AI challenges are high on CIOs' to-do lists for 2027, and they don't end when a tool is approved. They stretch from evaluating technology before rollout to managing how employees use it afterward.
Before rollout, CIOs need to know what problem a tool should solve and how they’ll judge whether it works. Once employees start using it, new questions can arise. A vendor might replace the model the company tested, or a business team might let an agent approve a transaction on its own.
To explore those and other AI challenges, TechTarget asked seven technology leaders what companies should prepare for in 2027.
Several focused on how to test agents and set limits on what they can do. One urged repeat testing whenever a vendor changes an underlying model. Others emphasized measuring business results, giving AI enough context to interpret company data and deciding which requests to send to outside providers and which to handle on self-hosted models.
Here’s what they had to say.
What AI challenge should companies prepare for now?
Michael Corrigan, CIO at World Insurance Associates
The biggest challenge is to ensure your AI investments are outcome-based.

This means they tie back to a specific business challenge the business or IT is trying to solve.
Make sure you can define what is measurable in that outcome. Be able to compare the new outcome to the old outcome based on those defined metrics. We shouldn’t lose sight of that.
Mojgan Lefebvre, chief technology and operations officer at Travelers, an insurance company
One of our most senior data scientists says, “AI is the victory lap you take if your data is in order.” So, if you want to take the victory lap, make sure you've got your data in order, meaning the quality of the data, the classification of the data and the understanding.
Ensure that the data can be translated into business context. The way AI understands data is different from how humans do.
Andy Boettcher, chief innovation officer at DoubleTrack, an IT services and consulting company
Running a mix of models rather than just one is a challenge to address. In 2027, companies will balance big frontier models with open models they host themselves, using routers that send each request to the right model based on classification rules.
I run three open models on my own machine next to a frontier model, and the open ones are right on their heels. Prepare now by deciding which data and questions can go where, so the most sensitive work never has to leave the building.
Frank Meltke, CEO of Contraco, a digital transformation consultancy
The model underneath an approved system will change without anyone approving it. Vendors update and replace models continuously, and an assistant or agent that a company tested and signed off in spring can behave differently by autumn while carrying the same name, the same permissions and the same place in the process.
Very few companies treat a model change as an event that triggers a new review. They should. Record which model and version each production system runs on; ask vendors to give notice before any change; and re-run the original acceptance tests whenever a change occurs, especially for systems that can approve, buy or change records on their own.
Sean Safieh, CIO of global platforms and digital solutions at Sedgwick, an insurance claims management company
One of the biggest challenges is governing AI agents as they move from summarizing information and making recommendations to taking actions, such as routing work or triggering transactions. That shift introduces a different level of operational risk, particularly in complex enterprise environments.

When a person makes a decision, accountability is generally clear. With an AI agent, organizations must establish who authorized its access, what actions it can take, how those actions are monitored and whether they can be audited and explained. These are not just technical questions. They affect risk, compliance and how much trust the business can place in the technology.
We view AI governance as a business discipline, not simply an IT control. CIOs should start by treating AI agents like any other identity in the enterprise, with clearly defined permissions, appropriate oversight and a reliable audit trail. They also need to work with business and risk leaders to establish accountability before agents are given greater autonomy.
Boris Kolev, global head of technology at JA Worldwide, a youth education nonprofit
A big challenge is the shift from AI that answers questions to AI that takes action.
For the last few years, the main question has been whether we can trust what an AI tells us. With autonomous agents, the more important question becomes, “What are we willing to let the AI do?”
Once an agent can modify data, initiate transactions, communicate externally or trigger other systems, an AI mistake becomes an operational risk rather than simply a bad answer.
Companies need to start designing identity management, permissions, audit trails, spending limits, escalation rules and human override mechanisms specifically for AI agents, before autonomous systems become deeply embedded in critical workflows.
Daniel Herrington, chief AI officer at The Zebra, an insurance comparison service
If you didn't invest in synthetic evals, you're behind, and you'll need to ramp that up to ensure the agentic agents handle data correctly. It's naive to think that traditional security solutions like two-factor authentication or CAPTCHAs will matter. The security platforms will ultimately lose the cat-and-mouse game.
Tim Murphy is a reporter covering IT strategy for Informa TechTarget, with a focus on IT leadership, governance, workforce skills and AI strategy. His enterprise technology coverage has earned multiple Azbee Awards.
Editor’s note: Responses were edited for length and clarity.