You have heard the phrase from every provider you have talked to.
“Our AI is built on best practices.”
It sounds rigorous. It sounds like someone did the hard work. It sounds like the provider surveyed the landscape, identified what works, and built the foundation on the highest standard available.
Here is what it actually means: nobody filtered anything.
“Best practices” is the average of what everyone does. It is the collection of every technique, framework, opinion, and approach that gained enough consensus to seem credible, packaged together without anyone asking the one question that matters. Does it actually work?
That is not the highest standard. It is the lowest. And when an AI system is built on it, the output tells you exactly what the foundation is made of. Competent. Interchangeable. Generic. The same result you would get from the next provider, because the next provider built on the same thing.
The question this article answers is specific. When a provider claims their AI is built on proven methodology, what does that mean? And when they claim it is built on best practices, what does that mean? The distinction is not semantic. It is the distinction between two completely different foundations, and it is the single best predictor of whether the output will sound like your business or sound like the internet.
Why the Foundation Matters More Than the Technology
Other articles in this series cover how to evaluate a provider’s discovery process, how to prepare your own business for AI implementation, how to tell calibration from configuration, and how to vet providers overall. This article covers the layer underneath all of that: the raw material the AI is built from.
There are two ways an AI foundation fails. The first is the empty foundation: generic AI with nothing real underneath it, producing the internet’s statistical average and calling it done. The second is the blended foundation: AI stuffed with every methodology the provider could find, proven and unproven at equal weight, producing output that sounds sophisticated and means nothing. Both produce generic results. Both charge custom prices. And both hide behind the same phrase: “built on best practices.”
This article is about the blended foundation specifically, because it is harder to detect. An empty foundation is easy to spot. The output is obviously flat. A blended foundation is insidious. The output sounds credible. It cites frameworks. It uses the right terminology. But nothing underneath was filtered, which means nothing underneath is reliable. The volume creates the illusion of depth. It is the more expensive mistake.
The technology is mostly the same. The major AI models are available to everyone. Any competent provider can access GPT-4, Claude, or Gemini. The differentiator is not the model. It is what the model is built on.
Think of it this way. Two restaurants can buy the same commercial oven. One fills it with pre-made frozen meals. The other uses ingredients sourced from a chef who spent twenty years learning which producers, which cuts, and which techniques actually produce results. The oven is identical. The output is not. And nobody eating the food would confuse the two.
Two providers can use the same AI model, the same API, the same deployment infrastructure. If one builds the foundation on curated, proven methodology and the other builds it on “best practices,” the output will be structurally different. Not cosmetically different. Structurally.
The foundation determines the output. That is an engineering reality, not a sales talking point.
What “Best Practices” Actually Tells You
Michael Porter published a landmark article in the Harvard Business Review in 1996. The core argument was a warning about the exact phenomenon the AI implementation market is reproducing today.
Porter observed that as companies benchmark against each other, they converge. They adopt the same processes, outsource to the same vendors, implement the same tools. The result is a race down identical paths that no one can win. He called it competitive convergence.
The phrase “best practices” is the engine of that convergence.
When a provider says their AI is built on best practices, they are telling you the system was built on whatever gained consensus across the market. Not what was tested against real outcomes. Not what anyone filtered for effectiveness. What gained consensus. In every field, consensus and effectiveness are different things. They overlap sometimes. They diverge more often than anyone in the room wants to admit.
The competitive strategy angle is plain. If you implement what competitors are doing, assuming you implement it correctly, you become indistinguishable from your competitors. The point of strategy is to be different. The point of best practices is to be the same.
W. Edwards Deming, the architect of modern quality management, warned against the practice decades before AI existed. He taught that copying without understanding is dangerous, that “best practice” is a static idea that leads organizations to imitate rather than improve. Taiichi Ohno, the engineer behind the Toyota Production System, pushed in the same direction: everything you need to improve your organization is already within it. Looking outside at what everyone else does is looking in the wrong place.
The logical conclusion is uncomfortable but precise: benchmarking best practice is the fastest way to mediocrity. That is what happens when an entire market builds on the same undifferentiated base.
Apply this to AI. When every provider says “built on best practices,” they are telling you the foundation is undifferentiated. Not individually. Collectively. If Provider A, Provider B, and Provider C all built on the same consensus-based body of information, the output from all three will be structurally similar. The tone might shift. The formatting will vary. But the substance underneath will converge, because the foundation underneath already converged.
You have already experienced this. Every proposal you received from AI providers sounded similar. Every demo produced competent but interchangeable output. That was not a coincidence. That was the foundation expressing itself.
Picture the evaluation. Three providers. Three demos. Each one ran your content through their system. The output from Provider A was clean and professional. The output from Provider B was clean and professional. The output from Provider C was clean and professional. You sat in the conference room afterward and could not articulate what made any of them different from each other, or from what you could get by typing the same prompt into ChatGPT. The formatting varied. The headers changed. The substance did not.
That is what a best-practices foundation produces. Not bad output. Indistinguishable output. And indistinguishable is the most expensive version of generic, because you are about to pay a premium for it.
What “Proven Methodology” Actually Means
Proven methodology is not more information. It is filtered information.
The evidence-based management movement drew a hard line between two approaches that most people assume are the same. Consensus-based decisions ask: what does everyone do? Evidence-based decisions ask: what does the evidence show actually works?
These are not the same question. They produce different answers.
David Sackett, the physician most associated with evidence-based medicine, defined the approach as “the conscientious, explicit, and judicious use of current best evidence in making decisions.” Three words carry the entire definition: conscientious, explicit, and judicious. Conscientious means deliberate, not passive. Explicit means documented and traceable, not assumed. Judicious means filtered through judgment, not included by default.
That is what proven methodology looks like when it is real.
A real foundation asks five questions about every framework before it gets included. What is this grounded in? Is the research authoritative? Has it been tested against real outcomes, not just theorized? Does it conflict with anything already in the system? And does it survive when the audience is someone who will push back?
A blended foundation asks none of those questions. It collects. Every framework, every opinion, every approach goes in at equal weight. Proven and theoretical. Tested and speculative. Authoritative and amateur. All mixed together. The output inherits every weakness because nobody filtered for strength.
Alvan Feinstein, a pioneer of clinical epidemiology, warned about exactly this pattern. The authority given to collected evidence, he argued, leads to inappropriate guidelines and rigid dogma. Not because the collection contains no good material. Because the collection treats good material and bad material as equals. The volume creates the illusion of rigor. But rigor requires exclusion, not inclusion.
Here is the distinction that matters. “Best practices” is a blended foundation. Everything in, nothing filtered. “Proven methodology” is a curated foundation. Specific frameworks, grounded in authoritative research, tested against real outcomes, structured for repeatable results, and selected so that only what works makes it in.
The difference between the two is not quantity. It is curation.
Why More Sources Do Not Make AI Smarter
There is a persistent assumption in the AI implementation market that breadth is a strength. “Our system draws from thousands of sources across dozens of industries.” It sounds impressive. The research says otherwise.
A 2024 study from the University of Massachusetts Amherst investigated the relative impact of training data quality versus quantity on language model performance. The finding was clear: data quality plays a more significant role in overall model performance than data quantity.
This is not an edge case finding. It is the consensus across the machine learning research community. A well-curated dataset that covers the problem space comprehensively will outperform a vast, unrefined one. Volume without curation does not produce depth. It produces noise.
The contrast between IBM Watson’s oncology AI and Google DeepMind’s AlphaFold tells the story at industrial scale. Watson was trained on synthetic and limited data. It failed in real-world hospitals, producing inaccurate treatment suggestions. AlphaFold was built on meticulously curated, validated experimental datasets. It predicted 200 million protein structures. Same era. Same underlying technology class. Opposite foundations. Opposite results.
The early promise of deep learning was that you could feed any dataset into a neural network and skip the curation step. That turned out to be wrong. The success behind GPT-4, the model that changed the public conversation about AI, had as much to do with investments in data curation as algorithmic research. The teams that built the models everyone now uses did not win by having the most data. They won by having the best data.
One researcher framed the principle in a way that any buyer can understand. Training AI with undifferentiated data is like training a chef with hundreds of random recipes that have missing ingredients and unclear instructions. You do not get mastery. You get confusion.
The AI implementation market runs on this exact dynamic. When a provider says their system draws from a broad base of frameworks, industry knowledge, and extensive sources, that is not a selling point. That is a description of a blended foundation. Breadth without curation produces the same result as best practices: output that is competent, generic, and structurally identical to what the next provider will give you.
The question is not how much is in the foundation. The question is who decided what belongs there and what does not.
How to Test Any Provider’s Claims About What the AI Is Built On
The four questions below probe the foundation directly. They are not the same as the evaluation framework in Article 1 of this series, which covers the provider’s overall process. These focus on one specific layer: what the AI actually knows when it produces output, and whether anyone curated that knowledge or just collected it.
“What specific frameworks is your AI built on, and who authored them?”
A curated foundation can name its sources. Not “industry best practices” or “leading research.” Specific frameworks. Specific authors. If the answer is vague, the foundation is blended. Vagueness is not modesty. It is the absence of curation. A provider who built on proven methodology can tell you exactly what is in there and exactly why. A provider who collected everything cannot, because there is no selection logic to explain.
“What did you leave out, and why?”
This is the question that separates collection from curation. A blended foundation has no exclusion criteria. Everything went in. A curated foundation has a filter, and the provider can describe it. What got rejected? On what basis? This question works because it asks the provider to show the negative space. What is not in the system tells you more about the system than what is.
“How do you handle conflicting methodologies?”
Every domain has competing frameworks. Sales methodology alone has a dozen established systems that contradict each other on fundamental principles. A blended foundation treats them as additive. All twelve go in. The AI averages them. A curated foundation resolves the conflicts before the AI ever encounters them. The provider chose one over another, or synthesized a specific position, and can explain the reasoning. If the answer is “we include multiple perspectives so the AI has a broad view,” the foundation is blended. Breadth without resolution is not a strength. It is the mechanism that produces generic output.
“Can you show me the difference between your output and what a default AI produces on the same prompt?”
This is the acid test. Take a real piece of your content. Run it through generic AI with a basic prompt. Then run it through the provider’s system. Compare the two outputs side by side.
If the provider’s output is structurally different, meaning it contains strategic reasoning, specific frameworks, and operational logic that the generic output does not, the foundation is doing real work. If the output is cosmetically different, meaning the tone shifted but the substance did not, the foundation is not curated. It is configured. And configuration is not the same thing. (Article 5 in this series covers that distinction in detail.)
Any provider who has done the work of building a curated, proven foundation will welcome this test. The output proves the foundation. That is how proof works.
The Difference You Can Trace
Strip the marketing from both approaches and one difference emerges that tells you everything.
A blended foundation cannot explain its own output. Ask why the AI recommended a particular approach, and the honest answer is that it averaged everything in the system. No specific framework produced it. No specific rationale drove it. The output is a statistical composite of every methodology that went in, contradictions included. The buyer pays custom prices. The output delivers commodity results. And nobody can trace any specific recommendation back to a specific source with a specific reason.
A curated foundation can. Ask the same question and you get: this recommendation follows a specific framework, grounded in specific research, selected for specific reasons. The buyer can evaluate the rationale. They can push back on it. They can test it. Because the foundation is traceable, the output is accountable.
That traceability is the signature of a curated foundation. If you cannot trace the output back to a specific rationale, the foundation is blended. If you can, someone did the work.
The Question That Cuts Through Every Sales Conversation
Every article in this series hands you a question that exposes the truth about a provider’s process. This one cuts to the foundation.
“What is this built on, and how was it curated?”
The answer separates the two foundations instantly. A provider who names specific frameworks, describes the selection criteria, explains what was excluded and why, and can show you the difference in the output has a real foundation. A provider who says “best practices,” “industry-leading research,” or “a broad base of knowledge” has a blended one. And blended foundations produce blended output. The same competent, interchangeable, generic result you can already get for free.
The most important thing about any AI system is not what you see in the output. It is what was decided before the AI ever produced anything. What went in. What stayed out. And whether anyone made those decisions deliberately, or whether they just collected everything and let the AI average it into the same middle that everyone else already occupies.
The next time a provider tells you their AI is “built on best practices,” you will know exactly what that phrase means. And you will know the question that turns a sales conversation into an evaluation.
“What is this built on, and how was it curated?”
