On November 16, Gartner published a survey in London. The sample was 298 infrastructure and operations leaders in the United States and the United Kingdom, surveyed between June and July 2026. Put three of those numbers together, and it reads almost like a tongue twister.
95% of organizations have encountered at least one agent-related incident in the past 12 months. Among the same respondents, 70% believe their agents currently operate within safe boundaries. 57% admit that existing governance is thin.
95%
At least one agent-related incident in the past year
70%
Believe that the agent operates within safety boundaries
There is no contradiction. It only shows one thing: many things happened, but most were not recorded as "accidents".
1. Three Types of Accidents That Don't Count as Accidents
In most enterprises, the mishaps caused by agents do not fit into existing failure categories, because they do not look like failures.
First type: cyclic invocation
An agent responsible for reconciliation encountered an invoice in the wrong format and retried forty times. The system did not go down, and business was not interrupted; it was just that the month's model invocation bill was a bit higher. Finance recorded it under cloud costs, IT treated it as "an increase in usage," and no one recorded it as an incident. Yet it was clearly a loss of control.
Second type: Unauthorized configuration modification
To make the process work end to end, the agent changed a required field to optional, or slightly expanded the scope of application of a record rule. Nothing went wrong in the business that day—it even ran more smoothly. Three months later, a certain report did not reconcile, and no one remembered which change it was. In change management, this kind of thing is usually called "optimization," not an incident.
Third type: Context contamination
An outdated policy got mixed into the knowledge base, and the agent approved an expense according to the old rules. A human reviewer found it during review and corrected it, so it became “a human caught it” rather than “the system made an error.”
All three types of incidents share one thing in common: no ticket number, no severity level, and no recovery time. Traditional monitoring looks at availability—whether the system is online and whether the API returns 200. What often breaks in an agent is not availability, but judgment. And judgment has no error code.
II. Why Existing Monitoring Cannot See It
There are three more numbers in Gartner's survey: 55% of respondents said existing IT risk tools are "completely outdated"; only 15% have automatic circuit breakers installed that can stop an agent when it falls into a runaway loop; and without strong governance, by 2028, 40% of I&O organizations using AI at scale will experience a business-critical service outage.
Behind these three numbers is the same structural problem. Traditional systems report an error when they fail, while agents often return "success" when they fail; traditional systems have a clear owner, while agents are attached to a service account; traditional systems have change windows, while agents run 24/7. If you look at agents the way you look at servers, all you can see is whether they are still alive, not whether they are doing the right thing.
Recording is difficult also because it requires every action to have a business subject. When many intelligent agent projects go live, they only apply for an API key and do not even write the owner into the asset table. When something goes wrong, all that can be found is a call record and a bunch of token counts; it is impossible to find out 'for whom it is working.'
Nathan Hill's assessment in the press release is that agents are not another kind of automation script; they require a proactive approach to infrastructure governance. The report provides seven pillars: identity, guardrails, observability, agent FinOps, escalation, containment, and data boundaries.
All seven are important, but if we can only establish three first, I would choose identity, observability, and containment. The reason is not that they are the most advanced, but that they all point to the same thing: records. Without identity, records don't know who it is; without observability, records don't know what was done; without containment, records can only be reviewed after the fact and cannot save the present moment.
III. Oracle Turns Configuration into a Skill, Then Adds a Manual Approval Step
On October 26, Oracle released the Fusion Setup and Configuration Registry (FSCR) at AI World. On November 19, Fusion Insider published another introductory guide that explained its usage more clearly.
It is a public Git repository that turns Oracle Fusion configuration knowledge into machine-readable, executable AI skills, covering configuration options, dependencies, execution order, testing, and validation. AI assistants or coding agents can read business requirements, process documents, policies, and existing configurations, translate them into implementation runbooks, and then plan, execute, test, and validate directly in Fusion.
The most noteworthy part of this press release is not “AI can configure ERP now,” but the second half of the sentence: experts can review completed configurations, request adjustments, and approve deployment.
Oracle did not let the agent have the final say. It made "execution" very fast while leaving "approval" to humans. This division of labor is more concrete than any slogan about "human-machine collaboration."
Oracle Executive Vice President of Applications Development Chris Leone explained that in the past, every time customers and partners carried out an implementation, they had to rebuild critical configuration knowledge or rely on static documents that would quickly become outdated. What FSCR aims to solve is knowledge reuse, not replacing sign-off.
For people doing ERP implementation, the lesson of this matter is not in technology, but in the form of deliverables. In the past, what was delivered was documentation, screenshots, and operation manuals; now what must be delivered is executable configuration skills, corresponding test cases, and approval records for each item. These three things happen to also be the materials that can be brought out to reconstruct the scene when an incident occurs. Documentation can only prove "we said it"; testing and approval can prove "we verified it."
IV. When crossing organizations, identity issues are even harder
On November 19, the Wuhan Agent Trusted Interconnection Center launched trial operation and settled in Dongxihu District. It was initiated and operated by the China Academy of Information and Communications Technology (CAICT) and is a node in central China after Beijing, Xiamen, and Chongqing. Huang Zipei, deputy director of the CAICT Central China Research Institute, likened it to issuing “network ID cards” and “doorplate numbers” to agents, using unified identifiers to link them with developers, operators, and responsible entities.
This center must answer four questions: who it is, whom it represents, what it can do, and who is responsible when problems arise.
These four questions also hold true inside an enterprise, except that internally one can rely on institutional rules—who opened the account, who approved the permissions, and which department to contact when something goes wrong. Across organizations, this does not work: when agents from two companies call each other, emails and verbal commitments cannot make things clear. Therefore, what the industry did first was turn identity into a verifiable identifier.
Don't get the order backwards. For enterprises that cannot clearly answer the four internal questions, connecting to external networks will only amplify the chaos. First write down “who opened the account, what it can do, who is responsible,” then discuss interconnection.
V. In Odoo, how to make incidents leave a trace
ERP is the place where agents are most willing to take action, because there are documents, amounts, and approvals here. Precisely for this reason, it is best suited to play the role of an "accident ledger." The table below is categorized by action type, not by tool.
| Actions the agent needs to perform | Who approves beforehand? | What to keep during the process? | How to roll back after the fact |
|---|---|---|---|
| Read data, produce reports | Don't approve case by case; define the authorization scope once | Query and export records | Revoke permissions |
| Create and edit orders | Set approval steps by amount and status | Record chatter + field tracking | Cancel or reverse documents |
| Change configuration, permissions, master data | Must be approved manually, ideally with dual review | Configuration change request + before/after values | Roll back to a configuration baseline |
There are five things worth doing first when it comes to concrete Odoo practices.
- Dedicated service accounts with record rules. Create a separate res.users for the agent, assign groups and ir.rule by role, and don't reuse the administrator account. The account name should make it clear which agent it belongs to — don't call it "automation".
- Enable tracking on key fields. Odoo's mail.thread writes field changes into the chatter: who changed it, when, and from what to what. Watch out for two bypasses: direct SQL writes that don't go through the ORM, and writes with the tracking_disable context, neither of which leaves a tracking record.
- Route configuration changes through approval. Put switch values in ir.config_parameter, push changes through an approval request (Approvals in Enterprise; the OCA approval module or a custom state machine in Community), and retain the before and after values. For field-level auditing, you can add a layer of OCA's auditlog.
- Log attachment replacements separately. Replacing drawings, model files or firmware packages is an add/delete of attachments in Odoo, and it doesn't automatically go into the chatter. Either add version numbers and change notes to attachments, or attach key files to records that have tracking enabled — don't let "which version of the drawing was replaced" become an unsolved case.
- Give analytics agents a read-only replica only. For scenarios such as reports and data queries, let it connect to a read-only database; don't grant write access on the main database. Once permissions are given out, taking them back is often harder than giving them.
VI. Five Questions for the Incident Ledger
There's no need to rush out and buy a governance platform. First use five questions to test your current foundation. All five are retrospective; if you can't answer them, it means your records aren't sufficient.
- In the past seven days, which orders did it change? Can you pull the full scope in one go?
- Which of those were approved by a person? Who approved them?
- Was there any change that didn't take effect, or that took effect too broadly?
- Which version of the rules, prompts and knowledge base was in use at the time?
- If you wanted to stop it right now, how many steps would it take? Would you be stopping it, or the entire business line?
If you can't answer three of the five questions, improve your records first — don't rush to a platform. A platform solves "seeing"; records solve "having something to see."
The ledger itself doesn't need to be complex. Five fields are enough: time, action, impact, how it was discovered, and how it was recovered. Scan the distribution once a week and you'll see things that never show up in weekly reports — for example, a certain agent changes a few extra orders every Friday afternoon because that's when batch price adjustments happen; or a certain type of error is always caught by human review, so it never makes it into the failure statistics. These patterns tell you where to change things next better than any governance framework.
One action you can take this week: replay testing. Randomly pick an order written by an agent and replay it in a test environment under the rules in effect at the time, to see whether you get the same result. If you don't, it means at least one of the rules, data or model isn't versioned. That says more about the problem than any governance plan.
In that Gartner survey, the number most worth copying down isn't 95% — it's 70%. 95% shows incidents are widespread; 70% shows we haven't even agreed on "what counts as an incident."
An agent doesn't become safe because it's trusted; it becomes manageable because it's recorded. Give it a name, give it a ledger. Once incidents have names, governance has truly begun.
Data sources in this article: Gartner press release dated 2026-11-16, "Gartner Says Weak AI Governance Will Lead to Business-Critical Service Disruptions for 40% of I&O Organizations by 2028" (298 I&O leaders in the US and UK, surveyed June–July 2026; 95% incident rate, 70% awareness of safety boundaries, 57% weak governance, 55% outdated risk tools, 15% automatic circuit breaking, 40% forecast; Nathan Hill quotes and seven governance points); Oracle press release dated 2026-10-26 and the Fusion Insider 2026-11-19 overview (FSCR public Git repository, configuration skills, expert review and approved deployment, Chris Leone quotes); Hubei Daily 2026-11-20 report on the trial operation of the Wuhan Agent Trusted Interconnection Center (initiated and operated by CAICT; Huang Zipei's "network ID card / house number" phrasing and the four questions). Odoo models and fields are illustrated using versions 17 to 19; version 20 may change them, so check the official documentation before implementation; the three types of incidents in the article are a combined description of common situations and do not correspond to any single company.
Zhongdi Information · Enterprise Digital Management Observation · The company cases cited in this article are quoted from public reports and are not accompanied by third-party audits.
