AI App Delivery Top 10: Incompatible delivery policies

Lori Mac VittieDistinguished Engineer and Chief Evangelist | F5

Incompatible delivery policies used to be one of those “architecture hygiene” topics that only got airtime after something broke. You know the pattern. One team makes a routing change in Cloud A, another team updates security rules in Cloud B, and suddenly prod behaves like a different species than staging. Nobody can reproduce it consistently, and the only thing you can say with confidence is that it’s “probably networking.”

ADC07 calls this out for what it is: when delivery policies don’t line up across environments, you get unpredictable performance, unreliable failover, scaling bottlenecks, and a whole lot of operational misery.

Now add AI, and this stops being a slow-motion mess and becomes an accelerant.

Impact on performance

Incompatible delivery policies show up as performance problems that feel random, and random is always the worst kind of performance problem because engineers can’t “fix random.” They can only chase it.

AI makes troubleshooting truly painful because AI interactions are often multi-step chains. A single user request might trigger retrieval, reranking, inference, tool calls, and then another inference pass. If any stage crosses an environment boundary where policies differ, you’ve just introduced jitter into the whole experience.

Incompatible policies introduce inefficient routing and increased latency across providers even when delivering traditional or modern applications. That’s the baseline reality. In AI delivery, that inefficiency shows up as a system that feels hesitant and uneven even when everything is technically “up.”

Impact on availability

This is where incompatible delivery policies graduate from nuisance to outage generator. Availability depends on consistent behavior under stress, and incompatible policies tend to surface precisely when you are under stress.

Failover is the classic example. If one provider’s failover logic doesn’t align with another’s, you can end up with traffic that can’t be redirected cleanly or gets redirected into a region that is technically alive but operationally impaired. ADC07 calls out this problem directly in hybrid multicloud setups with mismatched redundancy and failover approaches.

AI piles on another layer: orchestration and agent behavior. When an agent experiences delay or failure, it often retries or attempts alternate paths. If your delivery policies differ by environment, those alternate paths might not behave the same way. That creates “soft failures” where the system returns partial results, times out mid-workflow, or gives different answers depending on which tool endpoint it hit.

Impact on scalability

Scaling in a multicloud world only works when policy is consistent enough that traffic can move freely to where capacity exists. When delivery policies are incompatible, scaling becomes constrained by where the “correct” policies live, not where the available resources live.

The result is some regions becoming overburdened while others remain underutilized. That is exactly what happens in AI systems when you scale infrastructure but don’t scale policy alignment. You end up with plenty of compute in one place, but the routing rules, rate limits, or security constraints prevent you from using it.

AI also breaks the mental model of “requests per second.” Your load is sessions, tokens, and multi-step chains. If one environment has stricter limits on payload sizes, concurrency, or timeout thresholds, you can accidentally create a scalability ceiling for the entire system. Worse, when some sessions succeed and others fail based on where they land, user retries amplify load and you get the classic death spiral where scaling pressure creates more scaling pressure.

Impact on operational efficiency

Incompatible delivery policies are how you end up with an ops team doing translation work instead of engineering work. Every environment boundary becomes a place where you have to reinterpret intent. Every cloud has its own knobs, its own defaults, and its own special way of breaking your day.

Different monitoring and reporting standards across providers already make it hard to get a holistic view and slows incident response. AI makes it worse because troubleshooting is end-to-end. If a user says, “the assistant is slow,” you have to figure out whether it was inference, retrieval, routing, tool calls, or policy enforcement.

And policy drift is not theoretical. AI delivery pipelines magnify this because people bolt on new components fast, and policy consistency tends to be “we’ll clean it up later,” which is famous last words in operations.

Best practices for mitigating incompatible delivery policies

The fix is boring, which is why it’s hard. You have to treat delivery policies as first-class artifacts, not tribal knowledge living in someone’s console clicks.

That means policy needs to be declarative, portable, and consistently enforced across environments. Routing rules, failover behavior, caching decisions, rate limits, and security controls should follow the application and be validated the same way you validate code. If you rely on manual translation between clouds, you are signing up for drift, and drift is the root cause of “it worked yesterday.”

You also need unified observability, because you cannot manage what you can’t see. Fragmented monitoring across clouds creates blind spots, and AI systems will absolutely exploit those blind spots. Consistent telemetry and consistent policy validation checks are how you stop troubleshooting from becoming archaeology.

Finally, in AI systems specifically, you want policy that understands modern traffic patterns. Agents and tool calls create behaviors that traditional delivery policies were never designed to anticipate. If you enforce policies as if every request is a neat, user-driven REST call, you’ll either block legitimate AI workflows or let dangerous ones slip through. Neither outcome is great for uptime or risk.

Conclusion

ADC07 is fundamentally about inconsistency, and inconsistency is the enemy of reliable delivery. AI raises the stakes because it multiplies interactions, chains dependencies, and makes users far less tolerant of unpredictability.

If you want AI systems that feel fast, stable, and trustworthy across hybrid and multicloud, delivery policies have to stop being environment-specific folklore. They need to be aligned, portable, and enforced consistently, or the whole stack will behave like a patchwork of “mostly works” that collapses under real-world load.

Read more about the Top 10 Application Delivery challenges faced by organizations across the globe.

About the Author

Lori Mac VittieDistinguished Engineer and Chief Evangelist | F5

Lori MacVittie is a Distinguished Engineer and Chief Evangelist in F5’s Office of the CTO with deep expertise in application delivery, automation strategy, and infrastructure. She is known for turning complexity into clarity whether she’s defining guardrails for AI agents, dissecting brittle multicloud architectures, or probing the limits of scalable systems. She brings more than thirty years of industry experience across application development, IT architecture, and network and systems operations. Before joining F5, she served as an award-winning technology editor. MacVittie holds an M.S. in Computer Science and is a prolific author whose publications span security, cloud, and enterprise architecture. She is also an avid tabletop and video gamer with unapologetically strong opinions about cheese.

More blogs by Lori Mac Vittie