What repetitive tasks can be automated with AI and which ones need human approval?
· CompaniesAutomation
Which repetitive tasks AI can do alone, which it prepares for a person to approve, and which remain solely for people, by department. Three levels of authority, the questions to decide the level, and how to relax the gate.
Repetitive tasks that have stable rules, high volume, and a reversible error can be automated with AI from start to finish: classifying emails, extracting data from documents, preparing drafts, updating records, answering already documented questions. Tasks that involve money, contracts, people, sensitive data, or critical communications, and any action that cannot be undone, require human approval. In the middle, the AI prepares and a person approves.
What rule separates what AI does alone from what needs approval?
Three conditions for automating from start to finish: stable rules (most cases are solved the same way and exceptions are recognized), high volume (dozens or hundreds of cases per month), and reversible error (if the agent makes a mistake, it is corrected without damage). If all three fail, it is not automated; if only the third fails, it is automated with approval.
And five signs that a person is needed before something goes out: it involves money (a payment, a discount), it involves a contract (conditions, deadlines, signatures), it affects people (a hiring, a dismissal), it touches sensitive data (health, payroll, bank details), or it is a critical communication (a complaint, an important client). To these, irreversibility is added: deleting, sending, signing, and paying cannot be undone. To order which process to address first, the framework of which processes to automate first crosses hours and criteria.
What AI does alone, what it prepares, and what stays with people, by department
The table lists the most common tasks of an SME in six departments. The middle column is the longest in almost all companies and where most of the savings are found: the agent produces and the person approves in seconds.
| Department | AI does it alone | AI prepares and a person approves | Only people |
|---|---|---|---|
| Admin and Finance | Extract data from received invoices, classify expenses, match invoices with orders, prepare bank reconciliation | Proposed entries, payment orders, issued invoices, cash flow forecasting, responses to the consultancy | Unusual payments, negotiating with bank or supplier, deciding on a bad debt |
| Sales | Qualify leads with written criteria, log activity in the CRM, follow-up reminders | Quotes and proposals, responses to price requests, discounts within the authorized margin, follow-up on important accounts | Final negotiation, discounts outside policy, signing a contract |
| Customer Service | Answer documented questions (hours, order status), classify and route tickets, request missing data | Undocumented cases, returns and credits within policy, responses to public reviews | Serious complaints, customers threatening to leave, compensation outside policy, legal cases |
| Operations | Transfer data between systems, generate delivery notes and labels, status notices, stock alerts, daily reports | Orders to suppliers, route or shift planning, changes in committed deadlines | Stopping a line or a shipment, accepting an order that cannot be fulfilled, safety decisions |
| Human Resources | Initial screening with written criteria, answers about internal policy (vacation, leave), administrative onboarding | Job offers, communications to staff, absenteeism reports, training proposals | Interviews and hiring, evaluations, dismissals, sanctions, any health data |
| Marketing | Content drafts, schedule approved posts, campaign reports, segment lists | Posts and newsletters, budget changes in campaigns, responses to comments | Positioning and pricing, crisis communication, agreements with partners or media |
More detail in AI agent use cases by department.
The three levels of authority of an agent
The three columns translate into three levels of authority, and each action is assigned to one, not the entire agent: the same agent can read invoices with full authority and propose payments with prior approval.
- Low risk: executes. The agent performs the action and continues: reading, classifying, extracting, recording in a correctable field, sending an internal reminder. It is reviewed by sampling.
- Medium risk: executes and logs for review. The agent acts, but the action remains on a list that a person reviews later and can undo: closing a ticket, updating a price in an unsent quote, moving an opportunity stage.
- High risk: prior approval. The agent prepares everything and stops. Nothing goes out until a person approves it after seeing the result and the sources used. This is the level for money, contracts, people, sensitive data, critical communications, and everything irreversible.
How they are implemented with real permissions (own users, minimum scopes, limits by amount) is in AI agent governance: permissions and limits.
How to decide the level of each task?
With five questions per action, answered by the person who does it today and by the person responsible for it. If any points to high risk, the action goes to prior approval; it is not averaged.
- Is it reversible? If it can be undone in minutes without anyone outside noticing, it can be executed; if not, prior approval.
- Does it affect money or a contract? Any amount that goes out or condition that is committed, prior approval. A collection reminder does not move money; a payment order does.
- Does it touch personal or sensitive data? Reading to complete a task can be low level with limited permissions; sending, exporting, or deciding on that data, prior approval.
- How many times a month does it occur? If it's few, it's not worth automating yet.
- What happens if it makes a mistake? A wrong piece of data in an internal field is corrected; a wrong email to an important client is not. Write down the worst specific error and the level comes out on its own.
Tasks that do not pass these questions in any way are not automated even if technology allows it; typical cases are in when not to automate with AI.
How is the gate lowered over time?
With data, not with trust. Each task with prior approval records how many results were approved without being touched and what was corrected. When the correction rate for a type of case stays low for several consecutive weeks and the remaining corrections are cosmetic, that type of case moves down a level: from prior approval to execute and log, and later to execute.
The lowering is done by case type, not by the entire task: in quotes, those for a standard catalog can go out alone while those including custom items remain gated. And it is reversible: if the correction rate goes up, the case type returns to the previous level until the cause is fixed. Three things never move down from prior approval regardless of how good the rate is: payments, contracts, and decisions about people.
This is how Companies Automation designs authority limits
Companies Automation builds AI Factories (agents that produce repetitive work while the team reviews and decides) with these limits from the start: each agent has its own user with minimum permissions in each system, each action is traced with what data was used and who approved it, and human approval is mandatory where there is money, contracts, or sensitive data. Client knowledge is not used to train third-party models, and GDPR and the AI Act are considered in the design, not at the end. This design is applied first in their own businesses ("we are our first client"). How the entire system is protected, from permissions to logs, is in AI agent security and governance.
Frequently Asked Questions
Can AI send emails to customers without anyone reviewing them?
It depends on the email. An order status notice, an appointment reminder, or a documented answer to a frequent question can go out alone because the content is fixed and the error is minor. A response to a complaint, a quote, or an email to an important account goes through prior approval.
Doesn't human approval eliminate the savings of automating?
No, because approving takes seconds and producing takes minutes: the person checks the points they know fail and approves; searching for data, writing, and formatting is no longer done by them. Additionally, the gate is relaxed by case type when the correction rate stays low, so the cost of reviewing decreases over time.
What tasks should never be automated?
Decisions about people (hiring, firing, evaluating, sanctioning), signing contracts, unusual payments, and crisis communications. AI can prepare information for all of them, but the decision and responsibility remain with a person. It is also not advisable to automate processes that change every month.
How do you know an agent hasn't exceeded its authority?
Because technically it can't: the user it acts as in each system only has the permissions of its level, and high-risk actions require a person to release them. Furthermore, every action is recorded with the date, data used, and approver, so any case can be audited later.
If you want to know how many hours of your company are in the middle column and what the first task would be, the free diagnosis (9 questions, 2 minutes) returns the euros you lose per year on repetitive tasks, your AI First maturity level, and the first step.