The AI Tool Map Looks Nothing Like It Did a Year Ago
When the reference changes, your stack should too
Some guides age. Others act as thermometers. Ethan Mollick’s guide belongs firmly in the second category.
Simon Willison, one of the most rigorous observers of the AI ecosystem, noted on his blog on July 27th something worth paying attention to: Mollick’s opinionated guide to which AI to use for what has changed substantially compared to the version from a year ago. What was essentially a chat comparison — ChatGPT, Claude, Gemini — with some advanced reasoning models as an alternative, is today an entirely different map. More complex, more specialized, and more useful if you know how to read it.
For a company already using AI in its operations, or in the process of scaling it, that change is not a footnote. It’s a direct signal that the ground has shifted.
What has changed in the AI tool map over the past year?
A year ago you could simply pick one of the three major chat models and maybe add a synthesis tool; today the ecosystem has specialized by task type and mode of use. It’s no longer about which model has the best general benchmark, but about which tool, in which mode, for which specific output.
A year ago, the conversation about AI tools in business settings was easy to simplify: you picked one of the three major chat models, maybe added Deep Research for information synthesis tasks, and called it done. The friction was mostly around adoption, not selection.
What Mollick documents now — and what Willison highlights as significant — is that simplicity is gone. The ecosystem has evolved toward a logic of specialization by task type and mode of use. It’s no longer just about which model has the best general benchmark; it’s about which tool, in which mode, for which specific output.
This has direct implications for any organization that built its workflows around decisions made twelve months ago.
What’s the risk of not updating your team’s AI stack?
A stack that isn’t reviewed accumulates two silent debts: capability debt — you’re using a model or mode that’s no longer the most precise, fast, or cost-effective for your case — and design debt — your workflows don’t take advantage of what current models can do. The opportunity cost is real even if it’s invisible day to day.
Many companies that jumped on the AI bandwagon early made an understandable mistake: they chose a tool, integrated it into their processes, and left it there. That worked reasonably well when the pace of change was slower. It doesn’t anymore.
An AI stack that isn’t reviewed regularly accumulates two kinds of debt:
Capability debt. You’re using a model or mode of use that’s no longer the best fit for your situation. Not because it’s bad, but because more precise, faster, or more cost-effective alternatives have emerged for exactly what you do. The opportunity cost is real, even if it’s invisible day to day.
Design debt. The workflows you built around last year’s capabilities don’t take advantage of what current models can do. It’s like designing an assembly line for 1990s tools and never touching it since.
The point isn’t to change for change’s sake. It’s to review with intention.
Reading Mollick’s guide well: what to look for beyond the specific recommendations
Mollick’s guide has value as a reference map, but its real usefulness for a business isn’t in following its recommendations to the letter. It’s in the thinking framework it proposes.
What has changed in his approach — and this is what’s worth extracting — is that he no longer organizes tools by brand or model, but by task type and mode of use. That’s exactly what any operations or digital transformation leader should be doing when evaluating their stack.
The questions that matter are:
- What kind of reasoning does this task require? Is it information synthesis, structured content generation, data analysis, decision-making under ambiguity?
- Is the output for direct human consumption, or is it an intermediate step in an automated workflow?
- Do I need integration with internal systems, or is this task self-contained enough for an external tool?
- What level of consistency and auditability do I need over the results?
When you structure your analysis this way, the specific recommendations in any guide — including Mollick’s — become useful inputs rather than definitive answers.
What this means if you’re scaling with agents
If your company is already running AI agents — or evaluating doing so at a departmental scale — the speed of ecosystem change introduces a management dimension that didn’t exist before: stack governance.
It’s not just about making the right choice the first time. It’s about having a process to review those choices systematically. When to review, who reviews, by what criteria, and how transitions are managed when you swap out a component without breaking the workflows that depend on it.
This is especially critical when agents aren’t support tools but real operational pieces: agents that qualify leads, manage customer communications, process documents, or make decisions within approval workflows. In those cases, an outdated or poorly chosen model isn’t a marginal productivity issue — it’s a service quality or operational efficiency problem with a direct impact on results.
The good news is that well-designed agent architecture allows components to be replaced without rebuilding from scratch. But that requires having thought it through from the start.
How do you choose your AI stack, and how often should you review it?
With three moves: audit the tools you’re using today against the current state of the ecosystem, map your tasks by output type, and set a review cadence. For most organizations, a quarterly review is reasonable.
If Mollick’s guide is useful for anything, let it be as a lever to do the following:
First, audit the tools you’re using today. Not to replace them by default, but to benchmark them against the current state of the ecosystem. Are they still the best option for what you’re doing with them? Have better alternatives emerged for any of your key tasks?
Second, map your AI tasks by output type. Separate those that generate content from those that analyze data, from those that reason about complex situations, from those that act within automated workflows. That map will give you a much clearer picture of whether you’re using the right tools — or solving everything with the same hammer.
Third, establish a review cadence. It doesn’t need to be monthly. But in an ecosystem where the reference guide changes substantially in twelve months, going two years without a review means taking on silent risk. Quarterly is reasonable for most organizations.
The map keeps changing: what matters is knowing how to navigate
The most valuable thing about Ethan Mollick updating his guide — and Simon Willison flagging it as worth attention — isn’t the specific recommendations. It’s the demonstration that the AI ecosystem is still in active motion, and that decisions you made a year ago deserve to be revisited with fresh eyes.
For a company that wants to use AI as a real lever for operational scalability, the critical skill isn’t having made a good choice once. It’s maintaining the ability to keep making good choices, continuously, with a clear process — and without relying on someone on the team constantly tracking every industry development.
The map has changed. The question is whether your organization has the mechanisms to detect that and adapt, or whether it will keep navigating with last year’s version.
Sources
- Mollick, E. – «An opinionated guide to which AI to use to do stuff», One Useful Thing
- Willison, S. – commentary on Mollick’s guide, Simon Willison’s Weblog, Jul. 27, 2026
- Mollick, E. – «Using AI right now: a quick guide» (previous edition), One Useful Thing
Frequently asked questions
Why does it matter that Mollick’s guide has changed so much in a year?
Because it reflects that the AI ecosystem can no longer be summed up as choosing between three chat tools: today there are specialized modes, models, and tools that change what’s possible and at what cost.
How does a company decide which AI tool to use for each process?
The key is to start from the output you need, not the model’s name: define the task type, the level of reasoning required, and whether it needs integration with your systems before selecting a tool.
Does it make sense to standardize on a single AI tool for the entire company?
Less and less so: the best results come from matching the right tool to each type of task — just as you wouldn’t use the same software for accounting and for design.
What’s the risk of not updating the AI stack my team is using?
Working with outdated models or modes means leaving available capabilities on the table, which translates directly into lower output quality and more time spent per task.
Turn your content into customers, without it depending on your time
At Yuniax we build the system that attracts, qualifies and nurtures your customers automatically: content, funnels and automation working together so your business grows without you being in every step. If you want to see how to apply it to yours, book a call with our team and we will show you where to start.