Case 36
When the Assistant Needs to Act: BLUEPRINT Readiness
Scenario
Your operations team wants an AI that not only drafts responses but also updates case records and sends emails. That moves it from answering to acting.
Why this framework
Uses PIN: what it's for, the context and source of truth, and a narrow ask. Enough for most everyday work.
Example context
Proposed agent description: The Case Assistant would read incoming complaint emails and the case record in our case management system, draft a reply, update the case status and priority, and send the reply to the customer after a short delay. It would run on weekdays between 8am and 6pm for the Customer Resolutions team of 22 people. The system holds names, addresses, account numbers and complaint history. The vendor says the agent can also issue goodwill credits up to £50. Current process: An advisor reads the email, searches the record, writes a reply, a team leader checks about one in ten replies, and the advisor updates the status by hand. Goodwill credits over £20 need team leader approval. Complaints involving vulnerable customers or legal threats go to a senior caseworker. Replies must go out within two working days.
Prompt
Purpose: Help me list the decisions we must make before letting an AI act in our systems, so we can work through them with BLUEPRINT.
Context: Use the attached description of the proposed agent and our current process.
Narrow the ask: A checklist of open questions: what it is for, what it may read and change, what it must never do, who checks its work and when, and how we stop or roll it back. Mark which ones we can already answer from the description.
Materials (pasted here in place of the attachments):
Proposed agent description: The Case Assistant would read incoming complaint emails and the case record in our case management system, draft a reply, update the case status and priority, and send the reply to the customer after a short delay. It would run on weekdays between 8am and 6pm for the Customer Resolutions team of 22 people. The system holds names, addresses, account numbers and complaint history. The vendor says the agent can also issue goodwill credits up to £50.
Current process: An advisor reads the email, searches the record, writes a reply, a team leader checks about one in ten replies, and the advisor updates the status by hand. Goodwill credits over £20 need team leader approval. Complaints involving vulnerable customers or legal threats go to a senior caseworker. Replies must go out within two working days.Example output from ChatGPT
Case Assistant - Pre-Launch Decision Checklist
Decision required - Not ready for autonomous operation
Assessment: The Case Assistant has a defined purpose, operating window and proposed capabilities. But its authority to change records, contact customers and issue credits is not yet sufficiently controlled.
The checklist below follows BLUEPRINT v2.1, covering mandate, authority, operation and assurance.
1. Purpose and scope
Use Case & Outcome
Is the job defined?
Process complaint emails, draft replies and update cases for Customer Resolutions.
Known from materials
Who uses it, and when?
22-person team; weekdays, 8am-6pm.
Known from materials
What does successful completion mean?
Define accuracy, quality, turnaround and acceptable error rates.
Open decision
Does this need an autonomous agent?
Decide whether a fixed workflow with AI drafting is sufficient.
Open decision
2. Reading data and changing records
Permissions & Identity / Trusted Context
What data is available?
Complaint emails, case records, names, addresses, account numbers and complaint history.
Known from materials
Which records may it read?
Define assigned-case access, minimum necessary fields and data protection controls.
Critical decision
Which fields may it change?
Specify status and priority values, allowed transitions and prohibited changes.
Critical decision
What identity and access will it use?
Dedicated agent identity, least-privilege permissions and audit logging.
Open decision
How will it handle conflicting or stale information?
Define authoritative sources and when it must stop.
Open decision
3. Prohibited actions and authority
Limits & Autonomy / Behaviour
Which complaints require specialist handling?
Vulnerable customers and legal threats go to a senior caseworker.
Known from materials
May it send replies without approval?
Decide whether all outgoing messages require review or only approved low-risk types.
Critical decision
May it issue goodwill credits?
Vendor supports up to £50, but the existing process requires team leader approval above £20.
Critical decision
What must it never do?
Explicitly define prohibited actions, such as bypassing approvals, closing unresolved cases or making unauthorised promises.
Critical decision
What happens when a request is outside its authority?
Stop, preserve the case and route it to the appropriate person.
Open decision
4. Human review and escalation
Interfaces & Handoffs / Evaluation
What is today's checking process?
Team leader reviews approximately 10% of replies; advisors handle remaining replies.
Known from materials
Who approves AI-generated replies?
Assign named reviewers and decide whether approval is required before sending.
Critical decision
How are sensitive complaints detected and routed?
Define detection, referral to senior caseworkers and how uncertain classifications are handled.
Critical decision
What happens during the send delay?
Define its length, review opportunity, cancellation mechanism and expiry.
Open decision
How will review quality be measured?
Decide sampling rates, error thresholds, escalation and accountability.
Open decision
5. Operations, stopping and recovery
Navigation / Reliability & Recovery
What is the existing response deadline?
Replies must go out within two working days.
Known from materials
How can we stop the agent immediately?
Name the kill-switch owner; disable sending, credits and record changes.
Critical decision
How will incorrect actions be corrected?
Define status reversal, credit recovery and customer correction procedures.
Critical decision
What happens when systems fail or time runs out?
Queue safely, avoid duplicate actions, notify advisors and preserve deadlines.
Open decision
What evidence must be retained?
Record sources, proposed actions, approvals, tool results and intervention history.
Open decision
How will rollout and rollback be tested?
Set pilot stages, failure tests, success thresholds and a return-to-manual process.
Open decision
Decisions that should block launch
| Decision | Current position | Recommended interim control |
|---|---|---|
| Customer communication | Auto-send proposed | Human approval before sending |
| Goodwill credits | Vendor offers £50 | No autonomous credits; retain existing approval process |
| Case changes | Automatic updates proposed | Allow proposed changes, require confirmation |
| Vulnerable/legal cases | Senior escalation required | Mandatory human handoff |
| Shutdown and rollback | Not specified | No production access until tested |
These are recommended safeguards, not claims that the current policy already requires all of them.
Suggested rollout path
Phase 1
Draft only
AI reads authorised cases and drafts responses. Advisors approve and perform all actions.
Phase 2
Controlled updates
Permit selected status or priority changes with approval, logging and recovery tests.
Phase 3
Limited autonomy
Consider autonomous low-risk actions only after testing demonstrates acceptable performance and controls are approved.
Core recommendation: Start by treating Case Assistant as a human-supervised workflow rather than granting it full autonomy. The most important distinction is between generating a recommendation and having authority to execute it.
Before rollout, Customer Resolutions, Compliance, Information Security and the accountable business owner should review and approve the BLUEPRINT specification, including the test evidence and rollback procedure.
