12 Things AI Agents Cannot Do (And Never Will)

AI agents cannot be held accountable, cannot keep a promise from one session to the next, and cannot decide what a company stands for. Those limits are structural, so the useful question for a team is where to draw the line between agent work and human work, and who owns each side of it.

Most writing about AI agent limits treats them as temporary. Some are. An agent that writes weak code today may write good code next year. But a second group of limits does not move with model quality, because it comes from what an agent is: software with no stake, no memory it can be held to, and no identity that carries consequences. This guide lists twelve of those limits, then walks through what they mean for the boundaries, policy and oversight a team needs before it lets an agent act or answer.

The 12 things AI agents cannot do

1. Carry final accountability

An agent cannot be fired, demoted, or held to a past commitment. Accountability needs someone who stays the same person and lives with the result. When an agent acts, the accountability still lands on a human: the person who set it up, the team that runs it, or the company. No model upgrade changes that, because it is a fact about who can face consequences.

2. Hold a position over time

An agent's answer can shift between sessions, between prompts, or with small changes in wording. It has no settled view that it defends from one day to the next. That rules it out for any role that needs steady advocacy, such as pushing back on a bad plan for weeks or holding a negotiating line.

3. Take real risk

When a person recommends something, their name and reputation ride on it. When an agent recommends something, nothing rides on it. In any setting where advice matters because the adviser has something to lose, an agent cannot send the same signal.

4. Make commitments that bind the future

An agent can write the words "I will," but nothing holds the next session to them. The model can change, the setup can change, and tomorrow's agent may not honor today's answer. Contracts and promises need a party that can be bound forward in time.

5. Carry knowledge that comes from experience

A senior engineer often knows a service is fragile without being able to say exactly why. That kind of knowledge builds up through years of doing the work. An agent only has what is written down in its training and its context. Where expertise lives in experience, agents stay behind the people who have it.

6. Build relationships

A working relationship has memory, shared history and trust on both sides. An agent can recall past messages, but it has nothing invested. Roles that depend on relationships, such as large deals, partner work or senior hiring, stay with people.

7. Handle situations nobody has seen before

Agents do well inside the range of cases they have seen. A first-of-its-kind event, with no precedent and no playbook, is where they do worst and where people are most needed.

8. Make choices about identity or values

Should we serve this customer? Take this money? Work with this partner? Those are questions about who the company is, not what scores best. An agent can supply analysis. The choice belongs to people who answer for it.

9. Apologize in a way that means something

An apology matters because the person making it commits to doing better and puts their standing on the line. An agent can produce the words but cannot pay the cost, and customers can tell.

10. Earn trust over time

Trust grows when someone keeps their word through many small tests. An agent can build a history of outputs, but it cannot be trusted in the way a colleague is trusted, because it has nothing to lose by breaking that trust.

11. Act against its own instructions for a better reason

A person can choose quality over speed even when speed is what gets rewarded. An agent does what it was set up to do. Doing the right thing against the incentive stays a human act.

12. Decide when to stop

Given an open task, an agent keeps going until it is stopped or its stopping rule fires. Knowing when work is good enough means weighing time, cost and judgment, and that call still needs a person to frame it.

What these limits mean for your team

The twelve limits point to one practical rule: agents can prepare, draft and answer within lines that people set, and people keep the calls that need a stake, a promise or a judgment about values. The sections below cover the decisions a team has to make to put that rule into practice. Each one links to a full guide.

Set boundaries before you deploy

Boundaries are the written rules for what an agent may read, what it may answer, what it may do on its own, and when it must stop and hand off to a person. Most teams write them after something goes wrong. It is cheaper to write them first, because every boundary you skip becomes a judgment call the agent makes for you.

Good boundaries are specific. "Be careful with customer data" is not a boundary. "Read the support queue, never the billing system, and send any refund question to the support lead" is. Our guide to the boundaries every AI agent needs goes through them one at a time, with the questions to ask before you deploy. The AI agent readiness check scores how ready your team is.

Decide who owns AI policy

Someone has to own the rules: who approves a new agent, what it may touch, and what happens when it gets something wrong. In many companies that job falls between IT, legal, security and the people team, and so nobody owns it. That gap is where risky deployments slip through.

The answer depends on the company, but the owner needs the authority to say no and the time to review new tools. See who owns AI policy for the roles that usually hold it, and why the head of people is often missing from the room.

Keep a named human accountable

Because an agent cannot carry accountability (limit 1), every agent needs a named person who does. That person reviews what the agent did, fixes mistakes, and answers for the result. The risk to watch for is quiet drift, where the agent keeps answering and nobody remembers who signed off on it.

The same applies when AI speaks for a colleague. If a tool answers questions in someone's name, that person should know what it said and be able to stand behind it. Our guide on how to keep human accountability with AI covers the review habits and ownership rules that make this work.

Keep a human in the loop where it counts

Human in the loop (HITL) means a person checks, approves or corrects an AI system's output before it takes effect. It is the most direct way to apply limits 1, 7 and 8: the agent does the routine part, and a person makes the call that carries consequences.

The trap is putting a human on every step, which wastes time and trains people to click "approve" without reading. Put people where an error would be costly or hard to undo, and let the rest run. The guide to human in the loop AI explains the patterns and where each one fits.

Prefer AI that says when it does not know

The most dangerous answer an AI tool can give is a confident wrong one. A tool that says "I don't know, ask Priya" is safer than one that fills the gap with a guess, because a gap you can see is a gap you can fix.

That is why grounding matters: an answer should come from a named source, and when no source covers the question, the tool should say so. Read why an AI that says "I do not know" is the safer one for what to look for when you choose a tool.

Know how agents fail

Agents rarely fail loudly. More often they give an answer that is almost right, credit the wrong person, or act on stale information. Each of those comes back to the limits above: no stake in being right, no experience to notice something odd, no habit of stopping.

Knowing the common patterns lets you design checks for them. Our list of AI agent failure modes covers the ones that show up most in engineering and operations teams, and how to catch each one early.

Add a governance layer

A governance layer is the set of rules and records that sits between your tools and the people relying on them. It answers three questions for any agent: what is it allowed to do, who decided that, and what did it do. Without one, every new agent adds its own rules in its own place.

The layer does not need to be a product. It can start as a short written policy and a log of decisions. See what a governance layer is for how engineering teams set one up.

Choose AI that helps people instead of replacing them

The limits above suggest a simple test for any AI tool: does it make a person better at their job, or does it try to stand in for their judgment? Tools in the first group save time and keep people in charge. Tools in the second group create more output for someone to check.

The guide to AI tools that help people instead of replacing them compares the two approaches and the questions that tell them apart.

The AI agent boundaries checklist

Copy this into your planning doc before any agent goes live. Every line should have a written answer and a named owner. If a line has no answer, the agent is not ready.

Area	Question to answer in writing	Owner
Purpose	What single job is this agent for? What is out of scope?	Product or team lead
Read access	Which systems may it read? Which may it never read?	Security
Write access	May it change anything? If yes, what, and with what limit?	Engineering lead
Answers	Which questions may it answer on its own?	Team lead
Hand-offs	Which questions must go to a person, and to whom?	Team lead
Unknowns	What does it say when no source covers the question?	Product
Sources	Does every answer show where it came from?	Product
Labeling	Do people always know they are talking to an AI tool?	Policy owner
Approval	Which actions need a human to approve first?	Engineering lead
Accountability	Which named person answers for its results?	Executive sponsor
Records	Where is a record of what it said and did kept, and for how long?	Policy owner
Review	How often are its answers sampled and checked?	Team lead
Stop switch	Who can turn it off, and how fast?	Engineering lead
Policy	Which written AI policy does this deployment follow?	Policy owner

Where StandIn fits

StandIn is not an agent. It is a personal knowledge base, built on the work that you already do, that answers your teammates when you're away, in chat and email. It answers only from what you have written down and the tools you chose to connect, in your words, with a source under every answer. When the answer is not there, it never guesses: it says so and points to the person who knows. Every reply is labeled as your StandIn, it writes nothing into your tools, and you can review everything it said when you're back.

That design follows the limits in this guide. The person stays accountable, the answers stop at what was shared, and nothing is decided on anyone's behalf. See how StandIn works for the details.