AI agent governance failures are on course to derail four in ten enterprise deployments by 2027, according to Gartner, which published its warning in May 2026. The analyst firm predicts that 40% of enterprises will demote or decommission autonomous AI agents because governance gaps go undetected until something breaks in production.
The Governance Gap Behind the Gartner Warning
Shiva Varma, senior director analyst at Gartner, attributed the root cause to a binary mindset: organisations treat agents as either fully locked down or fully trusted, rather than calibrating controls by autonomy level and scope. Gartner’s own model defines four levels of autonomy. Level 1, labelled ‘Observe’, covers agents with read-only access to defined data sources that display results only to the requesting user, with access controls treated as a separate variable from autonomy level itself.
The implication is that most enterprises are skipping the graduated steps and deploying agents at higher autonomy levels before the governance infrastructure exists to support them. Three digital leaders at Snowflake Summit 2026, held June 1–4 at Moscone Center in San Francisco with more than 20,000 attendees across 500 sessions, described how they avoided that trap. Their lessons break into three themes: build formal frameworks, rely on expert analysts, and treat data as something to monetise.
Frameworks First: How Whoop Scaled Its Agent Programme
Matt Luizzi, VP of analytics at wearable technology specialist Whoop, said the firm collects biometric data around the clock to power its health and wellness platform, with Snowflake supporting internal analytics. Whoop was listed among the initial customers for Cortex Code, the product that Snowflake has since rebranded as CoCo and formally announced at the summit on 2 June 2026 as a coding agent that automates workflows, develops apps, and operationalises AI on enterprise data through a simple prompt. The earlier Cortex Code launch described Whoop’s use case as covering complex data engineering, analytics, machine learning, and agent-building tasks in natural language.
Luizzi said Whoop started its CoCo rollout with the analytics team alone, people who could evaluate query responses quickly and flag errors. ‘Now we’re at the point where we’ve created more formalized evaluation frameworks and are starting to roll agents out at scale,’ he said.
Software engineers now use CoCo to run A/B tests, analyse results, propose the next feature, test it, and iterate. ‘This approach is rapidly accelerating the way that we’re shipping not only business value, by automating the experimentation framework, but also the customer value,’ Luizzi said. A core lesson from that journey: context is everything. ‘That meant really leaning into the semantic layer and making sure the context is in a structured place,’ he said. ‘Building repeatable frameworks that enable us to scale these AI workloads is something that we’re taking forward with us.’
Expert Coaching Closes the AI Agent Governance Failures Gap at Fanatics
Madeleine Want, VP of data at Fanatics, manages data engineering, data science, and machine learning across the company’s betting and gaming division. Her team’s experience reinforces the Gartner warning from a different angle: agent quality tracks data quality directly.
‘When we began experimenting, we weren’t sure what would stick and what would slip, but we found that what stuck was the better the condition of the underlying data and the better the governance of it, the more easily the LLM was able to derive meaning and answer questions effectively,’ she said.
Want said early wins came in domains that were tightly bounded in context and where expert analysts understood the business domain end to end. Those analysts coached the agents. Over time, the investment required in the context layer has fallen, as has the degree of supervision an agent needs before answering questions on its own. ‘Our ability to measure the accuracy of the answers is increasing, because we’re now introducing scaled evaluation frameworks, which are helping us have confidence in how agents are answering when we’re not looking, which is kind of the whole point,’ she said.
Fanatics is now embedding APIs and agent responses into third-party tools so that data-powered insights reach people through the channels where they already work. ‘Users want to go further and do more with operational use cases,’ Want said.
Monetising Data: The Synopsys Approach
Sriram Sitaraman, CIO at software specialist Synopsys, said his team recognised around 18 months ago that AI agents could handle tasks typical of junior employees: running quick queries, creating graphs, and surfacing insights. Synopsys deployed a revenue agent for the finance department and a debug agent for its data-centre ticketing system.
Evaluating agents across three dimensions (quality of results, time to results, and cost of results), the team found AI performed well on all three. Sitaraman said that combination had previously required trade-offs: ‘In the past, you had to sacrifice one or the other.’
His advice is direct: ‘Start with data, monetize your data using AI. It doesn’t matter how much volume you throw at the initiative, because AI is just truly a linear scale. The more data AI has, the better decisions it makes.’
He also issued a warning on scope creep. ‘You can roll out an agent and say, “This is a sales ops agent.” Often, there’s nothing to stop it from also becoming a sales analyst agent or another type of agent,’ he said. ‘So, it’s important to ask, “Is this what we want it to do?” Frameworks are very important, as are skills. You need to think the process through carefully.’
That caution maps directly onto Gartner’s governance levels. The question for any enterprise deploying agents is not whether to govern them, but at what autonomy level each agent is cleared to operate, and whether that decision was made before or after the first incident.