Agentic AI

Agentic AI means a system can look at what’s happening, choose an allowed next action, use a tool, and adjust after it sees the result. The useful question for your firm is what decision it can make without a person making it first.

In the 150+ franchise units in the delivery data I review, roughly 80% still use chatbots, 15% use trigger-based workflows, and about 5% use agents started by conditions. A chatbot waits for your request, a fixed workflow follows a path somebody wrote, and an agentic system gets to choose part of that path.

That doesn’t tell you how many agents are involved or whether they’re ready for client work. You still have to decide what the system may do, when it must stop, and who takes over when something goes wrong.

What makes an AI system agentic?

An agentic system directs part of its own process while working toward a defined outcome. It reads the current state, chooses an action, acts, checks the result, and chooses again until the work finishes or reaches a boundary.

Anthropic’s guide to effective agents makes this distinction: an AI workflow follows paths defined in code, while an agent chooses how to proceed and which tools to use. Depending on what it finds, it may search the firm’s records, look something up in the CRM, draft a document, or ask a person to decide.

One agent can do this, a model choosing among four tools is already making agentic decisions. Adding more agents changes the design and creates more handoffs, but the number of agents isn’t what defines the behavior.

The practical test is what the system chooses after it sees the current state. “Draft this client email” is one model call. “Watch for a stalled engagement, inspect the account history, decide whether a client touchpoint is due, prepare the right next action, and route a risk to the supervisor” contains model-led choices.

How is agentic AI different from an agent and a workflow?

Agentic AI describes the behavior, and an AI agent is the software pursuing the goal. The workflow is the path the work follows, while an AI employee is the role your firm defines and gives to software.

Think about a follow-up email. With a chatbot, a manager asks for the email, while a fixed workflow might collect the notes, draft it, send it for review, and file the result in a set order.

An agentic system gets to decide which stalled account needs attention and which allowed action fits. An AI employee is the role you give that software, such as watching client health, with its own starting conditions and a result the firm can check.

A fixed workflow may contain one agentic decision. An AI employee may run several workflows and use an agent for the parts that cannot be mapped in advance. Orchestration coordinates those parts. It tracks state, moves context across handoffs, handles retries, and knows when the run should stop.

A product label does not reveal this operating design. A chatbot sold as agentic still waits for a person. A multi-agent demo can remain a Stage 1 tool if staff carry every result into the real workflow by hand.

What does agentic AI look like in a business?

Take client health. A manager normally reviews tasks, messages, meeting notes, and overdue commitments to decide which account needs attention. The costly work starts before the email. Somebody has to notice the pattern and gather the facts.

In an agentic version, you might set a condition of four days without contact with the client. The system reads the approved account history and checks whether a meeting or delivery is already scheduled, then chooses an allowed next action.

It might draft a check-in or create a task, or it might ask the supervisor to decide because the account shows signs that the client could leave.

The run should end with a record in the system where you manage delivery. You set the job, ongoing access, starting condition, written rules, and way to check the result, and the model makes choices inside those limits.

That’s what lets the system complete a piece of work rather than leave you with more material to sort.

An agent working from pasted context can still help, but the person has to notice the problem, collect the information, start the run, and carry the answer into the next system. The agent is making choices inside assisted work, rather than running an autonomous production role.

When is a fixed workflow better?

Use a fixed workflow when the path can be drawn before the run starts. It will usually be cheaper to test, easier to trace, and more predictable when something fails.

A proposal process often has a stable backbone. Collect the approved transcript, match the CRM record, extract requirements, draft, review price, and file the final version. The model can write inside that path without deciding what the whole process should do next.

Choosing the next step with a model is worth its cost when the action depends on evidence discovered during the run. A support request may need a refund check, a policy search, a technical investigation, or a handoff. Writing every branch as ordinary rules can become harder to maintain than giving one bounded agent the tools and decision standard.

Anthropic recommends starting with the simplest design that works because agentic systems add cost and delay in exchange for flexibility. Fewer model-led choices can be the better design when you don’t need that flexibility.

With a fixed workflow, the team knows which path to investigate when it fails. An agent may reach the same wrong result through several paths, so finding out what happened becomes part of the cost.

What controls does agentic AI need in production?

Start with what the system may do, give it its own identity, only the access the job needs, and a record of its actions.

In February 2026, NIST raised agent identity and authorization as a production concern because agents can reach data, tools, and applications. Its concept paper calls for identifying agents, checking permissions, auditing actions, keeping evidence of who did what, and controls against instructions hidden in untrusted material, known as prompt injection.

Write the boundaries for each run:

• what starts the work;
• which sources and tools it can reach;
• the decision it may make;
• which action needs approval;
• what proves the work finished;
• how many times it may retry;
• who receives the work when it stops.

There isn’t much public safety evidence to work from. The 2025 AI Agent Index studied 30 prominent agents, 24 of which launched or received major agentic updates in 2024 or 2025.

Of the 13 agents in its frontier-autonomy group, only 4 disclosed any agentic safety evaluation. Twenty-five of the full 30 disclosed no internal safety results, and 23 had no third-party testing.

Those are findings about the public testing record, not a failure rate for your work. A demo still leaves you with questions you need to test inside your own firm.

How people supervise agents also changes with use. Anthropic’s February 2026 study of millions of agent interactions found full auto-approval in roughly 20% of sessions for new Claude Code users and more than 40% for experienced users.

Experienced users also interrupted the agent more often, even as they approved fewer actions one by one. In that study, less step-by-step approval didn’t mean people had stopped watching, it came with more intervention.

Keep a person where judgment or a hard-to-reverse action begins, use rules to enforce permissions, and use an eval to check the completed output against a written standard. Keep the actions and evidence so the owner can see what failed and resume without repeating finished work.

How should a business test an agentic system?

Before the first production run, write down the goal, allowed tools, expected result, stop rule, and how you’ll recover failed work. That gives you something to test beyond whether the demo looks good.

Build the tests from real work, including a clean case, missing information, conflicting client records, a tool timeout after a write, an instruction hidden inside an untrusted document, and a task the agent should refuse. Check the final business record, because a polished message attached to the wrong client is still a failed run.

Then increase what it can do a little at a time, starting with read access and a draft, and adding a reversible write after it passes representative cases. Keep approval before money moves, a client receives a promise, permissions change, or a record disappears.

Use one test for autonomy: can the firm check the result and undo the action? Model capability does not settle either question. Task design does.

Where does agentic AI sit on the Delivery Model Ladder?

Agentic AI belongs at Stage 2, Augmented, on the Delivery Model Ladder when a bounded agent starts from a defined condition, uses approved company systems, completes a named part of delivery, records the result, and routes judgment or failure to a person.

At Stage 1, a person starts the agent, watches the work, and carries the result into the next step. The model may make several choices, but the workflow still depends on a person to hold it together.

Agentic architecture does not create Stage 3. AI-native is a firm-level state where delivery and revenue stop tracking headcount, supported by shared context and work the firm can verify. A complex group of agents can remain at Stage 1. A plain run with one agent, limited tools, one approval, and reliable recovery can be solid Stage 2.

I’d use agentic choices where writing every path in advance becomes impractical, and only where the firm can test, monitor, stop, and recover the work. If a fixed workflow can do the job, start there.