# Utility AI pilot brief

Use this working document to agree on one workflow, a measurable baseline, and a bounded pilot. Replace bracketed text with your own evidence and decisions.

## 1. The workflow

- Pilot name: [name]
- Business owner and sponsor: [names / roles]
- Process owner and participating users: [names / roles]
- Problem to solve: [one specific problem, with examples]
- Start and end of workflow: [what triggers the work / what a completed task looks like]
- In scope: [tasks, teams, systems, data]
- Out of scope: [tasks, decisions, integrations]
- Pilot dates and review checkpoints: [dates]

## 2. Baseline and hypothesis

- Current task volume per month: [number and measurement period]
- Current median time per task: [minutes, including rework]
- Current quality or error rate: [definition, sample size, measurement]
- Current waiting time / service outcome: [measure]
- Proposed AI assistance: [what the system will do]
- Hypothesis: [we expect this change to improve this measure by this amount]
- Expected review time per AI-assisted task: [minutes]
- Adoption assumption: [% of eligible work that will actually use it]
- Evidence behind assumptions: [observations, time study, pilot sample]

Separate time made available for other work from cash savings. Record a cash-saving claim only if you can identify the corresponding budget change, such as avoided overtime or external spend, and who can validate it.

## 3. Data and access

- Source systems and accountable data owners: [list]
- Authoritative records and document versions: [list]
- Representative sample, including difficult cases: [description]
- Known quality issues and correction process: [issues / proposed controls]
- Sensitive data and access restrictions: [classification and rules]
- Approved environment, retention, and deletion rules: [details]
- Source traceability required for outputs: [citations / source IDs / audit events]

## 4. Operating boundaries

- Actions the AI may take: [explicit allowlist]
- Actions requiring human approval: [list]
- Actions it must never take: [list]
- Review owner and escalation path: [roles and contact process]
- Failure cases to test: [missing data, conflicting sources, unsupported claims, misuse]
- Fallback when unavailable or uncertain: [existing process]
- Rollback / stop mechanism and accountable owner: [details]

Keep operational control, safety-critical decisions, and changes to systems of record outside the pilot unless separately approved through the utility's established engineering, safety, security, and change-management processes.

## 5. Evaluation and success criteria

- Evaluation set: [representative cases; held out from prompt or workflow development]
- Quality rubric and acceptable threshold: [accuracy, completeness, traceability]
- Time measure: [end-to-end time including review, correction, and exception handling]
- User measure: [adoption / satisfaction / successful task completion]
- Safety and security criteria: [required checks and unacceptable failures]
- Minimum sample size and evaluation period: [appropriate to the use case]
- Success threshold: [pre-agreed measures]
- Stop conditions: [specific failures or unacceptable outcomes]
- Evaluators and decision owner: [roles]

## 6. Costs and value

- One-time implementation: [integration, setup, data preparation, training]
- Annual running cost: [software, usage, hosting, support, monitoring]
- Time returned after review: [hours; include negative results]
- Modeled staff capacity value: [hours multiplied by fully loaded hourly cost]
- Capacity deployment plan: [what useful work will the team do with the time?]
- Verified cash changes, if any: [budget lines and evidence]
- Other benefits and tradeoffs: [quality, response time, reliability; avoid double counting]

## 7. Decision and next steps

- Outcome: [scale / revise / stop]
- Evidence supporting the decision: [measures, exceptions, user feedback]
- Unresolved issues and owners: [list]
- Requirements for wider deployment: [controls, training, integrations, support]
- Next review date: [date]

---

Template from AI for Utilities — https://aiforutilities.org/toolkit/
Adapt this document to your utility's policies and review processes.