Plateaus & Playbooks: Opportunities in an AI Slowdown
By Bryce Murray, Director of Technology
An applied scientist with a PhD in machine learning and AI, Bryce leads AI and Data Science initiatives at Permanent Equity. His focus is on developing practical solutions using a combination of cutting-edge technologies, historical internal data, external resources, and original development.
A Plateau in AI?
Recent releases from OpenAI, Anthropic, and Google suggest that large language models (LLMs) are no longer improving at the rapid pace we’ve come to expect. Put simply, we may be hitting a performance plateau: each new version delivers only small gains and sometimes even setbacks despite ever-growing compute power.
Think of it like squeezing an orange. The first press gives you a large amount of juice. But each squeeze after that yields less and less, until you’re straining for just a few drops. That’s roughly where today’s leading AI models are. The easy gains are gone, and each new improvement takes far more effort for far less payoff.
If the engine under the hood is stalling, the real innovation will shift to how businesses deploy AI, including integrations, workflow design, and proprietary data, rather than raw model horsepower.
Why Are Model Gains Slowing?
One possible reason performance appears to be slowing is that we’re running out of fresh, high-quality training data. Think of it like this: the models have read every book in your local library, and now we’re asking them to improve by rereading the same shelves.
What changed?
We now see unprecedented levels of compute power and capital being poured into training the top models, with seemingly unlimited resources and talent focused on the task.
Model sizes keep growing. Today’s frontier systems are many times larger than the previous generation, yet in theory there may soon be more model complexity than there are real patterns in the data to learn from.
Most internet text worth learning from is likely already in the training mix.
Legal barriers are tightening around copyrighted material and private user data.
Synthetic data (models training on their own output) helps at the margins but doesn’t provide new patterns to learn.
What Isn’t Plateauing?
It’s worth noting that while model horsepower may be leveling off,product innovation is still accelerating. For leaders, the distinction is useful context because it helps frame where to invest attention.
Plateauing: Base accuracy of general-purpose LLMs, larger parameter counts, and one-size-fits-all chatbots.
Still Advancing: Domain-specific tools, integrations with existing software, application of retrieval augmented generation (RAG), fine tuned models for niche vocabularies, and employees upskilling to work effectively with these AI copilots in their daily tasks.
The takeaway: Don’t chase the latest model release. Instead, focus on tools that plug into your processes and vendors who can show you how to leverage your data with new AI technologies to maximize ROI.
Implication #1 – Specialization of Tools
With model performance stalling, AI companies will have to shift their approach to continue adding value. This may mean they begin solving end-to-end problems for defined niches, such as software development, legal services, accounting, or healthcare.
Early signs already point this way: niche startups are rushing in, and larger platforms are beginning to fold specialized functions into their broader offerings. As this happens, internal tool chains may start to consolidate. Rather than stitching together dozens of single purpose apps, many companies will shift toward fewer but deeper tools that can handle entire workflows.
The bottom line: For SMBs and beyond, this means evaluating not just whether a tool works today, but whether it can expand with you as AI products specialize and mature.
Implication #2 – Software Development & Internal Tooling
Software companies are going to have to adapt, but they won’t become obsolete. AI coding tools, while good, still struggle.
The companies that win will be the ones working with developers who become masters of these tools. Tools shift. And, importantly, this shift may also mean companies no longer have to settle for off-the-shelf CRMs, ERPs, or other generic platforms.
The bottom line: With AI accelerating development, expect lower costs for and time-to-execute on systems tailored to the business process, instead of forcing the business to conform to the software.
Implication #3 – Foundations & Processes
I’ve had a concern that establishing processes, software, or setting a foundation on top of AI too soon could be pointless. The uncertainty around what might be coming down the pipeline made it feel overwhelming to figure out where to invest time, sparking worries that the effort might be wasted if a new model might replace and outpace our gains.
But, if we are right about the plateau, the timing changes. With fewer game-changing features likely to appear, the ground is steadier to begin laying durable tracks for how your organization will use AI.
The bottom line: It’s not about rushing to do everything, but recognizing that the window is open to start codifying systems and practices without fear that tomorrow’s release will make them irrelevant.
Looking Ahead
The plateau in raw model horsepower is not the end of the story. It signals that the playing field is opening for smaller operators. Here are the shifts most likely to matter:
First movers will win. Teams that adopt AI fastest will start to outpace their competition. The real advantage will come from people who get creative, learn where AI is weak, and design workarounds that transform how their teams operate.
Prompts as playbooks. Prompts will evolve into a shared language across companies. They will not only guide AI tools but also capture and communicate process knowledge in a way that can be reused immediately and at scale.
Falling barriers to custom software. The cost and complexity of automation and custom systems will continue to shrink. With modern tools and AI in the mix, businesses may soon find that building software tailored to their process is as affordable as buying off the shelf.
The takeaway: The opportunity is no longer about chasing model performance. It’s about building momentum in your business by putting these tools to work.
Getting Started: A 30-Day Plan
Rather than a long rollout, think of this as a 30-day starter plan to build momentum. The goal is to make AI real for your team right away, not to design a perfect program.
Step 1: Log in. Have your team create an account with ChatGPT, Claude, or Gemini. Make sure everyone can access at least one of these tools.
Step 2: Try the prompt. Run this exercise together (if you’re using ChatGPT, we recommend toggling to the o3 model from 5-pro, which is the current default):
As an expert in growth for [INSERT COMPANY NAME], your job is to help me identify three new ways to find customers. Our ideal customer is [DESCRIBE IDEAL CUSTOMER].
We will go through the following process:
Brainstorm three new ideas for finding customers.
Evaluate the pros and cons of each idea.
Develop an action plan for enacting the best idea.
Step 3: Share results. Ask each person to share what they learned from the exercise. Compare notes and highlight at least one idea worth testing.
Step 4: Pick a pilot. Choose a single idea to try in the real world. Keep it small and time bound, but commit to measuring the outcome. Examples include:
Growth: Use the prompt above to brainstorm new ways of finding customers.
Operational efficiency: Use a prompt to plan your next team meeting agenda.
Researching new markets: Ask AI to map three adjacent industries and summarize potential entry points.
In 30 days, you will have moved from curiosity to hands on practice, with a tangible pilot underway.
The opinions and analyses in this article are those of the author and do not necessarily represent the opinions or positions of Permanent Equity Management, LLC, its officers, employees, affiliates, or portfolio companies. This content is for informational and educational purposes only and should not be construed as investment, legal, tax, or other professional advice. Nothing in this article is an offer to sell or a solicitation of an offer to purchase any security or interest, nor does it form the basis of a contract or create any fiduciary or advisory relationship with Permanent Equity Management, LLC. Any examples, references to third-party services (including AI platforms such as ChatGPT, Claude, or Gemini), or links to third-party content are for illustrative purposes only and do not imply endorsement. Do not input confidential, proprietary, or personal data into third-party AI services; Permanent Equity Management, LLC is not responsible for the availability, accuracy, or content of third-party services. You should obtain independent advice suited to your specific circumstances before taking action. Past performance is not indicative of future results.
