What is a Level 3 AI agent?
A Level 3 AI agent takes a goal, plans the work, and carries it out without you managing every step. You give it business context and feedback as it works.[1]
This article uses the framework by Feng, McDonald, and Zhang, published by Columbia's Knight First Amendment Institute. They call Level 3 "user as a consultant." It is a proposed framework, not a universal standard.[1] Google DeepMind researchers use Level 3 to mean shared human and AI collaboration.[4]
The agent leads. Your team guides it.
In this model, the agent takes the lead on planning and doing the work. Your team answers questions, explains company needs, and asks for changes. You stay involved without walking it through each step.[1]
That is more than just signing off on a finished task. The authors put "user as an approver" at Level 4. At Level 3, your input helps shape the work while it is in progress.[1]
A simple business example
Say you're managing shipments in your business. You ask an agent to look into late deliveries.
In this hypothetical example, you give it access to the records it needs. It plans the review, reads shipping records and support tickets, and looks for patterns.
It finds that many delays come from one warehouse. Before drawing a firm conclusion, it asks whether there is a warehouse limit it should know about.
You explain that the team changed its pickup times last month. The agent uses that context, checks the dates, and continues toward a report. Your team reviews the findings and decides what to change.
The same approach could apply to missed invoice payments or project delays. The task changes, but the pattern stays the same: the agent leads the review, and your team adds judgment.
An answer, a workflow, and an agent
Asking AI to explain one support ticket is a request for an answer. You still handle the next steps.
A fixed workflow follows a set path. An agent can choose its next steps and tools based on what it finds. This is the distinction Anthropic uses in its guidance on building agents.[2]
Autonomous AI agents still need limits. Choosing how to review records is different from having permission to change them.
For AI workflow automation, choose the simplest setup that fits the task. A fixed workflow may be enough when the steps rarely change.[2]
Set the rules before it starts
For the delivery review, start with access to read records, not change them. Keep refunds, customer messages, and shipping changes outside that task.
Write down the controls:
- Which records and tools can it use?
- When should it ask your team for context?
- Which actions need approval, and from whom?
- When must it stop and ask for help?
- Where can your team check its work and undo changes?
Test those rules before using the agent on live work. A Level 3 label does not prove that an agent will follow them.
Set up its instructions and stored company knowledge, too. Keep approved policies, terms, and past corrections where it can use them. Memory needs to be built and checked. It is not a promise that the agent will remember every detail or get every answer right.
Start with one useful task
For agentic AI for business, start with a task your team already knows well. Check the output against the records. Compare the time spent, errors, and work your team still has to do before expanding its scope.
Core Agentic builds custom AI agents and software around selected team workflows, with clear controls and human input.[3] The goal is useful help inside your company, not handing over the business.