How AI works

Two diagrams for the decision you are about to make.

Most AI decisions go wrong because nobody has drawn the thing being bought. These are the drawings: where the new layer sits in the technology you already run, and what happens to a single request inside it. Both come from systems we run in production.

Diagram 1

Where AI sits in what you already run

AI is not a replacement for your technology stack. It is one new layer added between the screens your people use and the systems that hold the record.

Diagram in three parts. First, the three layers every business already runs, shown in grey: applications where people meet your business, business systems where the work is recorded, and foundations of data, infrastructure and security. Second, the same three layers with one new orange layer added between the applications and the business systems, taking work that needs judgement, reading the record and writing the outcome. Third, that new layer opened up into seven parts: the model, what it can read, what it retains, what it can act on, where a person approves, continuous assurance, and the audit trail. The model is one part in seven.

Where AI sits in a business technology stack: three layers you already run, and one new layer between the screens and the systems of record.

Open the full-size diagram
  1. 1

    The three layers you already have

    Applications where people meet your business, the systems where the work is recorded, and the data, infrastructure and security underneath. Every business running today has all three, whether or not anyone has drawn them.

  2. 2

    The one layer AI adds

    A layer that takes the work needing judgement, reads the record, and writes the outcome back. It sits between the screens and the systems. It does not replace either.

  3. 3

    That layer, opened up

    Seven parts: the model, what it can read, what it retains, what it can act on, where a person approves, continuous assurance, and the audit trail. The model is one part in seven, which is why choosing a model is one decision in seven.

Diagram 2

How an AI system is actually put together

The same architecture behind every system on this site, drawn once: what happens to a single request from the moment it arrives to the moment the answer goes back.

Systems architecture diagram. Channels such as web, mobile, email and voice feed an application layer of interface, public API, access and billing. A request needing judgement passes into the AI layer, where it moves through four numbered steps: assemble the context, ask the model, check the answer, take the action. Those steps draw on memory, a model gateway holding GPT, Claude and Gemini with routing and fallback, and a set of tools. A control rail of guardrails, human approval, evaluation and an audit trail runs across all four steps. The AI layer reads and writes the systems the business already runs on, such as bookings, quotes and client records, and the data infrastructure beneath them. The answer returns to the application layer.

How the pieces connect: a request is assembled, sent to the model, checked and acted on, drawing on memory, a model gateway and tools, with control running across all four steps.

Open the full-size diagram
  1. 1

    Four steps, in order

    Assemble the context, ask the model, check the answer, take the action. Every request follows the same four. A system that skips the third step is the one that puts an invented figure in front of a customer.

  2. 2

    What the steps draw on

    Memory, a model gateway holding more than one model with routing and fallback, and a set of tools the system is allowed to use. The gateway is why a model going down is an incident rather than an outage.

  3. 3

    The control rail

    Guardrails, human approval, evaluation and an audit trail run across all four steps rather than sitting at the end. Control is not a gate the answer passes through last; it is present at every step or it is not present.

  4. 4

    What it reads and writes

    The systems the business already runs on: bookings, quotes, client records, and the data infrastructure beneath them. The AI layer uses the same paths a person does, which is what keeps one audit log rather than two.

Use the diagrams on your own systems.

The second diagram is a checklist as much as a picture. Take one process you are considering automating and put it through the four steps: what context gets assembled, what the model is asked, what checks the answer, and what action it is allowed to take.

The argument behind the drawings

Both diagrams were drawn for one article: what a single AI answer is made of, which parts you rent, which part you have to build, and the three ways this supply chain behaves unlike any other you run.

Read next

The AI supply chain: what you are actually buying

Most AI decisions go wrong because nobody has drawn the supply chain. Here is what one AI answer is made of, which parts you rent, which part you have to build, and the three ways this supply chain behaves unlike any other you run.

Read the article

Make the next AI project one the business can measure.

Thirty minutes with the founder, no slides. You will know what the right solution looks like, what it would take to build, what it should return, and which part to start with.

Replies come from a named person in Dubai within one working day.