As of May 2026, at least 70% of large enterprises are experimenting with AI agents internally before considering their use in customer-facing applications, according to a report by AI Business. This shift in approach reflects concerns about the potential risks and rewards associated with deploying AI technology.
Key Takeaways
- 70% of large enterprises are experimenting with AI agents internally before deploying them to customers.
- Enterprises are using smaller testing teams and strict governance to manage the risk of AI deployment.
- The majority of enterprises are focusing on developing and testing AI agents before considering their use in customer-facing applications.
- Enterprises are prioritizing the development of AI agents that can be used in a variety of industries and applications.
- The report highlights the need for enterprises to carefully consider the potential risks and rewards associated with deploying AI technology.
Enterprises Contain AI Agents Internally
The report by AI Business highlights the growing trend of enterprises containing AI agents internally before deploying them to customers. This shift in approach reflects concerns about the potential risks and rewards associated with deploying AI technology.
According to the report, at least 70% of large enterprises are experimenting with AI agents internally before considering their use in customer-facing applications. This internal testing allows enterprises to assess the potential risks and rewards associated with deploying AI technology and to develop and refine their AI agents before deploying them to customers.
Internal experimentation isn’t new — enterprises have historically sandboxed emerging technologies before rollout. But with AI agents, the stakes are higher. Unlike earlier automation tools, AI agents can make decisions, initiate actions, and interact with systems and people with minimal oversight. That autonomy demands tighter controls. Enterprises aren’t just worried about bugs or performance. They’re worried about unintended behavior, data exposure, and reputational damage.
In 2024, a major financial institution tested an AI agent designed to assist with customer service routing. During internal trials, the agent began escalating non-urgent cases based on tone analysis, overloading human staff. It wasn’t malicious, but it exposed how easily an agent’s logic could misfire without direct harm. The issue was caught before public release — exactly the kind of scenario internal containment is meant to prevent.
That case, while not cited in the AI Business report, mirrors the pattern it documents: test in private, learn fast, fix quietly. The 70% figure suggests this isn’t just a few cautious outliers. It’s becoming standard operating procedure.
Testing Teams and Governance
Enterprises are using smaller testing teams and strict governance to manage the risk of AI deployment. These testing teams are responsible for developing and testing AI agents, as well as assessing their potential impact on the enterprise and its customers.
These teams often pull from compliance, security, legal, and engineering departments. That cross-functional structure is intentional. An AI agent might be built by engineers, but its behavior touches data policies, customer trust, and regulatory boundaries. A single oversight can trigger audits, fines, or PR crises.
Governance frameworks vary, but most include version tracking, approval checkpoints, and audit trails. Some enterprises require dual sign-off before an agent moves from testing to production. Others limit access to sandbox environments, ensuring only authorized personnel can modify or observe agent behavior.
One retail enterprise described in industry discussions — though not named in the AI Business report — runs AI trials with a “two-person rule”: no agent can be retrained or redeployed without both a technical lead and an ethics reviewer signing off. That’s not required by law, but it’s a response to growing pressure from boards and regulators to treat AI like any other high-risk system.
The size of these teams matters too. Smaller groups move faster and document decisions more consistently. A team of five can align on edge cases; a team of fifty might miss them. That doesn’t mean enterprises are skimping on resources. They’re concentrating them — opting for precision over scale during the experimental phase.
And experimentation is intense. Agents are run through hundreds of simulated workflows: handling customer complaints, processing invoices, summarizing internal reports. Each test logs how the agent interprets inputs, what decisions it makes, and whether it stays within defined parameters. If it veers off-script, the team investigates why. Was the training data skewed? Did the prompt allow ambiguity? These aren’t just technical questions — they’re governance issues.
Focus on Development and Testing
The majority of enterprises are focusing on developing and testing AI agents before considering their use in customer-facing applications. This approach allows enterprises to refine their AI agents and to develop strategies for mitigating the potential risks associated with their deployment.
According to the report, enterprises are prioritizing the development of AI agents that can be used in a variety of industries and applications. This focus on development and testing reflects the growing recognition of the potential benefits and risks associated with AI technology.
The push for broad applicability isn’t just about efficiency. It’s about ROI. Building an AI agent is expensive. Training, infrastructure, monitoring — it adds up. Enterprises want agents that can be adapted, not rebuilt from scratch for every new use case.
So they’re designing modular agents: core decision engines that can be plugged into different workflows with minimal retraining. A single agent architecture might power internal helpdesk support, supply chain alerts, and compliance monitoring — just with different data inputs and guardrails.
That modularity requires rigorous testing. An agent that works in HR shouldn’t accidentally leak sensitive data when repurposed for procurement. So enterprises are investing in validation frameworks — automated checks that run every time an agent is reconfigured. These tests verify permissions, data handling, and decision logic. They’re not perfect, but they reduce the odds of cross-context errors.
Testing also includes stress scenarios. What happens if the agent loses connectivity? If it receives malformed input? If it’s asked to do something unethical? Some enterprises run adversarial simulations, where red teams try to trick agents into violating policies. These drills expose weaknesses in logic or oversight — and they’re becoming routine.
The report doesn’t specify how long testing phases last, but anecdotal evidence suggests they’re getting longer. In 2023, some companies deployed AI tools within weeks of development. By 2025, the average internal trial stretched to three to six months. That slowdown isn’t from lack of confidence in AI. It’s from a clearer understanding of what can go wrong.
Industry Implications
The report highlights the need for enterprises to carefully consider the potential risks and rewards associated with deploying AI technology. As AI technology continues to evolve and improve, enterprises will need to develop strategies for managing the associated risks and for maximizing the potential benefits.
Different industries are moving at different speeds. Financial services and healthcare, heavily regulated, are among the most cautious. They’re testing AI agents for internal reporting, fraud detection, and document review — but avoiding direct patient or client interaction for now.
Manufacturing and logistics are further ahead. Some are using AI agents to monitor equipment health, optimize routing, and predict supply delays. Because these systems often operate behind closed networks, the risk of public exposure is lower. That makes them ideal testbeds.
Retail and telecom are in the middle. They’re experimenting with AI agents for internal support and operations, but they’re also under pressure to deliver AI-powered customer experiences. That tension is forcing them to move carefully — innovate fast, but not too fast.
Across all sectors, one trend is clear: no one’s betting everything on AI agents yet. The 70% experimenting figure doesn’t mean 70% are close to deployment. It means 70% are dipping a toe in the water. And most are watching each other closely.
“Enterprises are recognizing the potential risks and rewards associated with deploying AI technology and are taking steps to carefully consider their options,” said a spokesperson for AI Business.
What This Means For You
As an enterprise, it’s essential to carefully consider the potential risks and rewards associated with deploying AI technology. By taking a cautious approach and focusing on development and testing, you can ensure that your AI agents are refined and effective, and that you’re able to mitigate the associated risks.
If you’re a developer or founder, you’ll need to consider the potential implications of AI deployment on your business and your customers. By understanding the risks and rewards associated with AI technology, you can develop strategies for maximizing the benefits and minimizing the risks.
For developers, this shift means your work is under greater scrutiny. Writing code for an AI agent isn’t like building a static app. The agent learns, adapts, and acts. That means your documentation, testing protocols, and error handling need to be rock solid. Companies won’t tolerate black-box systems — they’ll demand transparency, traceability, and control.
Founders should pay attention too. If you’re building AI tools for enterprise, your sales cycle just got longer. Enterprises won’t adopt your agent because it’s cool or fast. They’ll adopt it because it fits their governance model, integrates with their audit systems, and can be contained within their risk thresholds. That means your product needs built-in monitoring, permission layers, and rollback capabilities — not just smart features.
For enterprise leaders, this trend creates both opportunity and friction. The opportunity is clear: AI agents can cut costs, speed up decisions, and reduce human error. But the friction comes from alignment. Your legal team will want strict policies. Your engineers will want flexibility. Your board will want results — but not at the cost of a scandal.
One scenario: a mid-sized insurer tests an AI agent to process claims. It works well in trials, reducing review time by 40%. But when governance teams audit the decisions, they find the agent downgrades certain claim types based on zip code — a red flag for bias. The project stalls. That’s not failure. It’s the system working as intended. The risk was caught internally.
Another scenario: a startup builds an AI agent for scheduling sales meetings. It integrates with calendars, emails, and CRM tools. A large enterprise pilots it internally. But during testing, the agent starts rescheduling meetings based on inferred urgency — not user input. Sales teams revolt. The tool gets shelved. The startup learns it needs clearer user controls.
A third scenario: a manufacturing company uses an AI agent to predict machine failures. It works. Downtime drops. The agent is so effective, other departments want to use it. But retraining it for a new factory requires retesting, new data validation, and fresh approvals. Progress slows — not because the tech fails, but because the process holds firm.
These aren’t hypotheticals. They’re echoes of what’s happening now. The enterprises leading in AI aren’t the fastest movers. They’re the ones building the right guardrails.
Looking Ahead
As AI technology continues to evolve and improve, it’s essential for enterprises to remain vigilant and to carefully consider the potential risks and rewards associated with its deployment. By taking a cautious approach and focusing on development and testing, you can ensure that your AI agents are refined and effective, and that you’re able to maximize the potential benefits.
The next 12 to 18 months will be telling. Will the 70% of enterprises still testing begin to deploy? Or will new incidents — a data leak, a biased decision, a system failure — push them toward even greater caution?
There’s no sign yet of a rush to market. If anything, the opposite. Enterprises are treating AI agents like high-voltage systems: powerful, useful, but dangerous if mishandled. They’re not waiting for perfection. They’re waiting for confidence.
And confidence comes from control. From knowing what the agent is doing, why it’s doing it, and how to stop it if needed.
The report doesn’t predict when customer-facing AI agents will go mainstream. But it makes one thing clear: they won’t arrive by accident. They’ll arrive on schedule, on budget, and only after passing through layers of internal review.
That’s not slowing innovation. It’s shaping it.
Sources: AI Business
