AI is a loudspeaker, not an auto-tuner

LLMs keep getting more capable. They read messy data that older software would have rejected. The Model Context Protocol (MCP) lets an agent plug into almost any system.

Put the two together and the pitch writes itself. The platform and data work that has stalled for a decade can be skipped. The money goes straight to agents. That removes the unglamorous line from the technology budget.

The promise is half right. The models are new, but what they read has not changed.

A loudspeaker changes the volume, not the signal

A model tolerates a messy format. It does not tolerate inconsistent data.

Take two systems of record that hold different rates for the same customer. For years a person sat between them, noticed the mismatch and asked.

An agent does not ask unless someone designed it to. Most were not. It picks one rate and repeats it in every channel at once.

Even an agent built to flag the mismatch cannot say which rate is right. That is a business decision.

An auto-tuner makes up for a singer's weaknesses in real time, and the audience never knows. Nothing does that for an API. A missing field stays missing. A rate limit stays a limit. The agent reaches it sooner and more often.

I set the strategy for opening a global logistics platform's APIs and services to tools such as ChatGPT. It taught me that you have to choose, consciously, which APIs go behind an MCP server. An API designed for internal use cannot be opened to agents overnight. An API already struggling with rate limits will fall flat when MCP multiplies its traffic.

AI cannot make up for deficiencies in the API estate, and it makes them louder.

Gartner predicted in early 2026 that 60% of agentic analytics projects relying on MCP alone will fail by 2028, citing the lack of a consistent semantic layer. Agents need shared definitions and a common language that tell them what the data means.

Choosing the best model matters less than it appears. The model is the one part of the stack every competitor can rent.

The model is rented. The context is owned.

Think of the stack as a signal chain. Source data sits in the systems of record, and the platform and its APIs carry it. The AI tier amplifies it. The customer hears the result.

Build the organisation's context once, and every model, agent and channel reads it instead of starting from scratch.

Context here means how the business works. It covers customers, contracts, products, prices and processes, described the same way everywhere.

A model that reads it gives answers that fit the organisation. Swap the model next year and the context carries over.

I saw this while building a pricing platform for a global B2B business. Before an algorithm could price anything, it needed catalogue, master data, routing, capacity and order inputs from separate systems. The bulk of the effort went into getting those systems to talk to each other. The algorithm only amplified what those systems could tell it. With that work done, the platform went live in many countries and improved contribution margin by double digits.

A system of decision needs a common language

AI can be the connective tissue across your systems of record. But only once they share a common language.

On the pricing platform, one system held buy rates. A separate system held sell rates. The two schemas had nothing in common. The systems were speaking different languages.

We had to define a new language that each system's own language could be translated into. Only then could a model reason across buy and sell rates together. That common language is the context.

Once the systems share that language, AI can work as a system of decision, the layer that sits on top of the systems of record. It reads across them and recommends an action. On the pricing platform it recommended prices, and commercial teams accepted or overrode them. A person decides, and AI can then carry the decision out in each system.

A system of decision is only as good as what the systems of record expose.

Turn the volume up on the source that is coherent

"Fix the data first" has been the standard advice for a decade. It has also been a reason to ship nothing.

Do not wait for perfect data. Think of your AI budget as a volume knob. Turn the volume up on the source that is already coherent.

Let the loudspeaker show you where the rest of the stack is inconsistent. Fix that, and go again. Each turn of the loop widens what the AI tier can reliably read.

An AI answer engine I launched on a global B2B website went live in under a year. It could ship because it ran on one source that was already coherent. A single content layer fed every page on the website and the apps. We did not begin with the transactional systems that disagreed with each other.

Where the value sits on systems that disagree, scope the foundational work to one decision, not the whole estate. We did that with pricing.

The foundational work on data, APIs and a common language used to be the hardest line to defend in a technology budget. It now decides what the AI budget returns. No model will auto-tune it for you. Keep funding the work the loudspeaker exposes.

This is the first of three pieces on AI as an amplifier. The next turns to judgement, and to why the human stays in the loop as automation takes on more of the workflow.