April 16, 2026 · 6 min read
Infinitely Fast AI Still Won't Fix a Slow Workflow
A lot of AI discussion still focuses on the model.
How smart is it? How fast is it? How many tokens can it handle? How good is the reasoning?
But there is another problem starting to matter more:
what happens when the model gets faster than the tools around it?
That is the core idea behind a recent argument from Google Chief Scientist Jeff Dean. His point is simple: even if models get dramatically faster, the real-world gain can still be limited by the systems they depend on. If agents are forced to work through tools built for human-speed interaction, then the toolchain becomes the bottleneck.
Businesses also need to become agent-ready
This has a direct implication for service providers.
A website built for humans is often not enough for an AI agent. Important information may be spread across multiple pages, expressed inconsistently, or left unstated. Service scope, pricing logic, response times, constraints, and booking paths are often hard for an agent to interpret quickly and reliably.
That matters because agents do not browse like humans. If they cannot understand what a provider offers, whether the provider is a fit, and how the transaction can move forward, they are likely to move on to another option.
That means businesses risk losing demand not because they are worse, but because they are harder for agents to use.
The new bottleneck is not always the model
Dean's example is especially clear in coding.
An AI agent may generate the next action almost instantly, but it still has to interact with compilers, files, editors, APIs, and other tools that were never designed for machine-speed execution. Things humans barely notice become major bottlenecks when an agent operates far faster than a person.
This is basically Amdahl's Law applied to AI workflows: if a meaningful part of the task remains slow, speeding up one part of the system has limited effect on the whole.
This goes beyond coding
The same issue extends well beyond software development.
The broader point is that AI agents are increasingly entering workflows that were designed for humans: documents, spreadsheets, business systems, service booking, and operational coordination. As that happens, the surrounding systems matter as much as the model itself.
The next gains may not come only from better reasoning models. They may come from rebuilding the operational surfaces those models have to use.
What this means for Aune
This matters a lot for Aune.
In real-world services, the hard part is not only understanding user intent. It is moving from intent to execution through systems that are often fragmented, manual, and still built around phone calls, inboxes, and human-paced coordination.
A faster model on its own does not solve that.
If the surrounding workflow is still slow, unclear, or unstructured, then better model performance gets trapped inside a weak operating environment. The bottleneck shifts from intelligence to execution.
That is why Aune's opportunity is not just in better interpretation of demand. It is in helping create a more agent-ready transaction layer for services:
- clearer structured intake
- better scope handling
- stronger provider matching
- cleaner negotiation paths
- more executable booking workflows
The more AI moves toward real-world action, the more the surrounding infrastructure matters. In services, that may be where a large share of the value ultimately sits.
Sources
This post is based on Jeff Dean's discussion on AI agent tooling bottlenecks.
- Video: https://youtu.be/XlfumXPPrLY