Four misconceptions execs have about AI

Kunal AnandChief Product Officer | F5

Our 2026 State of Application Strategy Report found that 78% of organizations now run AI inference themselves, across an average of seven models. A year earlier, our AI Readiness Index put the share of organizations genuinely ready to operate AI at scale at 2%.

The gap between those two numbers is the most interesting thing in enterprise technology right now. Enterprises didn't wait for readiness. They brought inference in-house because they wanted control of their data, costs, and security. Now they're discovering what that control requires. Call it the inference path: everything that happens between a user's intent and a model's output. Retrieval, routing, tool calls, identity, context assembly, policy enforcement, the return trip. The model is a component on that path. The path itself is the system, and it is the part you own.

Four misconceptions I run into repeatedly are all variations of the same mistake: treating the model as the product and the path as plumbing.

Misconception #1: "It's all about the model."

Which model, which vendor, and which use case should I lead with? This is the first question in nearly every executive conversation, and it is the least durable one.

Models are the swappable component. Only 8% of organizations rely exclusively on public AI services. The rest are running diversified portfolios precisely because they expect to swap, route, and fall back based on cost, accuracy, and security. If your architecture makes switching a model a project rather than a configuration change, you have already made the wrong bet.

What does not swap is the inference path. The retrieval layer that decides what context the model sees. The routing logic that picks what model handles which request. The identity that travels with the call. The policy that decides what the model can do with the answer. That is where accuracy comes from, where cost is determined, and where your proprietary advantage lives. A competitor can buy the same model tomorrow. They cannot buy your path.

Misconception #2: "If the pilot worked, production will be easy."

The failure mode is not that production is harder. It is that production is complex, and length compounds. The vast majority of enterprises (93%) operate across multiple clouds, according to F5 research, while 86% distribute applications across hybrid multicloud environments. Your inference path runs through all of it, and that complexity keeps growing. As organizations add more models, applications, and agents, the number of possible inference paths multiplies.

Our research also found that 52% of organizations already chain or orchestrate multiple models. Take a chain and assume each step is 95% reliable, which is generous for anything involving retrieval, tool calls, and a probabilistic model. A three-step pilot chain succeeds 86% of the time. A 20-step production chain succeeds 36% of the time. The per-step number never moved. The system went from mostly fine to mostly broken because you added steps.

Now apply the same arithmetic to security. If each agent in a chain has a 5% chance of being manipulated, mis-scoped, or handed a credential it should not have, a 20-agent workflow gives an attacker a very strong chance that at least one link is exploitable. This is why more than nine out of 10 respondents report new categories of problems once agentic AI hits production levels: credential handling, action auditing, and the basic question of what an agent did and on whose authority.

Misconception #3: "Performance, security, and governance are separate priorities."

In traditional application architecture, this separation was defensible. Security ran as a parallel workstream. Governance was quarterly. Performance belonged to a different team with a different budget.

AI collapses all three into one budget, measured in milliseconds and tokens. Every guardrail you add to the inference path costs latency. Every governance requirement costs context window, which costs money on every single call. Prompt inspection, output filtering, PII redaction, model routing, retry logic: each is simultaneously a security control, a performance cost, and a line item. You can’t fund one without spending from the others, and you can't cut one without exposing the rest.

This is genuinely new. Security overhead used to be amortized invisibly. On the inference path, it shows up in the p95 and on the bill, which means the tradeoff is now explicit, and someone must own it. Today, the most common control in place is authentication for AI services and API access, used by 55% of organizations, and 47% plan to move toward identity-aware infrastructure for agents. Both are the right instinct. Both also add latency to the path. Budget for them deliberately, or your teams will quietly drop them to hit a performance target.

Misconception #4: "Point solutions add complexity, but we can make it work."

Take the median enterprise in our data that has seven models, operates in multiple clouds, and has applications spread across on-premises, public cloud, and colocation environments. Add a separate tool for API security, another for AI gateway functions, another for output governance. Every policy you write now must be expressed correctly in every dialect, across every environment, for every model, and re-verified every time any one of them changes. The number of places a policy can silently drift is the product of those changes.

When something goes wrong at 2 am, the question "which control was supposed to catch this" has no single owner. The alternative is not fewer capabilities. It is one enforcement point in the path, regardless of which model, environment, or protocol is on the other side.

Policy consistency and model freedom are not in tension. They are only in tension when policy lives inside the model integration instead of in front of it. Converging application delivery and security onto a single platform is what makes it possible to change models weekly and still answer an auditor's question in one place.

The questions worth asking now

Here are four questions for your leadership team as your organization works to manage the inference path:

1. How many models are in production today, and how long does it take to add or replace one?
2. What is the longest agent chain running in production, and what is its measured end-to-end success rate?
3. Where is AI policy enforced, and how many distinct places would you need to change to update it?
4. What percentage of your p95 inference latency is security and governance overhead, and who owns that number?

If those questions produce estimates instead of answers, that is useful information, and far better learned now than during an incident. My suggestion for everyone is to audit the inference path this quarter. The organizations that lead in this era will not be the ones running the most AI. They will be the ones that can say exactly what happens between prompt and response, everywhere it runs.

To go deeper on the data, download the 2026 State of Application Strategy Report. To see how F5 is extending its Application Delivery and Security Platform strategy to enterprise AI, visit the F5 AI Security Platform webpage. Also, be sure to check out today’s press release.

About the Author

Kunal AnandChief Product Officer | F5

Kunal Anand leads the F5 product organization as Chief Product Officer. Responsible for product vision, strategy, and execution, he ensures development of breakthrough solutions that solve critical challenges and create exceptional experiences for customers. In his previous role as Chief Technology and AI Officer, Kunal charted the company’s technology and AI strategy and vision. Prior to F5, Kunal held the dual role of Chief Technology Officer and Chief Information Security Officer at Imperva. His journey to Imperva began in 2018 with the acquisition of Prevoty, an application security startup he co-founded in 2013. Before joining Prevoty, he was the Director of Technology at BBC Worldwide. Kunal has a deep history of innovation and technical expertise, and has held roles leading security, data, technology, and engineering teams at Gravity, MySpace, and the NASA Jet Propulsion Lab. Kunal has over 15 years of experience in AI and machine learning, ranging from model training, employing AI-driven algorithms to enhance products, and designing and implementing AI architectures. Kunal holds a Bachelor of Science degree in computer science from Babson College.

More blogs by Kunal Anand