Complete Field Examples
Illustrative teaching artifacts, from the first-week company brief to the month-12 board story. Northstar Field Services and the figures below are presented as worked examples, not a verified customer case study. Use them to understand the format; replace every assumption and result with your own evidence.
Inherited Company Brief
Purpose: give the founder, operator, or advisor a single page that captures what was inherited before prescribing AI. This prevents tool-first theater.
Company Snapshot
| Company | Northstar Field Services |
|---|---|
| Size | 78 employees, 3 locations, $18.4M annual revenue |
| Operating Model | Field technicians, dispatcher team, sales reps, service managers, finance/admin office |
| Systems | QuickBooks, ServiceTitan, Microsoft 365, SharePoint, RingCentral, Excel reporting packs |
| Commander Hypothesis | The company isn't short on effort. It's leaking time through status chasing, duplicate data entry, slow quoting, and unclear exception handling. |
Top Outcomes
- Reduce quote-to-dispatch cycle time from 3.8 days to under 1.5 days.
- Cut manual reporting effort by 50% within 90 days.
- Improve first-response speed for inbound service requests.
- Lower dispatcher after-hours workload.
- Create reliable executive visibility without spreadsheet hunting.
Top Workflow Frictions
- Service notes copied from calls into tickets by hand.
- Sales quotes require context from three systems.
- Dispatchers chase technicians for job status updates.
- Finance rebuilds revenue reports every Friday.
- Managers disagree about which number is current.
Top Trust Risks
- Customer data scattered across inboxes and ticket notes.
- No standard approval gate for refunds and discounts.
- Shared passwords still exist for legacy vendor portals.
- AI use already happening in personal accounts.
- No documented prompt, model, or output review policy.
First Three AI Candidates
- Call-summary to ticket-note draft.
- Quote preparation brief for sales reps.
- Weekly executive operating report from trusted data exports.
Operating Friction Map
Purpose: rank the operating pain by frequency, effort, risk, and fit for AI or automation.
| Friction | Frequency | Effort | Risk | Candidate Fix | Owner |
|---|---|---|---|---|---|
| Manual call summaries into service tickets | Daily / 90+ calls | High | Medium | AI draft + human approval | Service Ops |
| Quote research across inbox, CRM, and PDFs | Daily / 15-20 quotes | High | Medium | RAG brief + quote checklist | Sales Ops |
| Friday revenue reporting | Weekly | High | Low | Automated extract + commentary draft | Finance |
| Customer status requests | Daily | Medium | High | Internal answer assist only | Customer Care |
| Technician ETA updates | Daily | Medium | Medium | Workflow automation + exception alerts | Dispatch |
| Refund approvals | Weekly | Low | High | Do not automate yet; build decision brief | GM |
Highest-Leverage First Move
Call-summary to ticket-note draft. It's frequent, measurable, reviewable, and low enough risk if the human stays in the approval loop.
Do Not Touch Yet
Refund approvals and public customer commitments. These require governance, approval gates, and clean data before AI can safely participate.
AI Workload One-Pager
Workload
| Name | Service Call to Ticket Note Draft |
|---|---|
| Trigger | Call recording/transcript lands after inbound customer call. |
| AI Job | Summarize issue, urgency, customer sentiment, promised follow-ups, equipment references, and missing information. |
| Systems Touched | RingCentral transcript, ServiceTitan ticket, Microsoft 365 account context. |
| Human Review | Dispatcher reviews and edits before ticket note is saved. |
Value Case
- Baseline: 4.5 minutes average manual note per call.
- Target: 90 seconds average review and save.
- Volume: 90 calls per day, 22 business days.
- Hard value: 99 hours/month reclaimed before quality improvements.
- Soft value: faster handoff, fewer forgotten promises, less dispatcher fatigue.
Red Lines
- AI may not promise a technician arrival time.
- AI may not approve refunds, discounts, or credits.
- AI may not classify legal threats without escalation.
- AI may not save directly to the ticket without human review.
Pilot Decision
Proceed. The workload is frequent, bounded, measurable, and reviewable. Launch as supervised draft mode with daily quality sampling for the first two weeks.
Pilot Charter
| Pilot Name | Service Call to Ticket Note Draft |
|---|---|
| Owner | Director of Service Operations |
| Executive Sponsor | General Manager |
| Timeline | 14-day controlled pilot, 5 dispatchers, 1 location, inbound service calls only |
| In Scope | Drafting internal ticket notes from call transcripts; urgency flag; missing-information checklist. |
| Out of Scope | Customer replies, refunds, scheduling commitments, technician performance commentary, legal/compliance classification. |
| Success Threshold | 40% reduction in note time, 90% dispatcher acceptance after edit, no high-severity hallucinations. |
| Test Cases | Routine request, angry customer, warranty question, missing equipment ID, refund demand, emergency request, unclear audio. |
| Kill Criteria | Two high-severity errors, review time exceeds manual time for three consecutive days, dispatcher trust score below 3/5. |
Value Readout
Time
Manual note time fell from 4.5 minutes to 1.7 minutes average. Estimated 92 hours/month reclaimed.
Quality
Dispatcher acceptance rate after edit reached 88%. Missed follow-up mentions fell by 31% in sampled tickets.
Adoption
Four of five dispatchers asked to keep the tool. One dispatcher requested more control over summary format.
| Question | Finding | Decision |
|---|---|---|
| What changed? | Notes are faster, more consistent, and less emotionally loaded after difficult calls. | Graduate to second location. |
| What failed? | Urgency labels were too aggressive for warranty language. Prompt revised. | Keep urgency as recommendation only. |
| What value was created? | 92 hours/month reclaimed, fewer handoff misses, better dispatcher morale. | Add to reusable workload library. |
| What comes next? | Quote-prep brief is now the strongest adjacent workflow. | Draft pilot charter. |
Governance Desk
A plain-English policy that lets the business scale AI without pretending probabilistic systems are ordinary software.
| Agent May Do | Agent May Recommend | Agent May Never Do |
|---|---|---|
| Summarize, classify, draft internal notes, prepare briefs, flag missing information, propose next steps. | Customer response language, quote assumptions, exception handling, escalation severity, workflow changes. | Send public messages, approve refunds, delete records, change pricing, commit legal positions, bypass human review. |
Who Decides What
Every request an agent submits carries its own routing authority. The route decides which human can act on it.
| Route | Who Acts |
|---|---|
| Log only | Nobody. The record is the point. |
| Standard approval | A different human than the submitter. No self-approval, ever. |
| Manager approval | The submitter's manager. |
| Team approval | Any member of the named team. |
| Admin approval | Any admin. |
Logging Standard
- User
- Agent
- Model
- Prompt version
- Inputs and retrieval sources
- Output
- Human decision
- Cost
Review Cadence
- Daily sampling during pilot
- Weekly quality review
- Monthly cost and risk review
- Quarterly model/vendor review
Incident Response
- Pause workload
- Preserve logs
- Notify owner and sponsor
- Classify severity
- Fix prompt/data/control
- Approve relaunch
Reusable Skills Library
Purpose: catalog proven capability as versioned skill packages an agent loads on demand. Each row is a folder in version control: a trigger description, a procedure, and any bundled scripts.
| Skill | Fires When | Bundled Tools | Controls | Owner / Version |
|---|---|---|---|---|
| Executive Summarizer | "Summarize this," "give me talking points," or a long paste asking for a tighter version | None; procedure only | Preserves the author's stance; banned-word list for AI slop; flags unsupported claims instead of smoothing them | Content desk / v2 |
| Database Operations | "Apply the migration," "run this SQL," or a pasted project reference | Four helper scripts; access token fetched from the OS keychain at runtime | Never echoes credentials; confirm before destructive statements; migrations additive by default | Engineering / v3 |
| Signal Curator | "Review fresh signals," "promote signal," "list angles" | Pipeline API scripts | Agent scores and stages; only a human promotes toward publish | Content desk / v1 |
| Ad Pre-Screen | New paid creative lands in the queue | Moderation API call | Clean pass auto-approves; flagged waits for a human; payment held in escrow until decided | Revenue desk / v1 |
Approved Pattern
Prompt first, in draft mode. The third reuse, or the first credential, forces graduation into a package. Version it, name an owner, review every edit like code.
Known Failure Mode
Skill drift. An edit that skips review silently changes the behavior of every agent that loads the skill. Treat a skill edit like a schema change, not a wording tweak.
Board-Ready Transformation Story
The Narrative
We inherited a hardworking operation with fragmented systems, spreadsheet-dependent reporting, and undocumented AI use. In year one, we stabilized the data flows, created an AI governance desk, piloted four supervised workloads, and promoted two into standard operating practice.
The result wasn't "AI magic." It was operating discipline: clearer ownership, faster handoffs, fewer missed follow-ups, and a reusable library of patterns that can now be applied across locations.
Board Metrics
- 92 hours/month reclaimed from first graduated workload.
- 31% reduction in sampled missed follow-ups.
- Two AI workloads graduated, one revised, one killed.
- AI usage moved from personal accounts into governed lanes.
- Prompt, model, logging, and escalation standards established.
Risk Retired
Untracked AI use, shared prompts, unmanaged keys, missing approval gates, and unclear accountability.
Value Created
Time savings, quality consistency, operating visibility, reusable assets, and stronger employee trust.
Year-Two Roadmap
Expand workload library, add deeper retrieval, integrate CRM and finance reporting, and formalize vendor/model review.