Not every problem needs AI. Most AI problems don't need an agent.

Six ways to build it

“Should we use AI?” is the wrong question. Almost any task can be built in one of six ways, and each step down the list buys flexibility at the cost of money, speed and predictability.

The companies that build these systems give the same advice: start at the top. Google's Rules of Machine Learning open with: (Zinkevich, Google)

Don't be afraid to launch a product without machine learning. (Zinkevich, Google)

Anthropic, writing about AI agents, recommends “finding the simplest solution possible, and only increasing complexity when needed,” and adds that for many applications a single, well-prepared AI step is usually enough. (Anthropic, 2024)

Most real solutions combine several rows: ordinary software for the known steps, with AI only where a step needs it. The rest of this guide shows how to find those steps.

Describe the task step by step

Before choosing any technology, write the task down as you would explain it to a new colleague.

List every step, in order, with what goes in and what comes out. Then mark each step with one of four labels:

Most tasks turn out to be mainly rules, with one or two steps that need language or judgment. Those are the only places AI is needed. Microsoft's guidance for AI strategy starts at the same point: begin with the business problem, before considering AI at all. (Microsoft)

Rules you can write down belong in ordinary software

If you can write the rules on paper, ordinary software is cheaper, faster and fully testable.

OpenAI's guide to building agents lists three situations where AI earns its place:

• complex decisions with nuanced judgment, exceptions or context
• rule sets that have become so large that every update is costly or error-prone
• heavy reliance on unstructured data: emails, documents, conversations

Its advice for everything else is plain: (OpenAI)

Otherwise, a deterministic solution may suffice. (OpenAI)

“Deterministic” means the same input always gives the same output, as ordinary software does. Google's guidance adds the other side: once hand-written rules become too complex to maintain, learned models are easier to update. Its rule: “Choose machine learning over a complex heuristic.” (Zinkevich, Google)

Predict from your history, or work with language?

Classic machine learning and today's AI models solve different problems.

Google's guidance on framing machine learning problems gives a rule of thumb: to classify something or predict a number, use predictive machine learning; to create content or understand language, use generative AI. (Google)

Classic machine learning learns from historical data with known outcomes: past orders and which customers cancelled, past sensor readings and which machines broke down. Without that history, it has nothing to learn from.

The line is not strict. Google notes that large language models “can frequently perform text classification tasks better than predictive ML trained for the specific task,” because they already understand language. Sorting emails by topic, for example, often works without any training data of your own. (Google)

A fixed workflow, or an agent?

An agent is only needed when the steps can't be known in advance.

Anthropic distinguishes two kinds of AI system. In a workflow, the steps are fixed in code and AI handles some of them. In an agent, the AI itself decides which steps to take and which tools to use. (Anthropic, 2024)

Agents suit “open-ended problems where it's difficult or impossible to predict the required number of steps.” If you can draw the steps as a flowchart, a fixed workflow does the same job more cheaply and predictably. (Anthropic, 2024)

If an agent is the right choice, start small. OpenAI recommends getting the most out of a single agent with a few tools before splitting work across several agents. (OpenAI)

Decide how much the answer may vary

AI is not a calculator. The same question can get a different answer, and sometimes a wrong one.

Birgitta Böckeler of Thoughtworks puts it precisely: AI models “are NOT compilers, interpreters, transpilers or assemblers of natural language, they are inferrers.” Microsoft's guidance draws the same line between generative AI, whose output can vary for the same input, and deterministic systems for tasks where the same input must give the same result. (Böckeler, 2025) (Microsoft)

Böckeler suggests three questions to decide how much that matters for a task: (Böckeler, 2025)

• How likely is the AI to get it wrong?
• If it gets it wrong and nobody notices, what are the consequences?
• Will anyone notice when it gets it wrong?

Where decisions affect people, the law sets limits too. Under the GDPR, people should not be subject to a decision based solely on automated processing that has legal or similarly significant effects on them, such as an automatic refusal of a credit application, except in specific cases and with safeguards, including the right to human intervention. (European Commission)

In practice: keep decisions that must be repeatable in ordinary software, let AI prepare or read the input, and hand over to a person where the stakes are high. OpenAI names two triggers for that hand-over: when the AI fails repeatedly, and before actions that are sensitive, irreversible or high-stakes, such as cancelling orders, large refunds or payments. (OpenAI)

Count the running costs

Ordinary software costs little to run once it is built. Every AI step adds cost and waiting time to every request.

AI models are billed per use, and each call takes time, from a fraction of a second to many seconds. An agent makes many calls for one task. Anthropic notes that agentic systems “often trade latency and cost for better task performance.” (Anthropic, 2024)

Estimate early: how many requests per day, how fast an answer must come, and what one request costs. A step that is affordable a hundred times a day may not be at a hundred thousand, and a customer waiting at a checkout cannot wait ten seconds.

Agents also need more around them: quality checks, guardrails on what they may do, and a hand-over to a person. Those are part of the running cost. (OpenAI)

Buy, adapt or build

Before building anything, check whether you can buy it.

For common tasks such as writing assistance, meeting summaries or searching internal documents, ready-made tools exist and work with little setup. Microsoft's guidance describes the trade-off across the options, from ready-made tools to building on your own infrastructure: “As you move from the first model to the last, you gain control and give up speed.” (Microsoft)

Building pays off when the value lies in what only you have: your own data, your own processes, specialist knowledge. In between, many companies adapt: their own process and data around a vendor's AI model.

Whatever you buy, the steps above still apply. A bought tool is one row in your description of the task, with rules and checks around it.

Worked example: customer support emails

A 40-person company that sells and services professional coffee machines receives around 1,200 support emails a month, in German, English and Dutch. Customers describe problems in their own words: “it makes a grinding noise”, “the display shows E14 since yesterday”, “the milk is lukewarm again”. Here is the task, step by step.

This is where AI earns its place. No rule set reliably reads “it sounds like gravel” as a grinder problem, in three languages, with spelling mistakes. Everything around those steps stays ordinary software, and no agent is needed: the steps are the same every time. Nothing the AI produces goes out or commits the company without a person: every reply is sent by a support agent.

Before you choose

Go through this with whoever will build or buy the solution.

• The task is written down step by step, with inputs and outputs.
• Each step is labelled: rule, prediction, language or judgment.
• Rule steps are built as ordinary software.
• AI is used only at steps that need language or judgment, with a check after each.
• An agent is used only if the steps truly can't be known in advance.
• Decisions that must be repeatable, or that significantly affect people, stay in rules or with a person.
• Cost and response time per request are estimated for the expected volume.
• Ready-made tools were checked before building.
• Quality is measured on real examples before going live.

What this would look like for you