AI agents

What are agentic experiences, and when are they worth building?

The test is whether the system finishes real work for the user, and what it does when it cannot.

Open Dimension AI3 July 20267 min read

What are agentic experiences?

Agentic experiences are products where an AI agent completes a task for the user instead of only answering a question. The user states a goal. The agent plans steps, uses connected tools, checks the result and asks for help when unsure. The hard part is safe action, not conversation.

The short version

An agentic experience is one where AI carries out a task for the user, not one where it only replies.
A chatbot answers, a copilot suggests, and an agentic system acts inside limits someone set.
A wrong action costs more than a wrong answer, so every action an agent can take needs its own permission and limit.
Agentic experiences suit frequent, bounded tasks with reversible outcomes and a clear route to a person.
General AI adoption surveys do not show how many businesses run agents in production.

What is an agentic experience?

An agentic experience is an interaction where the user states an outcome and an AI agent carries out the steps to reach it. The agent decides what to do next, calls the tools it has been given, checks the result and reports back. The user supervises the work instead of doing it.

Take a customer who wants to change a delivery address. On a website they find the order, open the right page and fill in a form. A chatbot explains where that page is. An agentic experience takes 'send it to my office instead', finds the order, checks it has not shipped, updates the address and confirms the change.

The term is loose. Vendors apply it to anything from a chat window with one tool to a system running long tasks across several applications. When someone offers you one, ask what it can change without a person approving it. That answer tells you more than the label.

How is an agentic experience different from a chatbot or a copilot?

The difference is who takes the final action. A chatbot gives information and stops. A copilot drafts or suggests, and a person acts. An agentic experience acts itself, inside limits someone set, and then tells the user what it did.

This changes what a mistake costs. A wrong chatbot answer can be ignored or corrected. A wrong agent action has already happened in a real system. So the design work moves from conversation quality to control over actions.

How do the four kinds of system compare?

System	What the user does	What the AI does	Typical failure
Fixed automation	Triggers a defined process	Usually nothing, or one narrow step	Stops on an input nobody planned for
Chatbot	Asks a question	Retrieves and writes an answer	Gives a wrong or unhelpful answer
Copilot	Reviews and acts	Drafts or suggests	A person approves a poor suggestion
Agentic experience	States a goal and supervises	Plans, acts and reports	Takes the wrong action, or stops halfway

What does an agentic experience need to work?

It needs six things: a clear goal, tools, task state, limits, a handover route and a record of what happened. The model is only one component. Most of the engineering sits around it, in the connections to other systems and the rules for using them.

What are the six parts of an agentic experience?

A clear goal: the agent confirms what the user wants before acting on anything ambiguous.
Tools: defined operations such as 'look up order' or 'issue refund', each with checked inputs.
Task state: a record of which steps are finished, so a retry does not repeat them.
Limits: permissions, spending caps and step counts enforced in code outside the model.
A handover route: a way to pass the task to a person with the full context attached.
A record: a log of each decision and action that someone can audit later.

Why is safe action harder than good conversation?

Because an action changes something outside the conversation, and a model cannot be relied on to police itself. Language models produce probable output, not guaranteed output. The limits on what an agent can do must sit in ordinary code that the model cannot talk its way around.

Three problems appear the first time an agent runs against real systems. A tool call can time out after the change was made, so a blind retry issues a second refund. A task can stop halfway, leaving an order updated and the customer not told. And text the agent reads, such as an email or a web page, can contain instructions that try to redirect it.

The responses are plain engineering. Give each operation a unique key so it can only apply once. Track task state so a stopped job can resume or roll back. Treat everything the agent reads as data, and give it the narrowest permissions the task allows. An agent with read access only cannot be misled into deleting anything.

What should the user see during an agentic task?

The user should see what the agent understood, what it is doing and what changed. Visible status is what makes delegation tolerable. A silent agent forces the person to check its work by hand, which removes the reason for using it.

Match the amount of checking to the size of the risk. Small, reversible actions can run and report afterwards. Larger ones should show the plan and wait for a 'yes'. Anything involving money, legal commitments or personal data usually deserves explicit approval.

Every task should end with a receipt: what was done, where, and how to undo it if that is possible. The exit matters too. A person who says 'I want to talk to someone' should reach someone, with the context attached and no need to repeat the story.

Where do agentic experiences go wrong with customers?

They go wrong when a customer needs a resolution and the system cannot deliver one. Customer service is the clearest warning. A Qualtrics XM Institute study reports that AI-powered customer service fails at four times the rate of other tasks. It surveyed more than 20,000 consumers across 14 countries in Q3 2025.

That study covers AI in customer service generally, not agentic systems specifically, so it is not a failure rate for agents. It is still a fair warning. Customer service is where people meet AI when they already have a problem. Giving that system the ability to act raises the stakes, because it can now fail while changing a customer's account.

How many businesses use agentic experiences?

The surveys cited here cannot say. They measure AI use broadly, not agents acting in production, so any single adoption number would be a stretch. Treat them as context on who is using AI at all.

The US Census Bureau's Business Trends and Outlook Survey finds that firms with at least 20 employees are the biggest AI users. Its data ran from December 2025 to May 2026. It measures AI use in any business function, not agents specifically, so it cannot separate a firm running agents from one using AI for something much simpler.

The Small Business & Entrepreneurship Council's September 2026 survey of 516 owners of firms with 2 to 99 employees reports use as 'regularly or occasionally'. It excludes solo owners, and the question wording is not published on the page. Occasional use is not the same as a deployed system.

Will customers send their own agents to businesses?

They may, but treat it as a direction of travel, not a settled fact. Visa's 2026 release is titled 'Visa Defines the Next Era of Commerce: When AI Becomes the Customer'. It draws on an online survey conducted for Visa of 2,000 US adults and 512 US business decision makers. A payments company has a stake in that future, so read the release as its view of the market.

The sensible response does not depend on the forecast. Keep product data, prices and stock accurate and easy for software to read. Offer key actions through a documented interface. Both help your own agent, and any agent that arrives from outside.

When is an agentic experience not worth building?

It is not worth building when a form, a search box or a fixed workflow already does the job. If the steps never change, ordinary automation is cheaper to run and easier to test. An agent earns its cost only where the path varies from case to case and a person currently makes the judgement calls.

Skip it in four other cases. The task is rare, so the build never pays back. The data or systems it needs have no usable interface. A mistake would be expensive and impossible to reverse. Or nobody will own the system after launch, review its failures and update its rules.

How should a team build its first agentic experience?

Start with one frequent, bounded task and widen the agent's authority in stages. Do not start with a general assistant. A narrow task gives you clear success criteria, a small set of tools and failures you can actually read.

What are the steps to a first agentic experience?

Pick one task with high volume, clear rules and a reversible outcome.
Collect real past cases, including the awkward ones, and use them as a test set.
Run read-only first, so the agent proposes actions and a person carries them out.
Allow low-risk actions with approval, then without approval once results hold up.
Review the logs regularly and add every new failure to the test set.
Widen the scope only when results on the current task justify it.

Questions people ask

Are agentic experiences just chatbots with a new name?

Sometimes, yes. Plenty of products carry the label and only answer questions. The real test is whether the system can change something in another system, such as updating a booking or issuing a credit, within set limits. If it only replies, it is a chatbot, which may still be the right thing to buy.

Do we need an agent, or would workflow automation do?

Start by assuming automation will do. If the steps are the same every time and the inputs are predictable, a fixed workflow is cheaper to run and easier to test. Choose an agent only when the path depends on judgement, such as reading a messy request and deciding which of several actions applies. Many good systems combine both.

Should an agent ever act without human approval?

Yes, but only for actions that are low risk and easy to reverse, and only after it has proved reliable with approval in place. Updating a delivery note might qualify. Moving money or changing a contract usually should not. Set the boundary per action, enforce it in code, and review it as evidence builds up.

What do our existing systems need before an agent can use them?

They need a dependable way in, usually an API, plus permissions narrow enough to give the agent only what the task requires. They also need data that is accurate enough to act on. If a human has to cross-check three systems because one is out of date, an agent may act on the stale one. Fix that first.

How do we know whether an agentic experience is working?

Measure completed tasks, not conversations. Track how often the agent finishes correctly, how often it hands over, how often a person reverses what it did, and what each finished task costs. Sample real transcripts too, because a task can be marked complete while the customer is still unhappy. A low handover rate alone hides that kind of failure.

Is it safe to let an agent handle customer data?

It can be, if the agent runs under the same access rules as the person it serves. Scope each tool to the current user, log every read and write, and keep secrets out of the model's reach. Decide where data goes, including any model provider, before the build starts. Assume the model can be misled, and limit the damage that could do.