The AI System Is the Product

From prompt to system

For the past few years, much of enterprise AI has followed a simple pattern: input, model, output. Give the model a prompt and get back a summary, classification, recommendation or piece of generated text.

That pattern works for isolated tasks. More consequential applications add context, reasoning, tools, workflow, action, evaluation and feedback.

A complete AI system can determine what is happening, remember what happened before, decide what needs to happen next, interact with other software, execute an action and check whether the intended outcome occurred. Building that system is the engineering work.

What an AI system needs

Imagine an AI system handling an incoming customer request.

A basic implementation sends the message to a language model and asks what to do. A production system needs to know much more: who the customer is, which products they use, what conversations already happened, which policies apply, the current state of the request, which tools the AI can access, which actions it may perform automatically, what requires approval, what happens when information is missing, what happens when the model is uncertain, and what happens after the action completes.

The model may perform some of the reasoning. Most of the value sits in the system around it.

OpenAI's guidance on building agents draws a similar line. It recommends agents for workflows with complex decision-making, difficult rule sets, or unstructured information, and notes that deterministic software may still be better when those conditions are absent.

Use AI where ambiguity or judgement makes traditional software insufficient.

The applied AI stack

A useful way to think about production AI is as seven layers.

• Context: user history, organisational knowledge, previous conversations, database records, policies, documents, workflow state, and prior actions. A stronger model with poor context still produces the wrong answer.
• Reasoning: interpreting ambiguous language, extracting information, comparing options, recognising patterns. Use it only where the step requires judgement. If policy says a transaction above a threshold always needs approval, encode the rule.
• Tools: query a database, retrieve a document, update a CRM, create a ticket, send a message, schedule an appointment, call another service, update workflow state.
• Workflow: sequences, dependencies, exceptions, and handovers. The system needs to know where it is in the process and what happens after each step.
• Control: permissions are product architecture. Categorise an enquiry automatically; require approval before issuing a refund, changing a medical plan, signing a contract, or transferring money.
• Evaluation: if an agent uses five tools across eight steps, a wrong final answer tells you little about where the system failed.
• Feedback: was the recommendation accepted, corrected, or reversed? Did the customer reopen the case? Did the clinician change the recommendation? Outcomes close the loop.

Stripe shows what embedded AI looks like

Stripe Radar shows embedded AI working inside a transaction.

When a payment occurs, Stripe's machine-learning systems analyse the transaction and help determine how it should be handled. The system combines model-generated risk information with payment signals, issuer information and business rules.

Stripe reported in 2025 that adaptive Radar rules increased payment success rates by 1.3 percentage points with minimal change in fraud rates. At Stripe Sessions, the company also reported reductions in unnecessary authentication challenges and fraud following improvements to its AI systems. These company-reported figures measure operational outcomes rather than model benchmarks.

The intelligence sits inside the transaction and acts before a user needs to ask whether the payment looks fraudulent.

The product is the system around the intelligence

Foundation models will keep getting faster, cheaper, and more capable. Access to a strong model is unlikely to remain much of a competitive advantage on its own.

The engineering work moves to giving intelligence the right context, connecting it to the right tools, constraining what it can do, embedding it in the workflow, measuring whether it worked, and improving it from real-world outcomes.

The first wave of enterprise AI helped humans produce work. The next wave will help organisations operate the work itself. Companies building the next generation of AI products should stop thinking primarily about AI features and start thinking about AI systems.