Your AI Strategy Is a Wrapper. Here's What a Moat Looks Like When Intelligence Gets 13 Cheaper Every Year.
Most AI strategies I come across share an architecture: a prompt, a frontier model behind an API, a decent UI, and a slide with the word proprietary on it. I'd call that a reseller agreement with extra steps. I'm not k
Most AI strategies I come across share an architecture: a prompt, a frontier model behind an API, a decent UI, and a slide with the word proprietary on it.
I'd call that a reseller agreement with extra steps.
I'm not knocking wrappers as such. Plenty of good businesses started as one. My worry is narrower. Wrappers are built on the part of the stack that's losing value fastest, and I don't think most founders have looked at how fast.
Nothing in history has deflated like this
Epoch AI tracked what it costs to reach a fixed level of AI performance over time. Since 2023, that cost has dropped about 47% per quarter, which is roughly 13ร a year. By their comparison that's faster than DNA sequencing, compute, lithium batteries and electricity ever got cheaper (Epoch AI, "The plunging price of thought"). An earlier Epoch analysis put the yearly decline anywhere from 9ร to 900ร, depending on the task (Epoch AI).
Epoch is upfront about the caveats. Benchmarks aren't real work, and real users don't hop to the cheapest model every quarter. Still, nobody serious disputes the direction.
So think it through. If the core capability inside your product gets an order of magnitude cheaper every year, any cost edge you have over a competitor lasts about a quarter. Any capability edge lasts until the next model release, because they rent the same model you do. And your prompt engineering IP is a few kilobytes of text that a good engineer can approximate in an afternoon.
That leads to a test I would run on any AI product. If some part of your product gets better on its own when the foundation model improves or gets cheaper, that part isn't a moat. Every competitor shares that tailwind. Strip those parts out, and whatever's left is your actual strategy. Sometimes that's very little.
The MIT report's real finding
In 2025, MIT's NANDA initiative published The GenAI Divide, a report that got quoted everywhere for its headline claim that 95% of organizations were getting zero return from generative AI (MIT NANDA, State of AI in Business 2025). That number has been argued over, and I'd read the methodology before repeating it.
The diagnosis is the bit I'd keep:
"The core barrier to scaling is not infrastructure, regulation, or talent. It is learning. Most GenAI systems do not retain feedback, adapt to context, or improve over time."
That's my test seen from the other side. A wrapper is a system that doesn't learn. On day 400 it's exactly as good as it was on day 1, except that the model underneath is now a commodity everyone has. So when people ask me what makes an AI product defensible, my answer is the slope of its improvement curve. Where it started barely matters.
Evals are the asset nobody budgets for
If I could make founders invest in one thing, it would be their own evaluation suite: the tasks, edge cases, grading rubrics and failure categories specific to their domain. It's wildly underrated.
Here's why I think it compounds. With good evals you can swap models in days instead of quarters. When a cheaper model ships, you know within hours whether it's good enough for your workload, while a competitor without evals is either stuck paying for the expensive model or shipping regressions blind. That turns the 13ร yearly price drop from a threat into margin, because you can actually capture the savings while they hand theirs to customers out of caution. The suite also holds domain judgment that took years of customer contact to learn, and nobody can download that.
No evals, no AI product. What you have is a demo with uptime.
Text is everywhere, outcomes are rare
Everyone has text. Very few companies have labeled outcomes. Did the suggestion get accepted? Did the claim get paid, did the ticket stay closed, did the code pass review, did the customer leave anyway?
Outcome data is the one signal that's proprietary and also causally tied to value. So I'd design the product so every interaction closes a loop: the agent acts, the world responds, and you capture that response attached to the action that caused it. That's precisely the "retain feedback" capability the MIT report found missing.
In practice this means treating your event log as part of the product, with action, context and outcome linked and queryable. Most teams log prompts and completions. Almost none log what happened next, and I think that's the most expensive omission in AI product design right now.
Owning the write path
Models generate. Businesses act. The position I'd want is ownership of the system of record and the permission to change it, whether that's the ERP write path, the claims adjudication step, the deployment pipeline or the procurement approval.
A model that drafts a purchase order is a commodity. A system that sits inside the approval chain, holds delegated authority and is trusted to execute is hard to rip out, because whoever replaces it has to rebuild the integrations, the permissions and the organization's trust from scratch. Distribution plus authority beats intelligence, since intelligence is the bit you rent.
Trust is turning into a product
As agents start taking real actions, I expect buyers to pay for provable control: audit trails, enforced policy, a human in the loop at the right moments, and evidence that holds up in front of a regulator.
That stopped being a pure cost center this year. In the EU, the AI Act's transparency duties under Article 50 have applied since 2 August 2026, and after the Digital Omnibus amendment the high-risk obligations land on 2 December 2027 for Annex III systems and 2 August 2028 for AI embedded in regulated products (Regulation (EU) 2026/1744). In regulated industries, the vendor whose product already produces that evidence often wins the procurement fight before it starts. Compliance built into the architecture, rather than written up in a PDF, is a moat that few founders take seriously. I suspect that's exactly why it works.
Cost structure, but only with the rest
Then there's your inference cost structure: routing by difficulty, caching, distilling your common cases onto smaller models, and deciding feature by feature what runs in batch and what runs in real time.
On its own, anyone can copy this. Paired with evals and outcome data it compounds, since the evals tell you which traffic a small model can safely take and the outcome data gives you what you need to train a specialist for your domain. Picture a team that runs most of its traffic on a cheap specialist model with measured quality, next to one that sends everything to a frontier API. The first team has structurally better margins, and the gap widens every time prices fall.
Lines I would retire
"Our prompts are our IP." Prompts leak, get reverse-engineered, and get made obsolete by the next model. They're configuration.
"We fine-tuned a model." On what, though? If it's public data, a competitor can do the same next month for less. Fine-tuning is only as defensible as the data behind it.
"We're model-agnostic." Good, that's hygiene. But being model-agnostic without evals just means you can switch models without knowing whether you should.
"We'll add the feedback loop later." I'd argue the loop is the product. Data you didn't capture in year one is data you'll never get back.
"Our UX is the moat." Great UX matters and it's copyable. It buys months, maybe not years.
If I were reviewing your product
I'd go through your components and ask which ones get better when the model improves, and then ask what's left. I'd ask how long it would take to move safely to a model that's 10ร cheaper, and whether the answer is hours, weeks or a shrug. For every AI action, I'd want to know whether you capture the real-world outcome and link it back. I'd ask whether your product actually executes inside a customer's system of record or hands them text to paste. Could you show a regulator, or a Fortune 500 security team, the evidence for why your system took a given action? And how much of your traffic truly needs a frontier model?
Mostly "no" and "we don't know" is fine for a prototype. It just isn't a company yet.
When intelligence gets an order of magnitude cheaper every year, the model is the one thing you can be sure every competitor will have. I'd stop building on the part that's deflating and put the effort into the parts that gain value with use: your evals, your outcome data, your authority to act and your ability to prove you can be trusted.
If you're working out what to build versus rent in your AI stack, or you need an eval and feedback-loop plan, I'm happy to think it through with you. If you think my test is wrong, argue with it below.
Michael, ArkNet Digital (https://www.arknet.digital)
Originally published by Dev.to AI. Aggregated on AIWithGhost for educational purposes โ full credit and traffic to the original publisher.