Leila Fielding
WHAT YOU'LL LEAVE WITH
- A readiness verdict, with the gaps to close before anything is built.
- A one-page agent outline: what it does, what it can reach and where a person stays in charge.
- A test plan covering the ways agents go wrong, not just the happy path.
- A list of questions for your IT, security and data teams, saved together as one PDF pack.
BEFORE YOU START
- An idea for something an AI agent could do, even a rough one.
- A rough idea of the process it would be part of.
No technical knowledge needed. Plain descriptions are enough; don't enter system names, passwords or security details.
Step 1 of 6
The more an agent can do on its own, the more has to be in place first. Start by being clear about what it would actually do.
FROM THE FIELD
AI Snapshot #002: Standardise processes now.
Mapping the steps shows exactly where the agent would act, and where a person has to stay involved.
| Step | Who does it | Row actions |
|---|---|---|
3 steps · the agent acts on its own in 0 · 0 exceptions listed
FROM THE FIELD
An agent is only as good as the information it reads, and only as safe as what it's allowed to reach.
Describe things in plain words. Don't enter system names, passwords or security details.
What will it read?
| Information | Kept up to date? | Owner | Row actions |
|---|
What will it be able to change?
If it will only read, add a row called “Nothing; read-only” and choose “Read only”. This makes the boundary explicit.
| What | Access | Row actions |
|---|
0 Not sure · 0 No on up-to-date · 0 Can change it
FROM THE FIELD
RAG Architecture Planner is for agents that answer from documents.
In plain EnglishGuardrails are rules, technical controls and safeguards designed to keep an AI system operating within acceptable boundaries. They can restrict what information the system accesses, which actions it performs and how certain requests are handled.Practical takeawayFor AI systems that take actions, enforce permissions in the connected systems rather than relying solely on written instructions to the AI.Read the full entry decided now are much easier to agree than limits added after something goes wrong.
A good demo isn't evidence. Test the required safety scenarios with specific examples, then record what each test actually did.
| Test | Example for this agent | Planned? | Result | Row actions |
|---|---|---|---|---|
| Required test | ||||
| Required test | ||||
| Required test | ||||
| Required test | ||||
| Required test | ||||
| Required test | ||||
| Required test | ||||
| Required test | ||||
0 tests planned · 0 passed
Mark a condition verified only after you have confirmed it is true.
Verify every go-live condition before calling the pilot ready; 3 remain unconfirmed.
FROM THE FIELD
Use the AI Agent Testing Checklist to plan a fuller test set.
In plain EnglishAI readiness describes how prepared an organisation is to introduce, use and manage artificial intelligence effectively. It includes having suitable processes, reliable data, appropriate technology, the right skills and clear responsibilities.Practical takeawayAssess the process, data, people, governance and expected outcomes before investing in AI technology.Read the full entry
NOT YET
Something fundamental is missing. Fix the items below before any build starts.
NOT YET if the idea fails the starting check, has no named owner, has no switch-off owner, changes information without an approval, or the access-limit answer is No.
CLOSE: FIX THE GAPS if any required process, information, access, limits, test-result or go-live check is incomplete. Every planned test needs a specific example and a passing result; every listed go-live condition must be confirmed.
READY FOR A SUPERVISED PILOT only when every listed requirement is explicitly met.
GAPS TO CLOSE
- Name someone who can switch it off.
- Complete the idea check and confirm this is suitable to assess.
- Complete the idea, intended user and agent capability.
- List at least one process exception.
- Confirm that you have checked the process map with the people who do the work.
- List at least one information source.
- Describe what the agent can change, or explicitly list a read-only boundary.
- Confirm that the agent can only see information the user is allowed to see.
- Answer whether content the agent reads could contain instructions that try to steer it.
- Involve the security or IT team before a build starts.
- Choose the agent's autonomy level.
- Plan at least eight tests; 0 are planned so far.
- Pass at least eight tests with specific examples; 0 qualify so far.
- Include and plan a test for missing information.
- Include and plan a test for conflicting information.
- Include and plan a test for access limits.
- Include and plan a test for actions it should not take.
- Include and plan a test for instructions hidden in content.
- Include and plan a test for a tool or system being unavailable.
- Include and plan a test for sensitive information.
- Include and plan a test for handing over to a person.
- Verify every go-live condition before calling the pilot ready; 3 remain unconfirmed.
QUESTIONS FOR YOUR IT, SECURITY AND DATA TEAMS
FROM THE FIELD
YOUR READINESS PACK
Your readiness report so far
Save it as a PDF and share it with the agent's owner and your IT, security and data teams.
1. NEXT STEPS
- Close each gap below.
- Take the questions to your IT, security and data teams.
- Check again when the gaps are closed.
2. VERDICT AND GAPS
NOT YET
- Name someone who can switch it off.
- Complete the idea check and confirm this is suitable to assess.
- Complete the idea, intended user and agent capability.
- List at least one process exception.
- Confirm that you have checked the process map with the people who do the work.
- List at least one information source.
- Describe what the agent can change, or explicitly list a read-only boundary.
- Confirm that the agent can only see information the user is allowed to see.
- Answer whether content the agent reads could contain instructions that try to steer it.
- Involve the security or IT team before a build starts.
- Choose the agent's autonomy level.
- Plan at least eight tests; 0 are planned so far.
- Pass at least eight tests with specific examples; 0 qualify so far.
- Include and plan a test for missing information.
- Include and plan a test for conflicting information.
- Include and plan a test for access limits.
- Include and plan a test for actions it should not take.
- Include and plan a test for instructions hidden in content.
- Include and plan a test for a tool or system being unavailable.
- Include and plan a test for sensitive information.
- Include and plan a test for handing over to a person.
- Verify every go-live condition before calling the pilot ready; 3 remain unconfirmed.
VERDICT: NOT YET GAPS TO CLOSE: - Name someone who can switch it off. - Complete the idea check and confirm this is suitable to assess. - Complete the idea, intended user and agent capability. - List at least one process exception. - Confirm that you have checked the process map with the people who do the work. - List at least one information source. - Describe what the agent can change, or explicitly list a read-only boundary. - Confirm that the agent can only see information the user is allowed to see. - Answer whether content the agent reads could contain instructions that try to steer it. - Involve the security or IT team before a build starts. - Choose the agent's autonomy level. - Plan at least eight tests; 0 are planned so far. - Pass at least eight tests with specific examples; 0 qualify so far. - Include and plan a test for missing information. - Include and plan a test for conflicting information. - Include and plan a test for access limits. - Include and plan a test for actions it should not take. - Include and plan a test for instructions hidden in content. - Include and plan a test for a tool or system being unavailable. - Include and plan a test for sensitive information. - Include and plan a test for handing over to a person. - Verify every go-live condition before calling the pilot ready; 3 remain unconfirmed. GO-LIVE CONDITIONS: - [NOT VERIFIED] Every access-limit test passes - [NOT VERIFIED] No action is taken without the approvals in Step 04 - [NOT VERIFIED] The owner has signed off the test results DECISION: [not answered]
3. AGENT OUTLINE
What it does: [not answered] Who uses it: [not answered] What it does for them: [not answered] Process checked with the people who do the work: No What it reads: - [no sources listed] What it can change: - [nothing listed] Autonomy: [not answered] Approvals: - [none listed] What it must never do: - Change records it hasn't been asked to change - Share one person's information with another - Make a decision about an individual person When it hands over: - When it isn't sure - When the request is unusual or sensitive - When someone asks for a person Who can switch it off: [not answered]
4. PROCESS MAP
| Step | Who does it |
|---|---|
| Request comes in | A person |
| Look up the information | The agent drafts, a person checks |
| Reply is sent | A person |
Exceptions
None listed.
5. TEST PLAN
| Test | Example for this agent | Planned? | Result |
|---|---|---|---|
| Normal requests | Not entered | No | Not run |
| Missing information | Not entered | No | Not run |
| Conflicting information | Not entered | No | Not run |
| Access limits | Not entered | No | Not run |
| Actions it shouldn't take | Not entered | No | Not run |
| Instructions hidden in content | Not entered | No | Not run |
| A tool or system is down | Not entered | No | Not run |
| Sensitive information | Not entered | No | Not run |
| Hand-over to a person | Not entered | No | Not run |
| The same test, repeated | Not entered | No | Not run |
6. GO-LIVE CONDITIONS
- Not verified: Every access-limit test passes
- Not verified: No action is taken without the approvals in Step 04
- Not verified: The owner has signed off the test results
7. QUESTIONS FOR IT, SECURITY AND DATA
- Can its access be limited to [first source] only, for the person using it?
- How would we switch it off quickly?
- What would we log, and who would review it?
- How do we stop content it reads from changing its instructions?
Answers are stored in this browser when local storage is available; they are not sent to us. Save the PDF to keep a copy.
AI agent readiness report
VERDICT AND GAPS
NOT YET
- Name someone who can switch it off.
- Complete the idea check and confirm this is suitable to assess.
- Complete the idea, intended user and agent capability.
- List at least one process exception.
- Confirm that you have checked the process map with the people who do the work.
- List at least one information source.
- Describe what the agent can change, or explicitly list a read-only boundary.
- Confirm that the agent can only see information the user is allowed to see.
- Answer whether content the agent reads could contain instructions that try to steer it.
- Involve the security or IT team before a build starts.
- Choose the agent's autonomy level.
- Plan at least eight tests; 0 are planned so far.
- Pass at least eight tests with specific examples; 0 qualify so far.
- Include and plan a test for missing information.
- Include and plan a test for conflicting information.
- Include and plan a test for access limits.
- Include and plan a test for actions it should not take.
- Include and plan a test for instructions hidden in content.
- Include and plan a test for a tool or system being unavailable.
- Include and plan a test for sensitive information.
- Include and plan a test for handing over to a person.
- Verify every go-live condition before calling the pilot ready; 3 remain unconfirmed.
1. NEXT STEPS
- Close each gap below.
- Take the questions to your IT, security and data teams.
- Check again when the gaps are closed.
2. AGENT OUTLINE
What it does: [not answered] Who uses it: [not answered] What it does for them: [not answered] Process checked with the people who do the work: No What it reads: - [no sources listed] What it can change: - [nothing listed] Autonomy: [not answered] Approvals: - [none listed] What it must never do: - Change records it hasn't been asked to change - Share one person's information with another - Make a decision about an individual person When it hands over: - When it isn't sure - When the request is unusual or sensitive - When someone asks for a person Who can switch it off: [not answered]
3. PROCESS MAP
| Step | Who does it |
|---|---|
| Request comes in | A person |
| Look up the information | The agent drafts, a person checks |
| Reply is sent | A person |
Exceptions: None listed.
4. TEST PLAN
| Test | Example for this agent | Planned? | Result |
|---|---|---|---|
| Normal requests | Not entered | No | Not run |
| Missing information | Not entered | No | Not run |
| Conflicting information | Not entered | No | Not run |
| Access limits | Not entered | No | Not run |
| Actions it shouldn't take | Not entered | No | Not run |
| Instructions hidden in content | Not entered | No | Not run |
| A tool or system is down | Not entered | No | Not run |
| Sensitive information | Not entered | No | Not run |
| Hand-over to a person | Not entered | No | Not run |
| The same test, repeated | Not entered | No | Not run |
5. GO-LIVE CONDITIONS
- Not verified: Every access-limit test passes
- Not verified: No action is taken without the approvals in Step 04
- Not verified: The owner has signed off the test results
6. QUESTIONS FOR IT, SECURITY AND DATA
- Can its access be limited to [first source] only, for the person using it?
- How would we switch it off quickly?
- What would we log, and who would review it?
- How do we stop content it reads from changing its instructions?
BACK TO ALL PLAYBOOKS