Which Processes to Automate With AI First: A Decision Matrix

Which Processes to Automate With AI First: A Decision Matrix

The first processes to automate with AI are the ones that happen often, do little damage when they go wrong, and follow a rule your team can explain in writing. A matrix with three axes, volume, risk and rule clarity, lets you compare candidates and pick a single one for the pilot. The goal is not to automate as much as possible. It is to learn from one real workflow, with limits and a record of what the AI did.

Key takeaways

• Pick the first process using three criteria: volume (how often it happens), risk (what happens if it goes wrong) and rule clarity (whether someone can write down how the decision is made).
• The best first candidate usually has high volume, low risk and a reasonably clear rule, with a person handling the exceptions.
• A process with a fully fixed rule often needs a simple deterministic workflow, and AI only enters where text must be read or judgment is needed.
• A high-risk process with a vague rule should not be first, however attractive it looks.
• Define the success metric and the stop criterion before you switch anything on.

Why the first process matters more than the tool

When a company decides to use AI, the conversation almost always starts with the tool: which model, which platform, which integration. In practice, the process you choose decides whether the project works. A well-chosen process tolerates an imperfect tool. A poorly chosen process sinks even the best technology.

There is also a learning effect. The first automated workflow shapes how your team sees AI for months afterward. If the first pilot touches something sensitive, makes a visible mistake and nobody can explain why, trust drops and the next proposal meets resistance. If the first pilot is modest, well measured and reversible, it becomes an internal reference.

Anthropic, in its material on building agents, recommends starting with the simplest solution that solves the problem and adding complexity only when there is a demonstrable gain. The matrix below applies that idea to choosing a process.

The three axes of the matrix: volume, risk and rule clarity

Each axis answers a different question. Use three levels on each one: low, medium and high.

Volume: how often does the process happen? High volume gives you material to measure, adjust and justify the setup effort. Low volume makes measurement hard, because you never collect enough cases to know whether the workflow works. Regularity matters too: a process that runs every day teaches you more than one that runs once a quarter.

Risk: what happens if the automation gets it wrong? Consider money, the customer relationship, legal obligations, security and reputation. A mistake in an internal summary is cheap to fix. A mistake in a collection message, a refund or a reply involving personal data can be expensive or irreversible. Risk also depends on reversibility: what can be undone with one click weighs less than what leaves the company.

Rule clarity: can someone explain in writing how the decision is made today? A clear rule means criteria that two different people would apply in a similar way. A vague rule means the outcome depends on who handles the case that day, on undocumented context or on intuition. The vaguer the rule, the harder it is to judge whether the AI got it right.

One important detail: a clear rule does not mean no judgment is involved. Reading a customer message in free-form language and classifying the topic takes judgment, but the final classification can still use well-defined categories. That is where AI tends to help most: a clear rule applied to a messy input.

How to read the matrix: a decision guide

Once you have the three levels for each candidate, use this guide to reach a decision.

• High volume, low risk, clear or partly clear rule: the ideal candidate for a first pilot. Automate the repetitive part and leave exceptions with a person.
• High volume, medium risk, clear rule: a good candidate, as long as a human approves any action that cannot be undone.
• High volume, high risk, clear rule: do not start here. Leave it for a second stage, with an approval policy and a record of every decision.
• High volume, any risk, vague rule: document the rule first. AI can help find patterns in past cases, but it should not decide yet.
• Low volume, low risk, clear rule: solve it with a simple workflow, without AI, or leave it as it is. The gain rarely pays for the effort.
• Low volume, high risk: avoid it. Little to gain, a lot to lose.

When a process has a fully fixed rule, ask whether it needs AI at all. A deterministic workflow in a tool such as n8n, Make or Zapier often solves it with more predictability. If you are comparing options, n8n vs Make vs Zapier for Business: Cost, Hosting and Control helps with that part. AI belongs where there is free-form text, varied documents or classification that fixed rules do not cover.

Step by step: scoring your processes

The exercise fits in one working session with the people who actually run the processes, not only management. Follow this sequence.

• List five to ten repetitive processes, written as concrete actions, for example: answering questions about opening hours, sorting incoming emails, checking registration documents.
• For each one, note who performs it, how many people touch it, and where it starts and ends.
• Give each a volume level: low, medium or high. Use the team's sense of it and, if possible, the system or inbox history.
• Give each a risk level by asking: if the AI gets this wrong, who notices, when, and what does it cost to fix?
• Give each a rule clarity level by asking two people to describe separately how they decide. If the answers differ, the rule is still vague.
• Apply the decision guide and set aside the candidates that remain.
• Among those candidates, choose only one. Break ties by picking the one with a clear owner, accessible data and a metric that is easy to measure.

The tiebreaker matters more than it seems. A process with no owner has nobody to approve changes, and a process without accessible data forces you to start with heavy integrations. Both delay the pilot.

Illustrative scenario: three candidates in a services operation

Imagine, as an illustrative scenario, a services company that serves customers on WhatsApp and has three candidates on the table.

The first is answering recurring questions about opening hours, address and required documents. High volume, low risk, clear rule, because the answers already live in a document. It is the natural candidate, with handoff to a person when the question falls outside known content.

The second is negotiating late payments. Medium volume, high risk, partly clear rule, because each case depends on history and on concessions. It waits, until a sequence and limits are defined. The article on AI debt collection on WhatsApp: sequence, limits and tone shows why this topic needs more care.

The third is deciding which customers get a discount. Low volume, high risk, vague rule. It stays out, and the recommendation is to document the rule before thinking about automation.

The result: the first process goes to the pilot, the second joins the queue for a later stage, and the third becomes a documentation task. None of these decisions depends on choosing a model or a platform.

Signs the process is not ready yet

Even a well-scored candidate may not be mature. Before you start, check this list.

• Nobody in the company owns the process and can approve changes.
• Example cases are scattered across personal chats, with no organized access.
• The right answer changes depending on who handles the case, and nobody recorded why.
• Personal or sensitive data is involved and there is no legal basis or retention definition yet. For conversations, see LGPD compliance for AI on WhatsApp: legal basis, consent and records and the guidance from Brazil's data protection authority (ANPD).
• No human is available to take over when the AI cannot answer.
• Nobody has defined what a good result looks like, or when the pilot should stop.

Each item can be solved, but solving them beforehand is cheaper than discovering them with the workflow in production. If human handoff is the bottleneck, the WhatsApp AI agent human handoff playbook explains how to structure it.

How to map, limit, pilot and hand over

After choosing the process, our method has four moves. Map: describe the current workflow as it really happens, with inputs, outputs, exceptions and a baseline of what is done today. Limit: define which data the AI may read, which actions it may take and what requires human approval. Pilot: run a real workflow, with a record of decisions and a success metric agreed beforehand. Hand over: give the team the operation, the documentation and the criteria for deciding whether to expand, adjust or end it.

The central point is that the pilot needs an exit criterion. If results do not appear, stopping is a legitimate decision, not a failure. To structure that agreement, the AI pilot charter template helps you record scope, metric and limits. For follow-up, also read how to measure an AI automation pilot.

To think about risk in a structured way, the NIST AI Risk Management Framework offers a useful vocabulary for identifying, measuring and managing the risks of AI systems, and it can serve as a reference when you define the risk axis of your matrix.

Choose the process you can measure and undo, not the one that looks most impressive.

Frequently asked questions

Can I start with customer support on WhatsApp? Yes, as long as the scope is limited to recurring questions with a documented answer, and there is handoff to a person. Support that involves financial decisions, exceptions or sensitive data should wait for a second stage. WhatsApp has the advantage of already concentrating the volume, which makes measurement easier.

Is AI worth using if the process rule is already completely clear? Often, no. If the input is structured and the rule is fixed, a deterministic workflow is usually more predictable and cheaper to maintain. AI makes sense when the input is free text or a varied document, or when you need to classify and summarize before the rule is applied.

What if two processes tie on the matrix? Break the tie by process owner, data access and how easy the result is to measure. If the tie persists, pick the lower-risk one that is easiest to reverse. The other stays in the queue, and what you learn from the first pilot tends to improve its score.

Who should take part in the choice? The people who run the process day to day, the person accountable for it and, when personal data is involved, whoever handles privacy and security. Deciding only at the management level tends to hide exceptions that the operations team knows well. Those exceptions are exactly what defines risk and rule clarity in the matrix.

Sources

• Anthropic: Building effective agents
• NIST AI Risk Management Framework
• ANPD: Autoridade Nacional de Proteção de Dados