New playbook: when your own AI model pays. Read
An AI agent governance framework gives each agent a narrow job, a named owner and explicit decision rights, logs every action, and gates every change with evals.
It sets which decisions each agent may make alone and records them. This page covers the CEO’s decisions, the autonomy options, cost drivers, and what Stripe, DoorDash, LG Uplus, Coinbase and Zurich built in 2026.
Four things: identity, decision rights, traceability and evaluation. Google’s October 2026 framing gives every agent its own identity stamped into logs, permissions approved by security, and one policy gateway for all traffic.
The Financial Stability Board’s October 2026 practices for financial institutions add two more: attribute every action and escalation to a specific agent, and test kill switches that range from a full stop to a graceful return to human operation.
The CEO sets decision rights: what agents decide alone, what they recommend, and what stays human. IBM’s September 2026 rule is least privilege plus least agency. If the task only needs a recommendation, the agent may not need execution rights, and thresholds such as refund limits are enforced in the workflow itself.
The CEO also names a human owner for each agent’s outcomes. A Harvard Business School working paper on 202 enterprises (October 2026) found top-quartile firms run recurring production evaluations in 63% of cases against 4% in the bottom quartile, and counts a named human owner for agent outcomes among the practices it scores.
Companies give agents more autonomy one class of decision at a time. The JPMorgan treasury model (June 2026) moves from insight to signal to recommendation to execution.
| Level | Agent does | Human does | Move up when |
|---|---|---|---|
| Shadow | Recommends on live cases with no write access | Decides and acts | Agreement with human decisions is proven on real cases |
| Decision support | Prepares a decision-ready file | Decides | Reviewers accept most recommendations unchanged |
| Constrained execution | Acts inside written thresholds | Reviews exceptions and samples | Eval scores hold and errors are reversible |
| Delegated authority | Acts alone on a defined class | Audits logs and outcomes | The class is low-risk and fully traced |
Eval sets cost expert time: someone who knows the work must define what correct looks like. The second cost is reviewer time, because each human approval waits for a person. Shadow mode, described by Microsoft in July 2026, runs the agent on live cases without touching records or customers and gives the evidence needed to remove approvals safely.
Stripe breaks each financial-compliance review into small pre-tested sub-tasks, runs agents on each, logs every tool output, and hands findings to human reviewers who decide. Stripe says median review handling time fell 26%, with over 96% helpfulness from reviewers (June 2026).
DoorDash built an evaluation harness for its ordering assistant with rubrics per agent, a conversation simulator and auto-grading of 2,000 live sessions a day. DoorDash says quality scores rose 8 points and the error rate nearly halved before nationwide launch (June 2026).
LG Uplus feeds production traces, user feedback and contact-centre experts’ judgments into evaluation datasets that drive every change to its contact-centre agents, Arize reported (August 2026).
Coinbase has a security agent assign each code change a risk level, and anything affecting the movement of money still gets thorough human testing and approval (August 2026). Cursor reports that agents create 75% of Coinbase’s pull requests (June 2026).
Zurich’s commercial-insurance agents turn broker submissions into decision-ready files across more than 20 countries. Cytora reports manual triage time down 80% (September 2026).
Pick one agent already in use. Write its decision-rights map: what it decides, recommends and escalates. Give it an identity and a named owner, log every action, and build a replay set of fifty real cases before the next change ships.
Other layers: Models and routing · Data and ontology · The learning loop
26%
shorter median compliance review time
An agent system for financial-compliance investigations that breaks each review into a graph of small, pre-tested sub-tasks, quality-tests each, constrains agents with rails, and hands findings to human reviewers who decide (source).
2,000
sessions auto-graded a day before launch
An evaluation harness for the multi-agent Ask DoorDash ordering assistant: rubrics per agent, a transcript builder, a conversation simulator and auto-grading of live sessions, used as the launch bar (source).
70% → 96%
agent accuracy through evaluation-driven development
Evaluation-driven development for contact-centre agents on Arize AX: production traces, user feedback and domain-expert labels feed evaluation datasets that gate every change (source).
75%
of pull requests created by agents
An agent-first development loop: agents turn Slack requests into pull requests and test builds, a security agent assigns each change a risk level, and simulator fleets verify behaviour against the spec before humans look (source).
80%
less manual submission triage time
An agentic intake platform that reads broker submissions, digitises them and triages risks into a decision-ready workflow for underwriters across more than 20 countries (source).
12,500
banker hours saved in the first month of operation, July 2026
Five specialist agents on Amazon Bedrock AgentCore that classify and read payslips and bank statements, run serviceability calculations and verify them against lending policy, then hand the banker a result to review for mortgage and credit-card applications (source).
$10M
EBITDA uplift from two agent systems, Partners Group says
Two agentic systems built on a cleaned proprietary policy-level dataset: one runs new-client policy processing, the other automates human-intensive policy checking; delivered by a seven-person team with Version 1 (source).
2–3 minutes
per ocean or air freight quote, down from one to three days
Hundreds of in-house AI agents embedded across the shipment lifecycle (quoting, orders, appointments, missed-pickup checks) and now ocean and air quoting, ocean booking and customs brokerage, built and run under a lean waste-removal discipline by 450+ in-house engineers and data scientists (source).
Evals are repeatable tests that score an agent’s output against what an expert calls correct. A good eval set mixes replayed production cases with rubric scoring and runs before every change. DoorDash cut a six-hour manual regression test to about 20 minutes with its harness (June 2026), which made testing every change practical.
Autonomy is set per class of decisions, based on results on real cases. Start in shadow mode, move to decision support, then to execution inside written thresholds. Spotify auto-merges a defined class of maintenance changes without human review while other changes keep review, which shows autonomy assigned per class (June 2026).
Traceability means every action an agent takes is logged with the agent’s identity, its inputs, tool calls and outcome, so anyone can reconstruct why it acted. Stripe requires every tool output to be logged as an observation, giving audit trails that meet examination standards (June 2026).
Each agent needs a named business owner accountable for its outcomes, plus a central function that runs identity, gateways and eval infrastructure. The October 2026 HBS working paper counts a named human owner for agent outcomes among the operating practices it scores. The business owner sets the quality bar, which the central team checks with evals.
