From assistant to agent
How is Agentic AI changing the way we work? A practical guide to moving from a tool that answers questions to a digital agent that carries out complete tasks, while review and the final decision stay with a person.
You will design a clear agent brief and a first prototype of an agent that helps you complete a real task at work, and you will set its permissions, limits and the point where a person steps in.
Understand Agentic AI, find the tasks worth delegating, and use the tools in a practical and safe way at work.
Administrative staff, people working in quality, human resources, patient experience, training and procurement, and members of operational and non-technical teams.
How the guide works: you try hospital tasks in ChatGPT Work or Claude Cowork, with no coding and no complex technical detail.
From ChatGPT to Agentic AI
It starts with a simple question: do we want an answer from AI, help with part of the work, or a complete result?
Chatbot
Answers a direct question with a piece of information or a short text.
AI Assistant
Helps with one specific part of a task, such as summarising or rewording.
AI Agent
Plans the task and carries out its connected steps to reach a defined result.
The usual way: summarise this file.
Clear delegation: analyse the files, compare the results, pull out the problems, then prepare the required outputs.
The fixed principle: the agent does the suitable work, and a person reviews the result and makes the final decision.
Think of the agent as a new employee
A vague instruction is not enough for a new employee. They need a brief that sets out what is required, the limits and how to hand over the result. The same applies to an agent.
What result do we want?
Which files, data and references does it rely on?
What order should it follow?
What can it read, use or do?
What must it not do or use?
When does the agent stop and ask for review?
What form: a report, a table, a presentation or an action plan?
Is the output complete, clear, reviewable and fit for use?
Goal, sources, steps, permissions, limits, approvals and output format. These elements take us from writing a short question to preparing a clear, complete brief.
A quick example of a completed brief
Analyse patient experience survey results and compare them with the previous period's results.
Current and previous result files, using dummy or training data only.
An executive summary, an actions and owners table, a management report, and a short presentation.
A hands-on demo with ChatGPT Work
In this example you use dummy or training data only to try a multi-step task, and you see the difference between summarising one file and delegating a whole task.
Where to find it: in ChatGPT, choose Work instead of Chat. Work is available on the web and mobile for eligible paid plans, and in the ChatGPT desktop app when your plan includes it.
Analyse the results
Read the patient experience survey results and organise the key information.
Compare
Compare the results with the previous period's results and show what changed.
Find patterns
Identify recurring problems and order them so they are easy to review.
Prepare the outputs
Turn the analysis into outputs management can use at the next meeting.
The outputs the agent prepares
Executive summary
A quick read of the main findings and priorities.
Actions and owners table
Each action, who owns it and its follow-up status.
Management report
An organised draft ready for human review.
Short presentation
Focused points for the next meeting.
Demo rule: never use real or sensitive patient data.
Agentic AI uses in a hospital
Examples close to daily work that help you spot tasks you can organise or speed up while keeping human review.
Meeting minutes
Organise the minutes and track decisions and actions.
Staff onboarding and training
Prepare structured programmes and supporting training material.
Patient experience
Analyse feedback and surveys and find patterns.
Quality reports
Prepare organised drafts before review and sign-off.
Supplier offers
Compare criteria, prices and notes in one table.
Awareness materials
Draft clear content suited to the target audience.
Research and sources
Summarise and organise research and trusted sources.
Action plans and follow-up
Turn goals into tasks, owners and follow-up schedules.
Teaching AI how we work
Knowing the task is not enough. Give the agent material that explains how your organisation works, so its output meets your requirements.
Approved report templates
So it keeps the required structure.
Policies and procedures
So it knows the reference points and limits.
House writing style
So it writes in the right tone and terms.
Checklists
So it checks the work is complete before handing it over.
Workflow steps
So it carries out the task in the right order.
Review and approval
So it stops at the points that need a human decision.
Where to keep it: in Claude, save this material as Skills or inside a Project so Claude uses it in every related task. Cowork capabilities are available on paid plans (Pro, Max, Ultra, Team, Enterprise), and you can use them from any conversation in the new Claude experience. In ChatGPT Work, start the task from inside a Project so Work uses that project's context.
From one agent to a team of agents
Complex work can be split into clear roles, and then a person reviews and signs off the result.
Research agent
Gathers the sources needed.
Analysis agent
Organises the information and finds patterns.
Writing agent
Turns the findings into clear content.
Review agent
Checks the results are correct and complete.
The person
Reviews and signs off the final output.
Privacy, safety and human review
We classify data and uses with a traffic-light model. The colour shows how much caution is needed, not how good the task is.
Green
Public content, early ideas, templates, training data and non-sensitive information.
Yellow
Internal reports, operational information, quality analyses, and work that needs review and approval.
Red
Data that identifies patients, medical records, sensitive health information, login credentials, and any data your organisation's policy forbids.
Decision rule: use the agent when the use is appropriate, ask for approval when policy requires it, and do not use it when the data or the task is prohibited.
Set the controls in the tool itself: in Claude, choose the Manual permission mode so it asks before it takes actions, then choose Allow or Deny. Set sensitive connector tools to "Always ask". Claude also asks for explicit permission before permanently deleting files. In ChatGPT Work, review each action that needs approval, then approve or decline it with the on-screen controls. In a company workspace, admins decide whether Work and Cowork are turned on.
Design your first AI agent
Pick a recurring task from your work, then fill in the eight boxes below to design an agent you can try.
What you will be able to do after this guide
Sources: OpenAI Help Center (ChatGPT Work and Codex), Claude Help Center (Get started with Claude Cowork, Using Skills in Claude, Claude Cowork and chat are one Claude). Checked November 2026.
Want your team to use this at work?
I'm Abdulaziz Alshamsi, a certified AI trainer and consultant in Abu Dhabi. I train government and company teams to use AI in their daily work. Message me to discuss a session for your team.