In the past eighteen months, 95 AI tools have shut down and 101 have been acquired. The AI tool you learned last year may not exist next year. The prompting technique you mastered last quarter may already be counterproductive on the model that replaced the one you practiced on. And the platform you built your workflow around may, without warning, delete your entire account because of a UX decision no one tested.
All three of those happened to us at 24/7 Teach. They are not hypothetical. They are operational lessons from running AI across our organization since 2022, when we started with a copywriting tool called Jasper before most people had heard of ChatGPT. We have since moved through multiple generations of tools, models, and workflows, and the lesson that keeps reasserting itself is simple: the tool is the most fragile part of the system. The skill that survives is knowing how to learn the next one.
The wrapper layer expires first
Jasper raised $143 million and hit a $1.5 billion valuation in 2022. By 2024, its revenue had dropped from $120 million to $55 million, a 54 percent decline in a single year. Its CEO and CTO both stepped down. The product did not get worse. ChatGPT arrived and did the same work for free, and the value of a user interface wrapped around someone else's model collapsed overnight.
Jasper is not an outlier. Google's Darren Mowry said it publicly in February 2026: "If you are wrapping very thin intellectual property around someone else's model, the industry has run out of patience for you." The underlying economics explain why. The cost of GPT-4 tokens fell from roughly $30 per million in 2023 to $0.10 per million by 2025. A 99 percent cost reduction. When the raw capability is that cheap, the margin for a tool that adds a nicer interface on top of it disappears.
We started with Jasper. We do not use it anymore. Not because we made a bad choice in 2022, but because the layer we were operating at, a wrapper around a foundation model, was structurally the most fragile layer. Anyone who anchored their AI fluency to a specific wrapper tool in 2022 or 2023 has already had to start over at least once.
Prompting techniques are next
In 2023, "prompt engineer" was a real job title with a real salary. Fortune reported in 2025 that the role is now functionally obsolete. Microsoft's Jared Spataro put it bluntly: "Two years ago, everybody said, 'Oh, I think Prompt Engineer is going to be the hot job.' You don't have to have the perfect prompt anymore."
The macro story is clear, but the micro story is where it hits daily. We run into this constantly in our own work. A team member recently tested the exact same prompt on two different models. One proactively identified gaps she had only hinted at. The other missed them entirely. The prompt did not change. The model did. The technique that worked on Tuesday was insufficient on Thursday because the underlying model had shifted.
This is not a one-time adjustment. Andrew Ng demonstrated at Sequoia's AI Ascent that GPT-3.5 wrapped in an agentic workflow scored 95.1 percent on HumanEval, while GPT-4 with a single prompt scored 67 percent. How you use the model matters more than which model you use. But "how you use the model" is itself a moving target, because every new release changes what works.
The person who memorized prompting frameworks is constantly relearning. The person who understands why prompting works, what the model needs to produce a good result, adapts once and moves on. That is the difference between learning a technique and learning how to learn.
Token management is a real operational skill
Most organizations treat AI as a subscription. You pay, you use it, the cost is flat. That stops being true the moment you run AI at any real scale.
We learned this the hard way when our team hit 80 percent of our weekly token limit with three days left in the cycle. The cause: a new model and a new design tool were both drawing from the same token pool, and no one had accounted for the combined draw. The team had to downshift to a lighter model mid-sprint to finish the work. That is not a billing issue. That is an operational disruption caused by a resource nobody was managing.
The discipline we built afterward is model-to-task cost matching at the unit level. For our teacher-facing product, we use one model for analysis and a lighter model for conversational interactions, with prompt caching to cut repeated costs by 90 percent. The result is roughly $0.20 per interaction at scale. That number only exists because someone sat down and mapped which model handles which task at which cost, and then verified the math against real usage.
This is the part that no vendor demo covers. Token pricing changes with every model release. Usage patterns shift as features get added to platforms you already use. A cost model you built last quarter is wrong this quarter. The professional who treats AI cost management as a set-it-and-forget-it decision will get surprised. The one who treats it as a recurring discipline will not.
The tool stack is compressing
Two years ago, producing an interactive self-assessment for a blog post required an instructional designer working in Articulate, a visual designer in Figma, and a developer to wire it together. Last month, we produced one in a single working session using Claude. The tool stack that used to require three specialized applications and three distinct skill sets collapsed into a conversation.
This is not theoretical. Our team now prototypes dashboards directly in Claude Design and hands them to Claude Code for implementation. The step where someone opens Figma, builds a mockup, exports assets, and passes a spec to a developer has been skipped entirely. Not because Figma got worse. Because the workflow no longer needs it for the kind of design work we do most often.
The same pattern applies to Canva and Adobe Illustrator. Both are still useful tools. We still use them. But for quick visuals, social graphics, and one-off images, AI image generation handles what we used to open a design application for. The tools moved from the center of the workflow to the edge, reserved for the production-quality work that still needs them.
Articulate Rise and Storyline tell the same story from the instructional design side. For SCORM-compliant, LMS-integrated branching scenarios, Articulate remains the right tool. For simpler interactive content, the kind that makes up the majority of what most ID teams produce daily, the authoring step has collapsed into the AI conversation.
The deeper pattern is that we no longer select one model for a project. We map different models to different components of the same output: one for image generation, another for reasoning, and another for evaluation. The selection is per task component, not per project. That level of matching only makes sense if you understand what each model does well, what it costs, and where it falls short. Which brings us back to the skill that ties all of this together.
The four layers of AI fluency
Not all AI knowledge decays at the same rate. There are four layers, and most professionals are anchored to the one that expires fastest.
The wrapper layer is the most fragile. This is where you learned a specific tool: Jasper, Copy.ai, a particular AI writing assistant or design platform. When the tool changes, pivots, or disappears, the knowledge goes with it. Shelf life: months.
The prompting layer decays next. This is where you learned specific techniques: chain-of-thought prompting, few-shot examples, and formatting patterns. These shift with every major model release. A technique that improved output six months ago may now clutter the context window, making the response worse. Shelf life: one to two years.
The platform layer is more durable. This is where you understand how AI systems work: model selection per task, API integration, cost management, evaluation methodology, and system prompt architecture. This knowledge transfers across tools and survives version changes. When a new model drops, you know how to test it against what you already have. Shelf life: years, with periodic updates.
The meta-skill layer compounds. This is the ability to learn how to learn new tools quickly: evaluate, adopt, integrate, and move on. It does not decay because it is not anchored to any specific tool, technique, or platform. It is the skill of adaptation itself, and every cycle of learning a new tool strengthens it rather than replacing the last thing you learned.
Most professionals stop at the wrapper layer. The ones who build durable AI fluency move down the stack. Each layer deeper is harder to reach, but it is also harder to displace.
Note: This article was researched and written by Justice Jones with AI assistance, then reviewed and edited by our team. The studies cited belong to their original authors. The examples from our own work reflect our organizational practice.
What a platform failure actually teaches you
One of our AI platforms had UX so poor that an attempt to delete a single project deleted the entire account. Every integration, every API connection, every configuration, gone. We rebuilt from scratch, re-adding every variable manually.
No amount of research, no vendor evaluation, no demo would have predicted that. The lesson is not "choose better platforms." The lesson is that tool durability is a real variable in your cost model, and the only way to manage it is to keep your dependency shallow enough that you can rebuild when something breaks. That is a meta-skill decision, not a tool decision.
We also had temporary access to a new model for a limited window. The team evaluated it against our existing workflows, found it stronger in some areas and weaker in others, and made the call to use it selectively rather than switching wholesale. When the access window closed, we were back on our standard models with no disruption. That only works if the team's fluency is anchored to the process of evaluation, not to the specific model being evaluated.
"So you are saying learn to code?"
No. The meta-skill is not technical depth. It is the evaluation, adoption, and integration speed. It is knowing how to look at a new tool and ask: what does this actually do better than what I have? What does it cost per usable result? What breaks if it disappears? How long does it take my team to absorb it? And is the answer to all of those "do nothing and keep what we have"?
The World Economic Forum projects that 39 percent of existing workforce skill sets will either transform or become outmoded between 2025 and 2030. METR estimates AI capability is now doubling roughly every 4.3 months. The tools will keep accelerating. The only skill that compounds at the same rate is the ability to keep learning them.
Where to start this week
Audit your current tool stack for wrapper dependency. If the tool you rely on is a user interface on top of someone else's model, ask what happens when the model improves enough to make the interface unnecessary. If you do not have an answer, you are at the most fragile layer.
Learn one layer deeper than where you currently operate. If you are at the wrapper layer, learn how prompting works and why. If you are at the prompting layer, learn how to select models per task and manage costs. If you are at the platform layer, build the habit of evaluating every new tool against what you already have, including the verdict of "keep what we have."
Build the evaluation habit. Before adopting any new tool, write down what you use now, how long it takes, and how often it is right. Without a baseline, "this new tool is better" is a feeling, not a measurement. We wrote about the evaluation discipline in detail in The three roles that decide whether AI works in your organization.
If you work in instructional design or L&D, the shifts in this post are already in your workflow. Articulate, Figma, and the authoring stack are being compressed in real time. Our Instructional Design Bootcamp is built for exactly this moment.
If you work in any other field and want to build the meta-skill this post describes, that is what the AI Fluency for Professionals Bootcamp teaches: not a specific tool, but the ability to evaluate, adopt, and integrate whatever comes next.
Human-led, AI-facilitated. The tools will keep changing. The advantage is knowing how to find the right tool, test it, and make it work for your team and your budget.
About the author
Justice Jones is an instructional designer, AI strategist, and former K-12 principal. He co-founded 24/7 Teach to close the gap between what schools teach and what teens and professionals actually need to succeed, and he serves as CSO of Naomi-AI. 24/7 Teach has supported more than 50 organizations and placed 600 or more adults in new careers, and its teen graduates have collectively earned more than $5.7 million in scholarships.