Harborline Services is a fictional 240-person business-services company operating in three cities. It supports about 320 client organizations and processes customer requests, supplier invoices, employee questions, recurring reports, and marketing campaigns through a mixture of email, spreadsheets, shared drives, and a service desk.
The company wants practical AI improvements, but its leadership has imposed one rule: no AI output may create a financial, contractual, employment, security, or customer commitment without an accountable human decision.
Company Snapshot#
| Area |
Current state |
Baseline problem |
| Customer operations |
Shared inbox plus service desk |
18% of cases are rerouted; urgent cases are sometimes found late |
| Finance |
Email invoices and spreadsheet trackers |
7% require clarification; duplicate checks are inconsistent |
| Internal knowledge |
Policies across shared drives |
Staff ask repeated questions and cite outdated files |
| Marketing |
Interview notes, CRM exports, and newsletters |
Research is scattered and claims are reviewed late |
| Governance |
Different tools owned by different teams |
Logging, retention, and review rules are inconsistent |
Operating Roles#
- Maya Chen, COO: accountable for operating performance and pilot approval.
- Jon Bell, Support Manager: owns queues, priorities, and service levels.
- Priya Nair, Controller: owns invoice, close, and approval controls.
- Elena Ruiz, Marketing Director: owns external claims and editorial approval.
- Sam Okafor, IT Lead: owns integrations, access, logging, and incident containment.
- Nora Patel, HR Manager: owns policy sources and employee escalations.
Synthetic Intake Records#
| ID |
Channel |
Message |
| H-001 |
Customer email |
“The amount on invoice 7842 does not match our contract. We will pause payment until someone explains it.” |
| H-002 |
Support form |
“I cannot sign in after changing phones. I have a client presentation in two hours.” |
| H-003 |
Vendor email |
“Please update our bank details to the attached account before Friday’s payment run.” |
| H-004 |
Employee chat |
“Can I carry unused leave into next year? The handbook and portal seem different.” |
| H-005 |
Customer email |
“Everything is broken again. Call me.” No account number or service is identified. |
| H-006 |
Internal request |
“Please send the latest onboarding checklist to the new Singapore team.” |
Synthetic Invoice Records#
| ID |
Vendor |
Amount |
Purchase order |
Control issue |
| INV-101 |
Northwind Office |
$4,820 |
PO-5502 |
Clean example |
| INV-102 |
Delta Hosting |
$12,000 |
Missing |
Purchase-order exception |
| INV-103 |
Northwind Office |
$4,820 |
PO-5502 |
Possible duplicate of INV-101 |
| INV-104 |
Brightline Events |
$26,450 |
PO-5529 |
Amount exceeds approval threshold |
| INV-105 |
Apex Advisory |
$8,700 |
PO-5540 |
Bank-detail change requested by email |
Approved Knowledge Sources#
| Source |
Owner |
Audience |
Review status |
| Leave and Attendance Policy v4.2 |
HR Manager |
All staff |
Current |
| Travel and Expense Policy v3.1 |
Controller |
All staff |
Current |
| Client Support Escalation Standard v2.5 |
Support Manager |
Support team |
Current |
| Information Handling Standard v1.8 |
IT Lead |
All staff |
Current |
| Onboarding Checklist v5.0 |
HR Manager |
Managers and HR |
Current |
| Leave FAQ v2.0 |
Former HR analyst |
All staff |
Stale; conflicts with v4.2 policy |
Customer Feedback Samples#
- “The team is helpful, but I have to repeat the account history every time.”
- “Reports arrive on time, although the commentary often tells us what the chart already shows.”
- “Your onboarding was smooth. The security questionnaire took too long.”
- “We like the service, but pricing changes are hard to understand.”
- “Support is fast for normal requests and slow when several teams are involved.”
Governance Constraints#
- Public AI services may not receive bank details, employee records, credentials, or unredacted contracts without an approved enterprise arrangement.
- High-risk outputs require source evidence and a named reviewer.
- The IT Lead may pause an AI workflow during an incident.
- Finance approvals and payment release remain in the accounting system.
- Customer-facing text must be approved by the owning team until a pilot proves a lower-risk path.
- Logs retain workflow IDs and decision evidence, but must avoid unnecessary sensitive content.
Baseline Metrics#
| Metric |
Baseline |
| First-pass support routing accuracy |
82% |
| Median intake-to-queue time |
24 minutes |
| Manual invoice touch time |
11 minutes |
| Knowledge-search time per repeated question |
8 minutes |
| Newsletter production cycle |
2.5 working days |
| High-risk cases missing required review |
6% in the last sample |
Suggested Cross-Module Capstone#
Design one governed AI program for Harborline that:
- improves support intake;
- retrieves approved policy knowledge;
- extracts invoice fields without approving payment;
- creates a human review queue;
- defines evaluation, monitoring, and incident response;
- reports business value against the baseline metrics.
The final capstone should contain a workflow map, output contracts, a fixed test set, a control matrix, a measurement plan, and a go/revise/stop recommendation.