Enterprise AI is turning into an operations problem, according to reporting and surveys published between September 14 and 18, 2026. Companies can buy strong models. What they struggle with is routing work between them, keeping data trustworthy and knowing who oversees the agents running on top. Call it the “enterprise becoming operations” shift. It is real, but the evidence needs a closer read than the headlines give it.

AI Business contributing writer Liz Hughes connected the pieces in a September 18 analysis, citing a Deluxe case study, a Collibra survey and an EY report. I went to the underlying sources instead of relying on that summary, including EY's survey page, Collibra's press release and Grant Thornton's survey page, which InformationWeek cites.

Here is the short version of what I found. The direction holds up across three separate surveys and one company case study. The exact percentages come mostly from firms that sell governance, data or assurance services, so I have labeled what is confirmed, what is a company claim and what is my own read.

By Alex Carter. Published September 19, 2026. The source reports were published September 14 to 18, 2026.

Background: From Picking a Model to Running Many

The old enterprise AI question was which model to buy. The 2026 survey data asks a different one: who runs all of it?

Grant Thornton surveyed 950 business leaders between February 23 and March 18, 2026, and its 2026 AI Impact Survey found 78% lacked strong confidence they could pass an independent AI governance audit within 90 days. EY surveyed 202 senior AI decision makers at publicly traded US companies with at least $1 billion in annual revenue between May 28 and June 15, according to Cybersecurity Dive. Collibra and The Harris Poll surveyed 306 data, privacy and AI decision makers from August 5 to 11, per Collibra's press release.

The publishing wave came later. InformationWeek ran its Deluxe piece on September 14. EY released its findings on September 15 and Collibra on September 16. CIO Dive covered Collibra on September 17, and AI Business tied the threads together on September 18.

That timing matters. The “this week” framing rests on fieldwork done between late February and mid August. The headlines are new, but the data is not.

I have seen this shape before. Cloud adoption followed a similar arc: once companies scaled, cost and security tooling became product categories, and the sellers were quick to publish surveys showing the need. That is my read from earlier cycles, not something these reports say.

The Current Picture: Enterprise Becoming Operations in Three Places

Model choice is now a standing job

At Deluxe, a Minneapolis payments and data company, code never goes straight to a model, InformationWeek reported. A gateway sits in between and decides where each request goes, depending on which of the platform's more than 50 AI agents is doing the work. The report lists GPT-5.6, Claude Opus, Claude Sonnet or a model still under evaluation as possible destinations. The developer does not choose, and IT does not quite choose either.

Deluxe weighs quality, risk, latency, economics and operability. Chief technology and digital officer Yogaraj Jayaprakasam told InformationWeek: “Put a gateway between your applications and your models before you scale, not after.” He also argued that a cheaper model that triggers more retries and human intervention can end up costing more. That is an executive's claim, not an independent measurement.

Outside advisers say the seams between models are the weak point. Grant Thornton partner Sumeet Mahajan said public benchmarks score most leading models similarly, and InformationWeek's report calls them easy to game. He suggested building a private benchmark from cases where a model already failed in production. Michael Adler, director of AI governance and data protection at the law firm Akerman, described each vendor logging in its own format, leaving no single record of what happened.

Data is where projects stall

Collibra's survey found that 72% of respondents agree that when AI initiatives fall short, the root cause almost always traces to an unaligned or poor data foundation. In the same release, 87% said their teams regularly re-verify that agent context is accurate and current, and 51% reported significant staff hours spent reviewing and correcting agent outputs before they go live. Meanwhile, 53% said their AI function's reporting line has moved closer to the data organization over the past 12 months.

CIO Dive adds an outside data point. An August report from Google Cloud and MIT Technology Review Insights, as CIO Dive describes it, found that AI can access an average of 45% of an enterprise's data, and that only half of organizations trust their agents' outputs to be relevant and accurate. I have not read that report directly, so treat those figures as secondhand.

Collibra CEO Felix Van de Maele told CIO Dive: “Without that visibility, governance gaps aren't discovered until something goes wrong.” He was talking about knowing which agents are running, who owns them and what they can access.

Oversight is lagging deployment

EY's AI Risk and Governance Survey, published September 15, found that 98% of respondents have formal AI governance policies. Among respondents whose organizations use agentic AI, 85% said they have at least a handful of agentic systems acting without real time human intervention, 49% said governance has not been updated for agentic risk, and 26% said they cannot detect unauthorized agents operating internally.

Cybersecurity Dive's report on the survey adds that nearly six in 10 of those respondents perceived that no single group oversaw agents after deployment. Across all respondents, 47% said their organization had bypassed its AI governance process for an urgent deployment, and 36% reported an AI incident with a materially negative impact.

EY Americas Technology Risk AI Leader John McLain framed it this way: “The biggest agentic AI risk is that human oversight hasn't evolved accordingly.” Grant Thornton's numbers lean the same way, with only 20% of its respondents reporting a tested incident response plan for AI failures.

Permissions are turning into their own product category, too. Amazon Bedrock AgentCore Identity now ships a managed consent portal for agents that handles end user OAuth consent, as BriefFlash has covered.

Reality Check: What Do These Surveys Actually Prove?

They prove direction, not magnitude. Every survey here points the same way, but each comes from a firm with a product or service aimed at the problem it measured.

Collibra calls itself the enterprise AI control plane in its own release. Its 72% figure measures agreement with a statement about data foundations, put to data management, privacy and AI decision makers. My read is that this group has a stake in the answer. EY and Grant Thornton both sell AI assurance or advisory work, and both reports end with steps for closing the gap they measured.

The samples are modest. Collibra reports 306 respondents and a margin of plus or minus 6.4 percentage points. EY's 202 respondents all sit at US public companies with at least $1 billion in annual revenue, so the survey says little about mid sized firms. The freshest fieldwork, Collibra's, ended August 11, about six weeks before this piece.

The surveys also pull in different directions on accountability. Collibra reports that 84% of respondents say their organizations have clearly defined executive accountability when agents produce flawed or harmful outputs. EY found nearly six in 10 agent using respondents perceived no single group overseeing agents after deployment. Different respondents, questions and months explain some of that gap, but anyone quoting one number without the other is picking a side.

Two sourcing flags. The Deluxe account, including the gateway design, rests on one InformationWeek interview, which AI Business then repeated, and I found no independent confirmation. Also, InformationWeek describes Grant Thornton's 950 respondents as senior IT leaders, while Grant Thornton's methodology lists 390 from operations, 313 from finance, 234 from IT and 13 CEOs or managing partners.

What holds up is the pattern. Companies running more than one model and more than a handful of agents need routing rules, an inventory of what is running, an incident plan and a named owner. None of that depends on a vendor's percentage.

Who This Affects

  • CIOs and platform teams. In InformationWeek's account, platform teams build the tools and guardrails while product teams own adoption and outcomes. Mahajan described a similar split, with a central team owning routing and business units making task level calls.
  • Data and governance leads. Collibra's 53% reporting line finding suggests the org chart is already moving toward the data organization. Whether that is a durable shift or a survey artifact is still open.
  • Business unit owners. Adler said each deployment needs one person who owns the routing decision and can pause it if something goes wrong. Nearly four in 10 EY respondents at agent using organizations called accountability for overseeing agents undefined, so many companies have not filled that seat.
  • Smaller companies and builders. EY sampled only companies with $1 billion or more in revenue, so its numbers should not be read as a picture of small teams. Gateways and agent inventories are large company answers, and I would not scale the percentages down to a 50 person shop.

What's Next

What would confirm the enterprise becoming operations shift is independent data showing the same gaps. What would undercut it is data showing the overhead is small or shrinking, which would make this a vendor narrative. Here is what I am watching.

  • Full reports. EY's full report sits behind a download form, and Collibra says its complete 2026 Hallucination Tax Report is on its site. I want the question wording on accountability, since it could reconcile the 84% with the nearly six in 10.
  • Independent numbers. A survey from a group that sells neither governance nor data tools would either confirm the scale or shrink it. I have not found one I can verify.
  • Audit rules. CIO Dive notes that California Governor Gavin Newsom recently signed two AI bills setting up a framework for independent audits of large language models, and that EU AI Act literacy obligations and prohibitions took effect in August. If audits become routine, demands for evidence that controls work move from consultant slides to compliance work. That is my read, not a forecast from those reports.
  • Open weight models. AI Business's roundup says calls to slow frontier AI development could leave enterprises with more responsibility for testing, monitoring and governance of open weight models. The next round of surveys should show whether companies are staffing for it.

Related BriefFlash Coverage

Corrections: if any figure here changes, I will add a dated “Updated” note at the top of this article.

Key Takeaways

  • Enterprise becoming operations is the pattern across reports published September 14 to 18, 2026: EY found 49% of respondents at agentic AI organizations had not updated governance for agent risk, and Collibra found 72% agree poor data foundations are usually the root cause when AI initiatives fall short.
  • Deluxe, a payments and data company, routes requests from more than 50 AI agents through a gateway that picks the model, according to InformationWeek. It is the clearest single example of model choice becoming an ongoing job.
  • The figures come from EY, Collibra and Grant Thornton, all of which sell services aimed at these problems, with samples of 202 to 950 respondents and fieldwork between February and August 2026.
  • The Deluxe details rest on one interview, and the surveys disagree on accountability: 84% of Collibra's respondents report clear executive accountability, while nearly six in 10 in EY's agent using group perceive no single overseer.

FAQ

What does enterprise operations mean?

In general business usage, enterprise operations means the running of an organization's core processes and systems across the whole company, not inside one team. The reports behind this story do not use the phrase directly. They describe the same idea for AI: choosing and routing between models, keeping data reliable, controlling what agents can access and deciding who oversees them once they are live. InformationWeek's Deluxe example is the clearest case, with a central gateway and a split between platform teams and product teams.

What are the four types of enterprise?

There is no single official list, so the premise needs a caveat. As general background rather than something from these reports, the most common “four types” answer in US business law is sole proprietorship, partnership, corporation and limited liability company, though other sources sort enterprises by size or ownership instead. None of the AI reports discussed here use that framework. In enterprise AI coverage, “enterprise” simply means a large organization. EY, for example, surveyed publicly traded US companies with at least $1 billion in annual revenue.

What is enterprise operations management?

Generally, it is the practice of planning, coordinating and monitoring how a large organization's processes, people and systems work together. For AI, the reporting shows it taking concrete forms: a routing gateway and independent audit record across model vendors at Deluxe, inventories of running agents with named owners, and tested incident plans. Grant Thornton found only 20% of respondents had a tested AI incident response plan, so by that measure the practice is still early.