AI Routines by Department: Examples for Sales, Support, Finance, HR and Operations

AI routines by department: ready-to-adapt examples for sales, customer service, finance, HR, product and inventory, with trigger, destination, autonomy level and success metric, and how to pick the first one.

Published on October 30, 2026. 13 min read

An AI routine is an operational task that runs on its own, at a set time or when something happens, and delivers the result where your team already works. The formula fits on one line: trigger + agent + destination. The trigger fires, the agent does the work (search, summarize, classify, decide or draft) and the result lands in Slack or Teams, in an inbox, in a spreadsheet or in the CRM.

This article is a practical catalog. For six areas of the company, each routine follows the same format: trigger, input, work, destination, autonomy level and what measures success. At the end you get a summary table, a way to choose the AI model, the tool categories and an eight-question exercise to design your team's first routine.

What is an AI routine?

In a chat, the AI waits for you to ask. In a routine, it works without anyone asking. It is the last step described in from chat to routine: the five steps of AI in operations, and it is where the hours come back, because the task no longer depends on someone remembering to do it.

AI routine: trigger, agent and destination

The trigger fires, the agent works on the input data and the result lands where the team already is.

To see how the agent is built inside, with instructions, connectors and tests, read how to build an AI agent for operations.

How to read each example in the catalog?

The field that needs the most attention is the autonomy level, which defines how far the AI goes without a person:

Suggests. The AI prepares a draft and a person decides.
Acts with approval. The action only goes out after someone's sign-off.
Acts and notifies. The AI acts and reports what it did.
Acts alone. The AI acts and only logs it.

The more expensive the mistake, the lower the step. Anything that talks to customers, moves money or affects a person starts at "suggests" or "acts with approval". The article on AI agent autonomy levels shows how to move up safely.

Sales and customer service open the catalog for a reason: in its study on the economic potential of generative AI, McKinsey estimates that about 75% of the value of the use cases analyzed falls in four functions: customer operations, marketing and sales, software engineering, and research and development.

Which AI routines work in sales?

Account research before the meeting

Trigger. Every day at 8 a.m., for that day's meetings.
Input. The company's website and news, CRM history.
Work. Summarizes what the company does and what has been discussed, and suggests two questions.
Destination. Direct message in Slack or Teams and a note in the CRM.
Autonomy. Acts alone. It is internal material, and the salesperson checks it.
Success. Meetings with a ready briefing and prep minutes saved.

Follow-up draft after the call

Trigger. Meeting ended and notes saved in the CRM.
Input. Notes or transcript (with consent) and the proposal.
Work. Drafts the email with a summary, next steps and agreed deadlines.
Destination. Draft in the salesperson's inbox.
Autonomy. Suggests. It goes to the customer, so it never leaves without review.
Success. Share of follow-ups sent the same day.

Which AI routines work in support and customer service?

This is where the evidence is strongest. In the NBER study Generative AI at Work, Brynjolfsson, Li and Raymond followed 5,179 support agents: with an AI assistant, issues resolved per hour rose 14% on average and 34% among novice and less experienced agents.

Ticket triage and category

Trigger. New ticket.
Input. Ticket text, customer data and the list of categories.
Work. Classifies category and urgency and flags cancellation requests.
Destination. Fields in the ticketing system and an alert in the team channel when there is churn risk.
Autonomy. Acts and notifies. A wrong category is fixed in seconds.
Success. Accuracy in a weekly sample and time to first response.

Suggested reply from the help center

Trigger. Ticket classified.
Input. Help center, policies and previously approved replies.
Work. Drafts the reply and points to the article it relied on.
Destination. Draft inside the ticket.
Autonomy. Suggests. With a high acceptance rate, simple questions can move to "acts with approval".
Success. Drafts accepted without edits and average resolution time.

Which AI routines work in finance?

Payment reminders for overdue accounts

Trigger. Every day at 9 a.m.
Input. Accounts receivable in the ERP and contact history.
Work. Decides who gets a reminder and in what tone, and drafts the message.
Destination. Email to the customer and a record in the ERP.
Autonomy. Acts with approval at first. Negotiating terms or discounts always stays with a person.
Success. Average days overdue and payments after the reminder.

Weekly revenue and cost summary

Trigger. Monday at 8 a.m.
Input. ERP, bank statements and budget spreadsheet.
Work. Compares with the previous week and the budget and points out deviations.
Destination. Email to leadership.
Autonomy. Acts alone. The numbers come from the system; the AI organizes and comments.
Success. Report-building hours eliminated.

Bank reconciliation

Trigger. Statement imported, once a day.
Input. Statement and payables and receivables.
Work. Matches entries and explains what does not match.
Destination. Spreadsheet of open items.
Autonomy. Acts and notifies on exact matches; suggests on discrepancies.
Success. Share reconciled without human touch and days to close the month.

If you already have several candidate routines and need to decide which is worth more, the AI map with an expert prioritizes the opportunities with your team, area by area.

Which AI routines work in payroll and HR?

Building and checking shift schedules and timesheets

Trigger. Weekly for the schedule; daily for timesheets.
Input. Planned schedule, clock-ins, vacations, days off and the rules of the collective agreement.
Work. Proposes the schedule respecting rest periods and days off, and flags missing clock-ins, short breaks or unplanned overtime.
Destination. Timekeeping system and a notice to each team's manager.
Autonomy. Suggests for the schedule; acts and notifies on inconsistencies.
Success. Inconsistencies caught before payroll and HR hours at closing.

This routine matters more as working hours change: companies leaving the 6x1 schedule (six days on, one off) need to redo schedules without leaving shifts uncovered. See the end of the 6x1 schedule: what changes, deadlines and affected sectors.

Résumé screening with human review

Trigger. New application.
Input. Résumé and the job's objective requirements.
Work. Compares against the requirements, summarizes and justifies each point.
Destination. Recruiting system.
Autonomy. Suggests. The AI never rejects anyone on its own.
Success. Time to first screening and agreement with the recruiter in a sample.

Care is doubled here. Models can reproduce bias, so remove from what the agent reads anything that should not weigh in (name, photo, age, address) and audit the results. Brazil's data protection law, the LGPD, limits processing to the minimum necessary (art. 6, III) and guarantees the right to request review of decisions made solely through automated processing (art. 20).

Which AI routines work in product and management?

Customer feedback grouped by theme

Trigger. Every Friday.
Input. Tickets, satisfaction surveys, app store reviews and notes from sales.
Work. Groups by theme, counts frequency and quotes real examples.
Destination. Product team document.
Autonomy. Acts alone.
Success. Themes that became decisions in planning.

Daily metrics summary

Trigger. Every day at 8 a.m.
Input. Analytics, database and CRM.
Work. Compares the previous day with the average and highlights only what is out of the ordinary.
Destination. Management channel in Slack or Teams.
Autonomy. Acts alone. A mistake in an internal summary costs little.
Success. Time to notice a drop and shorter status meetings.

Which AI routines work in operations, purchasing and inventory?

Inventory alert

Trigger. Every four hours or at every relevant stock movement in the ERP.
Input. Stock on hand, recent sales and supplier lead times.
Work. Calculates days of coverage left and decides who needs to know.
Destination. Purchasing channel.
Autonomy. Acts and notifies.
Success. Stockouts per month.

Purchase order draft

Trigger. Inventory alert fired.
Input. Purchase history, approved suppliers and negotiated terms.
Work. Drafts the order with quantity and supplier and explains the calculation.
Destination. Draft in the ERP and a notice to the buyer.
Autonomy. Acts with approval. The order commits cash.
Success. Time from alert to order and orders approved without changes.

What is the summary of routines by area?

Area	Routine	Trigger	Destination	Autonomy
Sales	Account research	Every day at 8 a.m.	Slack or Teams and CRM	Acts alone
Sales	Follow-up draft	Meeting ended	Email draft	Suggests
Support	Ticket triage	New ticket	Ticketing system	Acts and notifies
Support	Suggested reply	Ticket classified	Draft in the ticket	Suggests
Finance	Payment reminder	Every day at 9 a.m.	Email to the customer	Acts with approval
Finance	Weekly summary	Monday at 8 a.m.	Email to leadership	Acts alone
Finance	Reconciliation	Statement imported	Open items spreadsheet	Acts and notifies
Payroll and HR	Schedules and timesheets	Weekly and daily	Timekeeping system	Suggests
Payroll and HR	Résumé screening	New application	Recruiting system	Suggests
Product	Feedback by theme	Every Friday	Product document	Acts alone
Management	Metrics summary	Every day at 8 a.m.	Management channel	Acts alone
Operations	Inventory alert	Every four hours	Purchasing channel	Acts and notifies
Purchasing	Purchase order	Alert fired	Draft in the ERP	Acts with approval

How to choose the AI model for each routine?

Weigh quality, speed and cost. Providers charge per input and output token, so the same flow can cost very different amounts depending on the model.

Reasoning models. More accurate on multi-step tasks, such as reconciliation with exceptions, but slower and more expensive, because the reasoning also consumes tokens.
Fast, cheap models. For high volume and simple tasks, such as classifying thousands of tickets.
Models strong at writing. For what the customer reads: follow-ups, payment reminders and support replies.
Models with web search. When the routine depends on current information, such as account research.

It is common for a company to use two or three models, each where it gives the best balance. Start with the simplest one that does the job.

Which tools can run AI routines?

Real-time research. Assistants with web search, for prospecting and monitoring.
Work agents. They operate files, the browser and systems, such as the agent modes in Claude and ChatGPT.
Automation orchestrators. n8n and Make connect systems with AI in the middle of the flow, and both have time-based triggers: n8n's Schedule Trigger and Make's scenario scheduling.
Scheduled tasks in chat apps. Some apps already let you schedule a task for the agent to run every day, without building a flow.

The connection to your systems usually goes through the Model Context Protocol (MCP), an open connector standard in which read and write access are granted separately.

Off-the-shelf tool or custom-built agent

It is common to start with an off-the-shelf tool and move critical routines to a custom agent.

When the routine is critical, UEEK puts the AI agent in production, connected to your tools and with human approval, and handles the agent operations afterwards. If the flow depends on a system that does not exist yet, the path is custom software development.

How do AI routines help keep operations running on a 5x2 schedule?

Each routine gives hours back to the team. Added up, those hours are what make it possible to shorten the work week without losing output and without hiring at the same rate. The summary built by hand, the reconciliation that took the morning, the triage that filled the first hour: that time goes back to the work that needs a person. The math starts with the first question of the exercise below, and the article 5x2 schedule: how to keep operations running without losing productivity shows how to organize it.

How to design your first AI routine?

Gather the people who do the task today and answer in writing:

Repetition. Which task repeats, how often and how much time does it take?
Trigger. Does it start at a time or with an event?
Inputs. Where does the data come from? Can the AI access that source?
Work. Search, summarize, classify, decide or draft?
Output. What is delivered, in what format, to whom and where?
Limits. What must the routine never do?
Cost of error. If it gets it wrong, what is the impact? That defines the autonomy.
Success. How will you know it worked?

The answers become the routine's prompt: context, task, constraints and an example. Then place the candidates on an impact and effort matrix:

High impact, low effort. Start here: accessible data, frequent task, cheap mistakes.
High impact, high effort. Plan it. It usually needs integration or a custom agent.
Low impact, low effort. Good for testing tools.
Low impact, high effort. Drop it for now.

Only 17% of Brazilian companies with ten or more employees use AI technologies, according to Cetic.br's TIC Empresas 2025: 15% among small companies and 50% among large ones. One well-chosen routine already puts a company ahead of most. For a starting point, the AI opportunity map is free and takes about ten minutes.

Before turning on any routine, four precautions: customer and employee data only goes in with a legal basis and at the minimum necessary; human review wherever a mistake is expensive; keys and passwords never in a prompt or document; and the model chosen for the task, because reasoning and high volume increase the cost.

Frequently asked questions

Do I need to code to create an AI routine?

Not for the simplest ones. Chat apps with scheduled tasks and orchestrators such as n8n and Make build flows without code. Code comes in when there are systems without connectors, complex rules or fine-grained access control.

What is the difference between an AI routine and traditional automation?

Traditional automation follows fixed rules. An AI routine handles what does not fit a fixed rule, such as reading a badly written ticket or drafting a reply. The two usually work together, as in process automation.

Which routine should I implement first?

The one that combines high frequency, accessible data and cheap mistakes. A daily metrics summary, ticket triage and account research are usually good starting points.

Can AI send messages to customers on its own?

It can, but it should not start that way. Begin at "suggests" or "acts with approval", measure how many drafts are accepted and only move up for standardized messages.

Can I use AI to screen résumés without breaching the LGPD?

Yes, by processing only the necessary data, informing candidates, removing data that causes bias, auditing the results and keeping a person in every decision.

How long does it take to put a routine live?

With an off-the-shelf tool and accessible data, a few days. When it involves several systems and approval rules, UEEK puts the agent in production in up to 4 weeks.

Sources

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 what pays off: in the AI map with an expert, we interview each area, prioritize routines by impact and effort and define the approval rules. Then we put AI agents into production and handle the agent operations. For a first look, take the free AI opportunity map.

Book time with an expert

Topics

AI routines
AI automation examples
AI in sales
AI in customer service
AI in finance
AI in HR
automating repetitive tasks