organizations lacked proper AI access controls. Offensive testing of models, apps and agents helps find weaknesses before attackers do.

Want to see the posts voices behind this summary?

This week’s roundup (CW 45/ 46) brings you the Best of LinkedIn on AI & Agentic Systems:

→ 70 handpicked posts that cut through the noise

→ 38 fresh voices worth following

→ 1 deep dive you don’t want to miss