When it comes to customer communications, AI is leaving pilot territory as agent rollouts continue at pace. Already, organisations Already, organisations are seeing measurable gains (faster resolution times, more consistent responses across markets…) as AI takes on a growing share of customer interactions.
However, an AI rollback economy is emerging, with nearly three quarters (74%) of organisations already rolling back or shutting down customer communication agents once they have been deployed. In the race to push agents live, many enterprises are only discovering hidden challenges or flaws once they’re out of a test environment.
In many cases, these issues are often not quick fixes. As a result, agents need to be rolled back to resolve issues to ensure they are not repeated. Instead, agents need to be designed with failure detection in mind so that with the right observability framework from the get-go to enterprises can detect issues quickly, understand the resolutions that need to be made, and then re-deploy agents confidently.
The very nature of AI means that failures are to be expected; zero issues would be impossible. Rolling back an agent is, therefore, not a failure in itself. Done as soon as an issue is identified, rollbacks don’t hamper innovation. Instead, they lay the foundation to scale communications networks safely, without allowing small issues to compound into bigger ones once an agent is handling real volume.
Why robust observability is paramount
AI is inherently probabilistic, so we cannot always control what it will say next. As a result, even the most advanced AI systems can produce errors or hallucinations. When customers are the first to see these errors directly from customer agents, the risk to brand reputation and customer trust significantly increases.
Observability is key to ensuring any issues are picked up immediately and dealt with correctly. However, accurate LLM observability requires specific tools. Rather than monitoring the overall process of each action, this means capturing full inputs and outputs, so that teams can understand in detail what actions the agent decided to take, and why. With this information, enterprises are in a stronger position to reconstruct agent activity when something goes wrong. This means that engineering teams can quickly and precisely identify where the error came from, and at what layer the repair is needed.
For example, a model hallucination that leads to an agent giving out incorrect information requires a different restorative approach than a deep-rooted infrastructure failure that stops an agent working altogether. Observability means engineering teams are immediately notified of the issue at hand and can react accordingly rather than having to waste precious time figuring out the problem across multiple data sources.
Ensuring governance does not stifle innovation
Organisations that do not design for failure monitoring or sufficiently observe their communication agents, risk an emerging operational guardrail tax. With each issue that arises, engineering teams spend more time in firefighting mode than innovating or scaling models further.
In fact, research shows that 84% of AI engineering teams report spending at least half of their time building guardrails and safety controls, even before a single failure occurs. As a result, less time is spent on initiatives that will drive new revenue streams or find cost saving efficiencies.
Ahead of rolling out agents, enterprises must ensure secure governance frameworks are in place, so that engineering teams can guardrail AI operations securely. Having the correct monitoring and observability in place means that engineers can understand the initial risk of both infrastructure and model failures while being able to take the correct action to minimise rogue activity further.
Measuring success in a new era of customer communications
As AI communication agents become table stakes, the KPIs used to measure successful customer communications must also evolve. Deployment metrics alone are no longer sufficient to measure progress. An AI agent may be live and handling a high volume of interactions but that does not necessarily indicate whether the agent is delivering accurate and trusted customer communications.
Enterprises, therefore, need to pay attention to customer satisfaction including how quickly agents resolve issues, and how they bolster customer trust and loyalty. When looking to scale a communications network, organisations should also assess whether agents are operating consistently across different markets, languages and customer groups, rather than allowing strong overall averages to conceal pockets of poor performance.
These metrics should extend beyond the customer-facing agent itself. Enterprises need visibility into the operational burden AI places on engineering, compliance and customer service teams. If employees are spending a growing proportion of their time reviewing outputs, maintaining guardrails or responding to avoidable failures, an agent may be creating hidden costs. Measuring the time required to oversee, repair and improve AI systems will give leaders a more realistic picture of whether deployments are genuinely scalable.
Building the foundations for sustainable AI
The AI rollback economy is not a sign that enterprises should retreat from AI. Many enterprises are finding that the infrastructure they built customer communications on was not designed for the operational demands AI introduces. Now is the time for organisations to fix this.
Governance and observability should not be seen as barriers that slow innovation. Instead, they are what make continued innovation possible. When enterprises can understand how agents are behaving, identify where risks are emerging and resolve issues before they reach customers at scale, they create the conditions to experiment and expand without putting brand reputation and customer trust at risk. What’s more, the same observability data that flags a failure can be used to retrain the model, so each rollback makes the next deployment more accurate rather than just safer.
Building governance and observability into systems from the get-go that can learn, recover and improve over time is the first step to successfully scaling AI customer communications.
