From Chat to Routine: The Five Steps of AI in Operations

Almost every company already uses AI, but use tends to stop at the same point: someone opens a chat, asks a question, copies the answer and moves on. The gain stays with the person and never reaches the operation. Between that one-off use and a task that runs on its own every day there are five steps, and each one depends on the one before.

This text lays out those steps as an AI maturity assessment. You leave knowing what each step changes in practice, where each one stops, which step your company is on and what the next move is.

What is the AI maturity ladder?

It is a simple way to measure how much AI is part of the company's work. It does not count tools purchased: it measures how much the business context, the systems and the decisions are connected to AI.

Climbing does not mean abandoning the step below: a company on step 5 still uses chat every day. The difference is that repetitive tasks leave people's hands and start running under rules.

What do the numbers say about AI in companies?

• Widespread use. In McKinsey's The state of AI in 2026, nearly nine in ten respondents say their organization regularly uses AI in at least one business function.
• Chat ahead, agents behind. In the same survey, chatbots are the most widely scaled tool (47% of respondents), while about two in ten say they have reached the scaling phase with AI agents.
• Little effect on results. Only 37% attribute any impact on operating profit (EBIT) to AI. Among the best performing companies, nearly three quarters redesigned workflows because of AI, against one quarter of the rest.
• In Brazil, adoption is accelerating. According to IBGE's PINTEC Semestral, the share of industrial companies with 100 or more employees using AI rose from 16.9% in 2022 to 41.9% in 2024. The areas with the most use were administration (87.9%) and sales (75.2%).

Using AI has become the norm. The return shows up when it stops being a conversation and becomes part of the process, which is the path of the ladder.

Why does AI get things wrong, and what does that have to do with the steps?

Three fundamentals explain the ladder.

• The model predicts the next piece. A language model generates text by choosing, piece by piece, what is most likely to come next, based on learned patterns. It does not query your company's data.
• When information is missing, it fills the gap. That is where hallucination comes from: without the right data, the model fills the gap with what seems most likely. The answer sounds confident and may be wrong.
• What is not in the window does not exist for the model. The context window is everything the model can take into account when answering, as Anthropic's documentation defines it. The same documentation notes that more context is not automatically better: what matters is putting the relevant information there.

Each step puts more of the right information in the window and then gives the model the means to act on it.

Step 1: what changes when the company uses AI in a chat?

• What it is. One-off questions and answers in a chat app, with no company information beyond what the person types.
• What changes in practice. Individual gains in drafts, summaries, proofreading and quick research.
• Example. A salesperson asks for help writing a first contact email.
• Limit. Without context, the model answers with the internet's average case. The email comes out generic, does not mention your product and has to be rewritten.
• Signal to climb. People repeat the same explanations about the company in every conversation, or say that "AI doesn't understand our business".

Step 2: what is chat with context?

• What it is. The same chat, now with standing instructions, files and projects that carry the business context.
• What changes in practice. The answer starts talking about your company, in your tone, within your rules.
• Example. The support team creates a project with the help center and service policies, and AI suggests replies to tickets based on that material.
• Limit. AI only sees what someone pasted or attached. Today's data, in the CRM or the ticketing system, stays out.
• Signal to climb. The team spends more time copying data from systems into the chat than reading the answer.

This is where the difference between a prompt and a rule comes in. A prompt applies only to that conversation. A rule is saved and applies to every flow: project instructions, custom assistants and, in development teams, files such as AGENTS.md. The business context is written once and reused.

A good prompt, or a good rule, has four parts:

Using examples has a technical name, _few-shot_, and Anthropic's prompting best practices describe it as one of the most reliable ways to steer the format, tone and structure of the output.

Step 3: how does AI start accessing the company's systems?

• What it is. The chat gains connectors and starts querying systems: CRM, spreadsheets, email, ticketing tool, database.
• What changes in practice. Data enters the window without anyone copying and pasting, and the answer uses current information.
• Example. Finance asks which customers are more than 15 days late on payment, and AI queries the billing system to answer.
• Limit. AI depends on someone asking. It queries, but does not carry the task from end to end.
• Signal to climb. The same query is made every week, and the person always takes the same steps after the answer.

The standard that made this step simpler is MCP (Model Context Protocol), an open standard for connecting AI applications to external systems. The documentation compares MCP to a USB-C port: one standard, many systems. Read and write are separate permissions, and the right way is to start with read only. If your systems do not talk to each other yet, read the guide to systems integration first.

Step 4: what sets an agent apart from a chat?

A chat answers. An agent executes. It receives a task and repeats a loop until it is done:

• What it is. A program that uses the model to decide and tools to act, with defined permissions.
• What changes in practice. A multi-step task comes out whole: search, cross-check, classify and draft.
• Example. In sales, the agent receives an account name before the meeting, researches the company, cross-checks it with the CRM history and delivers a summary with points of attention.
• Limit. Someone still has to trigger the agent. If the person forgets, the task does not happen.
• Signal to climb. The agent is always triggered at the same time or always after the same event.

How much the agent does on its own depends on the cost of an error: the more expensive the error, the lower the autonomy. See AI agent autonomy levels, and the step by step in how to build an AI agent for operations.

Step 5: what is an AI routine?

A routine is the sum of three parts: trigger + agent + destination.

• Trigger. A schedule (every day at 8 a.m.) or an event (new lead on the website, ticket opened, invoice overdue).
• Agent. Searches, summarizes, classifies, decides or drafts.
• Destination. Where the result lands: team chat, email, spreadsheet or CRM.

As with the other steps:

• What changes in practice. Nobody has to remember to ask, and the team gets the result where it already works.
• Example. Every Monday morning, a routine reads the finance system, builds the weekly revenue and cost summary and sends it to management. In support, each new ticket is classified on arrival, and cases with signs of cancellation trigger an alert to the owner.
• Limit. A routine needs follow-up: measuring, adjusting rules and handling exceptions. Without it, it fails silently.
• Sign of maturity. Every routine has an owner, a success metric and a log of what it did.

There are examples for sales, support, finance and product in AI routines by area. To choose the first ones, UEEK's AI opportunity map lists where AI pays off in your business and in what order to tackle it.

Which step is your company on?

Answer yes or no. The first "no" shows where to focus.

1. Do people use AI at work frequently? If not, the focus is step 1: access, guidance and simple cases.
2. Is the business context written down and reused? Project instructions, policies, tone of voice and saved examples. If not, the next move is step 2.
3. Does AI query at least one company system without copy and paste? If not, the next move is step 3.
4. Is there a multi-step task that AI completes on its own, with defined permissions? If not, the next move is step 4.
5. Is there a task that runs without anyone triggering it, on a schedule or an event? If not, the next move is step 5.
6. Does every routine have an owner, limits on what it can do and a success metric? If not, the company climbed the step without the foundation to stay on it.

The most common mistake is jumping straight to an agent without organized context and without access to data. The agent inherits the gaps of the steps below.

Why does shorter working time raise the pressure to climb?

The end of the 6x1 schedule in Brazil (six days of work for one day off) raises a practical question for many operations: how to keep the same volume of work with fewer hours available. Part of the answer lies in process and team scheduling. Another part lies in taking repetitive work off people, and climbing from chat to routines is one way to do that with control. The topic is covered in the 5x2 schedule: how to keep operations running without losing productivity, and the background of the change in the end of the 6x1 schedule.

Which precautions apply to every step?

• Data protection (LGPD). Customers' personal data only enters a flow with a legal basis and at the minimum necessary, as required by the purpose and necessity principles of Brazil's Law 13,709/2018.
• Cost. Reasoning models and high volume raise the cost per task. In the McKinsey survey, one in five respondents says their organization limits AI use because of operating costs. Choose the model according to the task.
• Human review. Mandatory where errors are expensive: customers, finance and decisions.
• Credentials. Keys and passwords never go into a prompt, document or repository. The agent uses its own credential, with the least access possible.

Frequently asked questions

What is AI maturity?

It is how closely AI is connected to the business context, the systems and the day-to-day tasks. With low maturity, AI is used in a one-off way. With high maturity, there are routines that run on their own, with an owner and controls.

What is the difference between chat with tools and an agent?

In chat with tools, the person asks and AI queries a system to answer. With an agent, AI receives a task and decides the steps on its own, calling tools in a loop until it is done.

Can you skip steps?

In practice, no. An agent without written context gets things wrong like the step 1 chat, and without access to data it has nothing to execute. You can climb fast, but the foundation of each step has to exist.

Do I need a technical team to reach step 5?

For simple routines, off-the-shelf automation and scheduling tools may be enough. When the routine touches core systems, customer data or money, integration, permissions, monitoring and testing come into play.

Where do I start?

With the self-assessment above and a list of repetitive tasks, with frequency, time spent and cost of error. With that, it becomes clear which routine brings the most return with the least risk.

Sources

• McKinsey, The state of AI in 2026: On the road to ROI
• IBGE, PINTEC Semestral 2024: use of artificial intelligence by industrial companies (in Portuguese)
• Model Context Protocol, official documentation
• Anthropic, context window documentation
• Anthropic, prompting best practices
• Law 13,709/2018, Brazil's General Data Protection Law (in Portuguese)

How UEEK can help

UEEK has worked for 14 years with digital products, for more than 200 clients, in 6 countries, on more than 500 projects. In AI, we start with the assessment: the AI opportunity map shows which step the company is on and which routines bring the most return, and the AI map with a specialist takes that analysis deeper with your team. Once the routine is defined, we build and run AI agents in production, integrated with your systems and with clear rules on what they can do on their own.

Take the free AI opportunity map