AI · 3 HOURS
Setting Up and Supervising AI Agents
Setting Up and Supervising AI Agents teaches the person who configures an agent in a licensed tool to build it from a brief and test it before anyone relies on it. You finish with a supervision note for one real agent, naming who reviews its work, how often, and who can switch it off.
About this course
Many organisations can now configure an AI agent or assistant inside products they already license, by writing its instructions and connecting it to mailboxes, drives, and business systems. An agent set up with default permissions and tested only on the questions its owner expected can answer from the wrong source, invent details when information is missing, or act on instructions hidden in an email, and it can drift after launch without anyone noticing.
The course takes you through standing up one agent from a written brief: writing standing instructions that hold for inputs you have not seen, connecting only the tools the job needs, and running three recorded tests covering the normal case, the missing input, and the action it must refuse. You finish with a signed supervision note that names the reviewer, the review rhythm, the changes that trigger a retest, and who can switch the agent off, so a colleague could take it over.
What you will take away
- You will be able to write standing instructions covering the role, scope, sources, refusals, hand-over, and format, so the agent behaves correctly on inputs nobody anticipated.
- You will be able to connect only the tools a job needs, check what each connection actually permits, and choose the lowest permission level, using a service account where the product allows.
- You will be able to run and record a normal case test with outcomes written in advance, reading the whole record of each run as well as the final answer.
- You will be able to test whether an agent stops and hands over when information is missing, and whether it refuses both a direct request and an instruction hidden in a document.
- You will leave with a signed supervision note recording the agent's configuration, the three dated tests and their results, and a named reviewer, rhythm, retest triggers, and switch-off.
Who this course is for
- Operations and systems administrators who will configure an agent in a licensed product and remain responsible for how it behaves.
- Process owners who have a written brief for an agent and need to turn it into instructions, settings, and tests without writing code.
- Team leaders who will supervise an agent after launch and need a review arrangement that someone can realistically keep.
- Technically confident specialists who are asked by a manager or auditor to show how an agent in their area is tested and controlled.
What you will do
Each lesson teaches one part of the method, works through a realistic example, and ends with a check on a case you have not seen. The course closes with an assessment and then the piece of work you sign.
- 6lessons with a worked example and a check
- 8scenario questions in the course assessment
- Signedthe supervision note, with a record anyone can verify
- 3 hoursat your own pace, with progress saved
- 01
Lesson
Standing instructions
This lesson explains why standing instructions must hold for inputs you have not seen and sets out the six parts they need, starting with scope, refusals, and hand-over. You edit instructions for an expenses policy assistant so that they include a scope, a source, a refusal, and a hand-over.
- Instructions for inputs you have not seen
- The six parts
- What standing instructions are not
- The mistake people usually make
You repair a flawed draft so it meets the standard the lesson sets, and your revision is checked against it.
- 02
Lesson
Only the tools that job needs
This lesson shows that a connection often permits more than its label suggests and why leaving a tool off is stronger than an instruction about it. You compare two configurations of an equipment catalogue agent and choose the one that gives the job only the tools it needs.
- A connection comes with permissions you did not write
- The rule for this lesson
- Leaving a tool off is stronger than a rule about it
- The mistake people usually make
You compare two versions of the same piece of work, choose the stronger one, and see the reasoning behind the answer.
- 03
Lesson
Test the normal case
This lesson teaches you to test an agent on real, typical inputs with the right outcome written before each run, and to read the whole record rather than only the answer. You mark the results of an expenses assistant's test run as within the job or outside it.
- Real inputs, with the right outcome written first
- Read the whole record
- What this test is not
- The mistake people usually make
You mark 5 statements from a realistic case and receive an explanation for each one.
- 04
Lesson
Test the missing input
This lesson explains why an agent that fills a gap with a confident answer is the most dangerous failure, and how to test its stop condition deliberately. You choose the response to a question outside its source that correctly stops and hands over.
- Give it a gap on purpose
- Two labels for each response
- This is not a test of cleverness
You compare two versions of the same piece of work, choose the stronger one, and see the reasoning behind the answer.
- 05
Lesson
Test the action it must refuse
This lesson covers the two routes by which an agent can be pushed into a forbidden action: a direct request and an instruction hidden in content it reads. You mark each line of an HR policy assistant's refusal test record as refused or acted when it should have refused.
- Two routes to test
- Two labels for each line of the record
- Instructions in content are information, not commands
- The mistake people usually make
You mark 4 statements from a realistic case and receive an explanation for each one.
- 06
Lesson
Supervise after launch
This lesson explains how a tested agent drifts after launch and the four parts of supervision that catch it: a reviewer, a rhythm, retest triggers, and a switch-off. You add the missing parts to an arrangement and choose which of two supervision plans would catch drift.
- Why a tested agent drifts
- The four parts of supervision
- What supervision is not
- The mistake people usually make
You compare two versions of the same piece of work, choose the stronger one, and see the reasoning behind the answer.
- 07
Course assessment
Course assessment
This lesson recaps instructions, tools, the three tests, and supervision, then reads one school leave assistant against every step. You answer eight scenario questions set in facilities, finance, HR, customer service, procurement, IT, and operations, and you need six correct answers to pass.
- Instructions and tools
- The three tests
- Supervision and the assessment
You judge 8 new workplace situations, with feedback on every option, and need 6 correct to pass.
- 08
Final work and signed record
The supervision note
This lesson explains what the supervision note records, why it includes failed tests, and how each line is judged ready to hand over. You write the note for the agent you set up during the course and sign it for your record.
- What the note is for
- What the note is not
- Ready to hand over
- How the note is checked
You write the supervision note for your own work, part by part, and sign it as your record.
GOOD WORK.THE SUPERVISION NOTE, SIGNED.
WHAT CHANGES
What the course changes for you and your organisation.
- 01
Agents tested before anyone relies on them
The three tests check the ordinary work, the gaps, and the requests the agent must refuse before colleagues or customers depend on it. Weaknesses are found and fixed by the person responsible rather than by the first user who meets them.
- 02
Permissions that limit the damage
You learn to leave unnecessary connections off entirely rather than relying on an instruction not to use them. If a hostile email ever persuades the agent to misbehave, the narrow permissions still prevent it from sending, deleting, or revealing what it was never given.
- 03
An agent a colleague can inherit
The supervision note records what was configured, what was tested, and who is watching, including any tests that failed and how they were fixed. When you move on, a colleague can retest the agent, review it, or switch it off from the note alone.
- 04
Named oversight for every agent
Gartner predicts that at least 15% of day to day work decisions will be made autonomously through agentic AI by 2028. When agents act on their own, employers need a named person who tests each agent, reviews its work and can switch it off.
ROLES THIS SKILL SHOWS UP IN
Roles where these skills are already valued and paid for.
These figures are published salary bands and wage premiums, not a guarantee that finishing this course moves you to the top of the range. They are here so you can see what the market is already paying for the work you will practise.
AI Prompt Engineer
£62,750 to £115,000, midpoint £92,500
Robert Half's 2026 UK guide. London midpoint is £125,750.
Artificial Intelligence Engineer
Midpoint £65,750
The neighbouring engineering role in the same Robert Half guide, for people who take the brief into a built system.
Machine Learning Engineer
Midpoint £75,000
Same guide. The brief you write here is the instruction that role then tests.
AI enablement or operations lead
Premium on the role you already hold
PwC found specialist AI skills carried a 34.2% wage premium in 2025. Many of those postings sit inside operations, customer, and people teams rather than a lab.
Customer communications or account work
The same seat, with a skill the posting now names
The course artefact is a prompt card a colleague can run. That is the work those teams are hiring people to supervise.
Learning, knowledge, or product operations
Listed against the AI-skilled premium
These roles now ask for someone who can brief a model, check the output, and leave a reusable instruction. That is the check this course marks.
QUESTIONS
Questions about Setting Up and Supervising AI Agents
- What is Setting Up and Supervising AI Agents about?
- Many organisations can now configure an AI agent or assistant inside products they already license, by writing its instructions and connecting it to mailboxes, drives, and business systems. An agent set up with default permissions and tested only on the questions its owner expected can answer from the wrong source, invent details when information is missing, or act on instructions hidden in an email, and it can drift after launch without anyone noticing. The course takes you through standing up one agent from a written brief: writing standing instructions that hold for inputs you have not seen, connecting only the tools the job needs, and running three recorded tests covering the normal case, the missing input, and the action it must refuse. You finish with a signed supervision note that names the reviewer, the review rhythm, the changes that trigger a retest, and who can switch the agent off, so a colleague could take it over.
- Who is Setting Up and Supervising AI Agents for?
- Operations and systems administrators who will configure an agent in a licensed product and remain responsible for how it behaves. Process owners who have a written brief for an agent and need to turn it into instructions, settings, and tests without writing code. Team leaders who will supervise an agent after launch and need a review arrangement that someone can realistically keep. Technically confident specialists who are asked by a manager or auditor to show how an agent in their area is tested and controlled. No specialist background is assumed, and every term is explained before it is used.
- What will I be able to do after Setting Up and Supervising AI Agents?
- You will be able to write standing instructions covering the role, scope, sources, refusals, hand-over, and format, so the agent behaves correctly on inputs nobody anticipated. You will be able to connect only the tools a job needs, check what each connection actually permits, and choose the lowest permission level, using a service account where the product allows. You will be able to run and record a normal case test with outcomes written in advance, reading the whole record of each run as well as the final answer. You will be able to test whether an agent stops and hands over when information is missing, and whether it refuses both a direct request and an instruction hidden in a document. You will leave with a signed supervision note recording the agent's configuration, the three dated tests and their results, and a named reviewer, rhythm, retest triggers, and switch-off.
- How long does the course take, and how is it taught?
- The course takes about 3 hours and is studied online at your own pace. It has 6 lessons, each with a worked example, a practice exercise, and a check on a new case, followed by a course assessment of 8 scenario questions in which you need 6 correct to pass. Your progress is saved to your account.
- Do I get a certificate?
- Yes. When you pass, you sign a record that names you, the course, and the supervision note. Anyone you share it with can verify it online and download it as a PDF. The record confirms what you completed and does not claim compliance with any regulation.
- How much does it cost, and when can I start?
- The course costs £179, paid once by card through Stripe. Access begins as soon as payment is confirmed, and you can save a sign-in to return to the course from any device.
- Can my organisation train a whole team?
- Yes. Individuals can buy any self-paced course online, and organisations can book trainer-led courses for teams, in person or online, through the Experrt Academy.
More ai courses are listed on the self-paced ai courses for professionals page.
THERE IS A NEXT CHAPTER.
Start today, and finish with
the supervision note your organisation can use.
Checkout takes an email address and a card, and access begins as soon as payment is confirmed. When you finish, you sign a record that names you and the supervision note, which anyone you choose can verify online. The record confirms what you completed and does not claim compliance with any regulation.