What Founders Should Delegate to AI and What to Keep

What Founders Should Delegate to AI and What to Keep

Founders should delegate to AI the work that happens often, costs little when it goes wrong and can be undone. They should keep the decisions that are rare, high in value or hard to reverse. To decide case by case, cross three questions: how often does this happen, what does a mistake cost, and how much is the decision worth. AI prepares and executes. The judgment that sets the direction of the company stays with the person who answers for it.

Key takeaways

• Good delegation means sorting tasks by frequency, risk and decision value, not asking what the AI is capable of doing.
• Frequent, low-risk tasks are the first candidates, with sampling to check quality.
• Rare, high-value or irreversible decisions stay with the founder, and AI only helps prepare the analysis.
• Between the two extremes sits the draft model: AI prepares, a person approves before any external action.
• Every delegation needs a data limit, an action limit, a record of what was done and a named human owner.

Why founders need a filter, not a list

Lists of tasks to hand to AI age quickly and ignore the context of each company. What is routine at a clinic can be a sensitive decision at a construction firm. The same email can be a calendar notice or a negotiation that defines the year.

Founders often have another problem: they picked up roles out of necessity and, over time, started deciding things that never needed them. Answering repeated questions, organizing the inbox, summarizing meetings, checking whether an order is complete. None of this requires an owner's view, yet it eats the attention that should go to pricing, hiring, partnerships and strategy.

A filter solves this because it works for tasks that do not exist yet. When a new tool or process shows up, the question stays the same: where does this fall on the three dimensions? This is also consistent with the guidance of people who build systems on language models: start with the simplest solution, add autonomy only when it is needed, and keep oversight where mistakes are expensive. Anthropic describes this progression in its Building effective agents article, separating predictable workflows from agents that decide their own steps.

The three questions: frequency, risk and decision value

Every task goes through the same three questions. Answer with high, medium or low, without trying to calculate anything precisely.

Frequency: how regularly does this happen? A daily task repeated with little variation has enough pattern to be described, tested and monitored. A task that appears once a year does not justify the setup and rarely produces the data needed to adjust it.

Risk: what happens if the AI gets it wrong? Three points matter here. Is the error visible before it reaches the customer? Is it reversible once it has? Does it involve personal data, money, health, contracts or reputation? A wrong draft that you fix before sending is low risk. A wrong message sent to an important customer, or an incorrect charge, is another level. Personal data calls for extra attention to Brazil's LGPD, and the national data protection authority (ANPD) is the official reference on the subject.

Decision value: how much does this choice change the company? Picking a meeting time has low value. Deciding who gets a discount outside policy, who to hire or which customer to serve first has high value, because it sets precedent and shapes culture.

The rule of thumb that sums up the filter:

The more frequent, less risky and less valuable the act, the more it can go to AI.

The reverse also holds. If the task is rare, the mistake is costly and the decision carries weight, it stays with you.

Decision guide: what each combination calls for

In practice, the most common combinations lead to five different postures. Use this list as a decision guide.

• High frequency, low risk, low value: delegate with automatic execution and spot-check sampling. Example: sorting incoming messages by topic, summarizing voice notes and meetings, organizing files.
• High frequency, medium risk, low or medium value: delegate in draft mode. AI prepares, a person approves before sending. Example: customer replies, standard proposals, routine collection notices.
• Medium frequency, medium risk, medium value: delegate the preparation and keep the decision. AI gathers data, compares options and flags open questions. Example: supplier analysis, sales pipeline summary.
• Low frequency, high value: keep with the founder. Use AI as a preparation assistant to collect information, simulate questions and review arguments, without outsourcing the choice.
• High or irreversible risk, any frequency: keep with a human, or allow automation only with explicit approval and technical limits. Example: payments, dismissals, contract changes, crisis communication.

One important detail: high frequency alone does not justify delegation. A daily task can carry high risk, such as answering complaints that include a legal threat. In that case, AI can classify and escalate, but it should not decide the content of the reply.

What is usually good to delegate first

The safest starting point is preparation and organization work, where AI produces material that someone reviews. A few examples that tend to pass the filter well:

• Email triage and drafting, with no automatic sending. There is a step-by-step guide in AI inbox triage for busy founders.
• Summaries of meetings, calls and voice notes, with action items identified and assigned to people.
• First replies to repeated customer questions, with clear rules about what can be said.
• Checking orders, proposals and records for missing or inconsistent fields.
• Weekly reports compiled from the spreadsheets and systems the company already uses.
• Initial research on suppliers, competitors or regulations, always with sources the founder can verify.

What these items share: the output is verifiable, the error shows up before it causes harm, and the volume is large enough for the time savings to be real. Start with one of them, not several at once.

What the founder should keep

There is a set of decisions that should not leave the founder, even if AI can produce a good answer. The reason is not only technical. It is about accountability: someone has to answer for the choice to customers, partners, staff and the law. The article what not to delegate to AI in your business goes deeper on this. Here is the summary through the filter:

• Strategy and priorities: where to invest, what to stop doing, which market to serve.
• People: hiring, promoting, dismissing, resolving conflicts. AI can organize information, but the decision and the conversation are human.
• Exceptions and precedents: a discount outside policy, a concession to a strategic customer, a settlement in a dispute. What you decide once becomes an implicit rule.
• High-value relationships: partners, investors and the customers who sustain the operation.
• Legal and financial commitments: signing contracts, non-routine payments, positioning during a crisis.
• Brand values and tone: what the company accepts and what it does not, even when it would be profitable.

Keeping does not mean working alone. You can ask AI to list the risks in a proposal, point out contradictions in a contract or simulate an investor's objections. The right use is as a second pair of eyes, not as a substitute for the person who decides.

Illustrative scenario: a week for a distributor's owner

Imagine, purely as an illustration, the owner of a materials distributor who takes orders over WhatsApp and email. He spends much of the day answering delivery questions, checking orders and approving exceptions.

Applying the filter, he sorts what he does:

• Delivery status questions: high frequency, low risk, low value. These go to AI, which reads from the order system in read-only mode, gives a standard reply and escalates to the team when an order has a problem.
• Checking orders with incomplete fields: high frequency, medium risk. AI flags what is missing, and a team member confirms before release.
• Draft replies to customers complaining about delays: medium frequency, medium risk. AI drafts, the support owner reviews and sends.
• Commercial terms for a major customer: low frequency, high value. These stay with the owner, who uses AI only to gather purchase history and points of attention.
• Dismissing or warning an employee: high risk, human impact. This stays entirely with the owner and legal counsel.

The result is not the owner's absence from the operation. It is a change in the kind of task that reaches him. Less routine checking, more deciding what the routine cannot resolve.

How to apply it: map, limit, pilot, hand over

The filter only works when followed by a method. We use four steps as a base.

• Map: list what you do in a typical week, and note frequency, who is affected and what happens if there is an error. If the process is still unclear, start with mapping business processes before applying AI. Also record a baseline: how long the task takes today and what the current quality is.
• Limit: define which data the AI can see and which actions it can take. Read first, write later. Any action that leaves the company, such as a message, a charge or a record change, requires an explicit rule or human approval. Input, output and action validation are covered in guardrails for AI agents. For permissions, the principle of least privilege, as described in OWASP's authorization guidance, is the right default.
• Pilot: choose one real workflow, with a log of everything the AI did and a success metric agreed before you start. A small pilot reveals failures that no theory will show.
• Hand over: at the end of the pilot, deliver documentation, owners and a review routine. Then decide whether to expand, adjust or stop. Stopping is also a valid outcome.

To organize risks in a structured way, the NIST AI Risk Management Framework offers a useful vocabulary for mapping, measuring and managing the risks of AI systems, even in small companies.

One last check before delegating anything new:

• Is there a named human owner for this task?
• Do we know what the AI can access and what it cannot?
• Would the error be noticed in time, and can it be corrected?
• Is there a record of what was done, so we can audit it later?
• Have we agreed how to tell whether it is working and when to roll back?

If any answer is no, the delegation is not ready yet.

Frequently asked questions

Where should a founder start delegating to AI? Start with a frequent, low-risk task that is easy to check, such as email triage or meeting summaries. Measure the current time, run it for a few weeks with human review, and only then expand. Starting with a few tasks lets you learn without exposing customers to mistakes.

Can AI make decisions in place of the business owner? For low-value, reversible operational decisions with clear rules, yes. For strategic, people, exception or legally significant decisions, AI should only support with information and analysis. Responsibility for the outcome stays with whoever decides.

How do you stop AI from sending something wrong to a customer? Use draft mode with human approval while the workflow is new, restrict what the agent can say and do, and keep a record of every reply. For outbound messages, follow the channel's rules, as with WhatsApp. Over time, the lowest-risk replies can be released with sampling, while sensitive cases keep human review.

Does delegating to AI require care with data protection law? Yes, whenever personal data of customers, employees or candidates is involved. Define which data the system can access, for what purpose, who is responsible and how data subjects can be served. If you are unsure about legal basis or supplier contracts, consult a legal professional and the ANPD guidance.

Sources

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