Skip to content
Back to ReadyWorks Blog

Are Your AI Agents One Bad Command Away From an Outage?

Most AI failures in IT start before the AI does anything wrong. They start when an agent gets access to systems without clean data, scoped permissions, or a person who approves each change. The fix is governance built in…

Are Your AI Agents One Bad Command Away From an Outage?

Most AI failures in IT start before the AI does anything wrong. They start when an agent gets access to systems without clean data, scoped permissions, or a person who approves each change. The fix is governance built into the platform the AI works on: one continuously cleansed view of the estate, application context that shows what else a change will touch, human approval before every write, and a tamper-evident record of every action. For a CIO planning a large AI move, the partner you choose should bring all of that from the first day and should test its own AI before it reaches your environment. That is how ReadyWorks puts AI to work on infrastructure.

Why do agentic AI projects fail in IT operations?

Look back over the last six months and a pattern shows up. The headlines talk about AI going wrong. The details tell a different story.

  • May 2026: Law firm Seyfarth Shaw wrote up the PocketOS incident. An AI coding agent was given a routine fix in a staging environment. It hit a credential mismatch, found a token in an unrelated file, and used it to delete a storage volume at the company's cloud host. The token had broad permissions, and the backups lived on the same volume. One command took out the production database and its backups. The outage ran more than 30 hours before the data was recovered.
  • September 3, 2026: Claude, Grok, and ChatGPT all reported incidents on the same morning. Info-Tech Research Group found no public evidence of a shared cause. Its advice to enterprises was to map the full dependency chain behind each AI workload and decide in advance what each one does when its provider goes dark.
  • November 12 and 13, 2026: An Exchange Online outage delayed email across Microsoft 365. AI assistants that connect to Microsoft 365 failed for most requests, and Maersk warned customers that emails and updates could be delayed. Writing in Forbes, Robert Szczerba pointed out that an empty inbox and a stuck one look identical to an agent. An agent that reads silence as "nothing new" will keep acting on old information.

None of these failures came from an AI that went rogue. Seyfarth's headline says it plainly: the guardrails were never there. The PocketOS agent had access it never needed and no one approving a destructive change. The outages exposed AI that depended on systems no one had mapped. The Exchange outage showed agents acting on data no one could confirm was current.

The research points the same way. In an April 2026 Gartner survey of 782 infrastructure and operations leaders, 57% reported at least one failed AI initiative, and 38% said poor data quality or limited data availability directly caused an AI project to fail. Gartner also predicts that over 40% of agentic AI projects will be canceled by the end of 2027, citing escalating costs, unclear business value, or inadequate risk controls. And in its Innovation Insight on Agent-Native I&O, Gartner expects that by 2028, 99% of I&O-led investment in agentic AI without system data and operating model improvements will fail to achieve sustainable ROI.

The market is responding. ReadyWorks puts AI to work on the infrastructure you already run, on a foundation that is in place before AI touches production: integrations across every vendor in your estate, clean data, application context, ready-made reports, preconfigured workflows, and security and guardrails the organization sets. ReadyAI analyzes, makes recommendations, and carries out the work within those guardrails, across day-to-day operations and transformation programs on one platform. Each of the failures above points to a specific control, and each control is part of that foundation.

What guardrails should AI agents have before they change production infrastructure?

Here is how those controls line up with each incident in the timeline:

Incident

What was missing

How ReadyWorks addresses it

PocketOS database deletion

Scoped access and approval before a destructive change

Role-based access on every call, a tiered write policy the organization sets, and human approval before every write

Microsoft 365 outage

A way to confirm the data was current, and a shared record of what had already been done

A unified data foundation that is continuously cleansed, plus a tamper-evident audit trail of what changed, who approved it, what it touched, and the outcome

Overlapping AI provider outages

A map of dependencies and a plan for when the model is unavailable

Application context that shows what else a change will touch, model choice including customer-hosted models, and a platform that delivers full value with AI off

 

  1. Start with access. The PocketOS agent used a token created for something else, with permissions far beyond its task. ReadyWorks applies the signed-in user's role-based access on every call. On top of that, the organization sets a tiered write policy: read-only, confirm every write, or confirm with blanket approval up to a record ceiling. Today ReadyAI recommends, a person approves, and then it carries out governed platform actions.

  2. Next is data. An agent is only as good as what it reads. The Exchange outage showed what happens when an agent treats a gap in its data as a quiet period. ReadyWorks connects to the systems you already run, accesses data on demand and in place, and keeps one normalized, continuously cleansed view of the estate. Write-back to any source system is optional, per system, opt-in, and fully audited. Your systems of record stay in place.

  3. Then context. Before any change, your team should know what else it will touch. ReadyWorks adds application context to every asset: dependencies, usage, compatibility, and lifecycle stage. That is the dependency map Info-Tech recommends, built into the work instead of drawn once on a whiteboard.

  4. Finally, the record. Every action ReadyWorks takes is logged, explainable, and covered by a tamper-evident audit trail. When a regulator, an auditor, or your board asks what happened, the answer is already on record. The guardrails also include optional pseudonymization, prompt-injection defense, and a hash-chained audit.

Go back to the provider outages for one more point. ReadyWorks lets you bring your own model, including Anthropic, Azure OpenAI, AWS Bedrock, OpenAI, Google Vertex, or a model hosted in your own environment. With a customer-hosted model, no data leaves your network. And because the data foundation is the product, the platform keeps delivering full value with AI turned off.

How should enterprises choose an AI partner for infrastructure automation?

Large AI moves raise the stakes on every one of the controls above. A CIO is balancing the pace the business wants against legacy systems, cost pressure, and compliance risk. The partner you pick decides how much of that risk you carry. These are the questions worth asking before any contract is signed:

  • How is the AI tested before a release reaches us? ReadyAI is tested on every release against a fixed set of real questions on real environments, with a judge that can block the release.
  • Who approves a change, and who sets that policy? With ReadyWorks, the organization sets the guardrails and a person approves every write.
  • Where does the AI get its data, and how is it kept clean? Ask to see how the data is normalized and how quality is scored across every vendor in your estate.
  • Can we see what else a change will touch before it runs?
  • What happens when the model is down? Ask whether the platform still works with AI off.
  • Can we start small and widen the scope as trust grows?

Put those questions to every new launch, including the ones that make headlines. On November 13, Virtana launched Agentic Observability Enterprise Management, which brings AI agents from different platforms into one shared incident response, from investigation through remediation. It is available to select customers now, with general availability planned for Q1 2027. Read past the headline and it includes policy controls and approval-based remediation. Then ask where it starts. It is built for the moment after something breaks: an incident is detected, and agents work together to diagnose and fix it.

ReadyWorks starts earlier, before the change. It puts AI to work on daily operations and on the planned programs where change happens at scale, such as VMware exits, endpoint refreshes, and server decommissioning. In a VMware exit, ReadyAI analyzes the estate, recommends, and plans the waves. A named delivery team, whether a services partner, ReadyWorks, or your own staff, executes each wave with your sign-off at every phase. AI-planned. Expert-delivered. Incident response and governed change both matter. A CIO choosing a partner for a large AI move needs to know which one a product was built for.

Who trusts ReadyWorks with this work?

Global partners and Fortune 100 customers use ReadyWorks to take months and millions out of infrastructure work. ReadyWorks has appeared in 26 Gartner Sample Vendor reports across 10 Hype Cycles since 2022. Its founders have run enterprise infrastructure programs for over 30 years, and its executives have delivered them for Fortune 100 customers.

Large AI moves don't happen in one step, and they don't have to. ReadyWorks meets you where you are and delivers value at every stage. You set the pace. Every stage runs on the same platform, and you hand more of the work to ReadyAI as your confidence grows. Most estates are still working toward the first stage, Foundation: identify every system, connect them, clean the data, add application context, and set the guardrails. Only then do teams build workflows, test them, and run the first program. Map your own path with the ten steps to autonomous IT operations. The tool prints a plan built from your answers. Bring it to the ReadyWorks team and we'll help you take the first step.

Talk to someone who has done this before your next AI rollout

The incidents above were ordinary: a broad token, a data gap, and a dependency no one mapped. Each one is common in large estates today. ReadyWorks gives your AI a governed foundation before it touches production: clean data, application context, human approval before every write, a tamper-evident record of every action, and a model you choose. Your next AI rollout will run on the data and guardrails you have when it starts. Talk to us today and we'll put them in place first.


Frequently asked questions

Why do agentic AI projects fail in IT operations?

Most fail because of missing governance and poor data. Gartner's April 2026 survey found 38% of I&O leaders traced an AI project failure directly to poor data quality or limited data availability. Incidents like PocketOS show the other half: broad access and no human approval before a destructive change.

What guardrails should AI agents have before they change production infrastructure?

At minimum: role-based access on every call, a write policy the organization sets, human approval before every write, a continuously cleansed data foundation, application context that shows what else a change will touch, and a tamper-evident audit trail. ReadyWorks provides each of these on one platform.

How should enterprises choose an AI partner for infrastructure automation?

Ask how the partner tests its AI before each release, who approves changes, how data is kept clean across vendors, whether you can see what a change will touch, and whether the platform still works with AI off. ReadyWorks tests ReadyAI on every release against real questions on real environments, with a judge that can block the release.

How is ReadyWorks different from Virtana's AI agent war room?

Virtana's new product coordinates AI agents during IT incidents to investigate and fix problems once they occur. Both products include policy controls. The difference is where they start. ReadyWorks starts before the change: it gives AI a continuously cleansed, multi-vendor view of the estate with application context, and puts AI to work on daily operations and transformation programs, such as VMware exits, endpoint refreshes, and server decommissioning, within guardrails the organization sets. In a VMware exit, ReadyAI analyzes, recommends, and plans, and a named delivery team executes each wave.

Does ReadyWorks replace our CMDB or monitoring tools?

No. Systems of record stay. ReadyWorks connects to them, accesses data in place, and retires single-purpose tools such as readiness spreadsheets and custom reporting scripts.

Can ReadyWorks undo a change?

For endpoint refresh and Windows migration waves, EndpointReady includes automated rollback. Across the rest of the platform, every action is logged, explainable, and covered by a tamper-evident audit trail, so your team always knows what changed, who approved it, and what it touched.

Newsletter

Agentic ITOps, in your inbox.

Analysis and product news from the ReadyWorks team. No filler, unsubscribe any time.

By subscribing you agree to our privacy policy.

readyworks cog Keep reading

Related posts