Why having the most powerful AI model doesn't guarantee results — and what does

Four months ago we published on this blog an article titled Enterprise RAG: The AI Revolution Transforming How Companies Access Their Knowledge. We argued that Retrieval Augmented Generation was the architecture that would let companies extract real value from generative AI. It was. It still is. But the industry conversation has matured over these months and the conclusion is uncomfortable for those selling technology and reassuring for those of us who have spent decades integrating systems: the success of an AI project depends less on the model than on the hand directing it.

Every week we speak with technology leaders, operations directors and executive committees asking themselves the same question: «which AI model should we use?». It is the wrong question. The right question is another: «what context do we have to feed it?».

This article is the answer to that question, and to why so many companies that have invested well — in licences, in consulting, in infrastructure — still do not get the return they expected.

The illusion of the most powerful model

The model is the most visible piece of an AI project. It is almost never the piece that decides the outcome.

In the last twelve months, the major AI providers have released increasingly powerful models: Claude Opus 4.7 with a million-token context window, GPT-5, Gemini Ultra. Any company can access them by paying a subscription. And that is exactly the problem.

When every competitor in your sector can access the same model, that model stops being an advantage. It becomes a commodity. What you have and no one else has is your information, your history, your processes, your customers and the specific way your company makes decisions. That is the context. And context cannot be bought: it has to be built.

Anthropic's own engineering team, the creators of Claude, put it recently: «the best model wins the demo; the best context wins production at scale». It is a sentence worth pinning on the wall of any executive committee evaluating an AI investment.

Buying a Ferrari does not make you a Formula 1 driver. Buying an Enterprise subscription to ChatGPT does not turn your company into a productive AI organisation.

The difference between the companies extracting real value from AI and the ones who are not is not explained by which model they chose. It is explained by what they built around the model.

Why so many AI projects fall short of expected results

AI projects don't «fail». They fall short because the system is missing what only the company itself can provide.

You will read in many places that 80 % of enterprise AI projects fail. It is a dramatic way to tell the story, but it is unfair to the technology itself. Models do what they are asked to do. The problem is not the AI: it is how it is being implemented.

When a company deploys a conversational assistant that responds with generic answers, or an agent that suggests the obvious, or a chatbot that just repeats what an FAQ already says, what is almost always missing is not model power. What is missing is:

• That the system knows what every internal thing is called in that company.
• That it knows what the last decision was on a given customer and why.
• That it understands the nuances of the sector — what «high» means in logistics is not what it means in banking.
• That it is connected to live data, not a PDF uploaded three years ago.
• That it has clear rules about what it can say and what it cannot.

Without any of this, the most powerful model in the world will give plausible but useless answers. And the company will mistakenly conclude that «AI doesn't work». AI does work. What hasn't been done is to give it the context it needed to solve the specific problem of that company.

This is the conversation we are having every week with clients arriving frustrated from a first pilot. The good news is that, once the real cause is identified, the solution is reachable. The less good news: it requires something that cannot be bought off the shelf.

What «context» really means in a company

Enterprise context is not one single thing: it is five different dimensions that are rarely worked on together.

When we talk about «giving context to AI», many companies imagine uploading documents to a system. It is a dangerous oversimplification, because it leaves out the four most valuable dimensions. Real enterprise context has five layers:

1. Decision history

Every company has spent years making decisions that are almost never documented: why a commercial proposal was rejected, why a supplier was changed, why an exceptional discount was approved. That decision memory — the «why» behind the «what» — is what distinguishes a mature company. And it is what an AI without access to it can never replicate.

2. Proprietary taxonomy

Every organisation calls things in its own way. «Active customer» does not mean the same thing in a consulting firm as it does in a telecoms operator. Neither does «closed order». If the AI does not handle the internal taxonomy, its answers will be correct in general terms and wrong in practical terms.

3. Internal policies

Who approves what, with what thresholds, under what circumstances, with what exceptions. Every company has an unwritten book of policies that live in emails, in the heads of managers and in tacit agreements. Without translating them into the system's context, the AI will end up proposing actions that violate internal rules nobody has told it about.

4. Tacit knowledge

This is the hardest dimension to capture and the most valuable. What that employee with fifteen years of seniority knows and is about to retire. What that account manager learned over two years working with a complicated client. What was decided in a meeting without minutes because «everyone got it». Turning tacit knowledge into explicit context is one of the most profitable things a company can do today.

5. Live data

Finally, context needs to be connected to real sources: the ERP, the CRM, the ticketing system, the document repositories. Not a three-month-old export, not a PDF in a shared folder. The AI has to know what the company knows today, not what the company knew when the pilot was built.

Building the five dimensions — well — is what separates an AI pilot that dies in six months from a platform that scales for years.

The role of the experienced human: the piece almost nobody puts at the centre

Without all three types of experience working together, any AI project remains incomplete.

Everything above — identifying the context the company has, structuring it, connecting it, governing it — needs a piece no AI model can replace: people with experience who understand the business and the technology at the same time.

There are three types of experience that have to be at the table, and it is very rare to find all three in one person:

Business experience. Knowing how the company really works: why each thing is done, which processes are held together with duct tape, which decisions are delegated and which escalate, which metrics actually move the business. Without this view, AI ends up optimising what does not matter.

Technology experience. Knowing how to choose between RAG, agents, fine-tuning or a classic system. Knowing when a knowledge graph adds more than a vector store. Knowing where to store data so that it meets GDPR and the AI Act. Without this view, AI ends up being a promise that does not survive in production.

Transformation experience. Organisations have antibodies. Introducing a new tool without the organisation rejecting it requires having been through this many times before. It requires knowing where resistance will appear, which manager has to be convinced first, which pilot will demonstrate value in six weeks. Without this view, AI ends up filed under «that project that didn't work out».

Without these three kinds of experience working together, the most powerful AI on the market is an expensive toy.

This explains why so many projects led only by IT departments do not scale, and why so many projects led only by the business do not reach production. AI transformation is a hinge between disciplines.

The two pieces properly understood: private RAG and Context Engineering

The architecture that actually works: a secure, sovereign base (private RAG) on top of which a layer of curation, structure and prioritisation (Context Engineering) operates.

With all of the above clear — that context matters more than the model, and that we need experienced people to build it — we can talk about the two technological pieces that really make the difference.

Private RAG: the base

Private RAG is the first layer. It means that all the company's information — documents, databases, structured knowledge — lives inside the organisation's own infrastructure, without leaving for external systems to be processed by third parties. It is the minimum condition of data sovereignty.

It is not an aesthetic preference. There are three strong reasons for RAG to be private:

• Intellectual property. Strategic information, contracts, internal procedures cannot be exposed in systems that may use them to train external models.
• Regulatory compliance. GDPR imposes clear obligations on personal data flowing through AI systems. The European AI Act adds new ones. Having RAG inside your infrastructure simplifies traceability enormously.
• Operational sovereignty. If your AI provider changes its pricing policy, has an outage or decides to stop operating in Europe, your information remains yours.

A well-built private RAG is a solid base. But on its own, it does not solve the problem.

Context Engineering: the layer that multiplies

On top of that base, Context Engineering operates: the set of techniques and disciplines that curate, structure, prioritise and deliver the right context to the AI at the right moment.

While RAG limits itself to retrieving relevant fragments of information for each query, Context Engineering does something deeper:

• It decides what information goes in and what does not, based on the user's role and the purpose of the query.
• It structures the context so the model can use it well (because delivering a lot of poorly organised information is worse than delivering little well-organised information).
• It prioritises what the AI should read first — a current contract weighs more than an expired one, an email from last week more than one from two years ago.
• It maintains memory across sessions, so the system does not start from scratch every time a user opens a conversation.
• It governs the flow: which decisions AI can take alone, which need human supervision, which need to be logged.

It is the difference between having a library with a million unsorted books and having an expert librarian who knows exactly which book to hand the reader for what they are trying to solve.

Together, plus the experienced human, equal real transformation

The formula that works in production is not either piece on its own. It is the combination: private RAG (the secure and sovereign base) + Context Engineering (the layer that curates and multiplies) + human experience (the hand deciding how to build it).

That trio is what is making the difference between the companies turning AI into a real competitive advantage and the ones still trapped in endless pilots.

The right question

If you have made it this far, you probably have in mind a project, a pilot or an AI initiative that is not delivering the results you expected. The tempting question is: should we change the model?. Almost always, that is the wrong question.

The right question is another: what context are we feeding this system, and does it make sense for what we are trying to solve?. And the question right after that: who, inside or outside the company, has the combination of business, technology and transformation experience we need to build that context properly?.

If you are asking yourself those questions, you are doing the right thing. And if you need to talk them through with someone who has spent decades integrating systems and years working with enterprise AI technologies, at Tecnea we can help you. We do not sell models. We help companies build the context and the processes that make any AI produce real results.

If you want an honest conversation about where your company is today and what it is missing to get where you want to go, request a free AI context diagnostic. Thirty minutes in which we tell you which pieces you have and which ones you are missing — no commitment, no technology sale.

Frequently asked questions

If the model matters less than it seems, why do providers talk about it so much? Because it is the most visible and the easiest to sell. Announcing a more powerful model generates headlines and commercially differentiates the provider from competitors. Context, on the other hand, is built internally by each company and cannot be packaged as a product. That is why it gets much less media attention, even though it is what really determines results.

Does this mean it doesn't matter which model you pick? No. There are real differences between models in latency, cost, reasoning quality, instruction-following, language support or context window. What we are saying is that this choice, with current models, usually accounts for between 10 and 20 % of the final result. The other 80 % comes from the context and how the system has been built around the model.

What is the difference between private RAG and cloud RAG? The key difference is where the information lives and is processed. In a private RAG, the data and the processing stay inside the company's infrastructure (on-premise or in a controlled private cloud). In a public cloud RAG, the information may pass through third-party systems, which has implications for intellectual property, regulatory compliance and operational dependency.

Does my company need a Context Engineering project, or is a good RAG enough? It depends on the use case. For simple informational assistants, a well-built RAG can be enough. For critical processes, complex customer service, operational decisions or any use case where answer quality is important, it is worth planning a Context Engineering layer from the start. The good news is that it can be built incrementally on top of an existing RAG.

How long does it take to implement a solution like the one you describe? A well-scoped initial pilot can deliver visible results in two to four weeks. Building a complete enterprise AI platform — with private RAG, Context Engineering and context governance — usually rolls out in phases of two to five months, prioritising the highest-value use cases. The investment depends on scope, but ROI is usually visible from the first use case if it has been chosen well.

Sources

1. Effective context engineering for AI agents — Anthropic
2. Context architecture is replacing RAG — VentureBeat
3. Context Engineering vs. RAG: Key Differences in 2026 — Atlan
4. AI Memory System vs RAG — Atlan
5. Introducing the Model Context Protocol — Anthropic
6. AI Trends 2026: agentic era and EU regulation — Opplus