Sheet 09 · Business case
Value, Measurement & Risk
How success gets measured, what could go wrong, and what we ask of you.
No AI initiative should be judged on enthusiasm. Every pilot in this plan starts by measuring the current state — how long a report takes today, how long staff spend hunting for documents, how long a quote takes to turn around — and is then judged against that baseline with numbers, not anecdotes.
Just as important: each pilot has explicit exit criteria agreed before it starts. If a pilot does not clear its bar, the honest move is to stop, learn, and redirect the budget — and this plan says so in writing. Nothing below promises a result; it promises a measurement.
How each pilot is measured
Field-service reporting
Baseline first
Hours from job completion to filed report; reports filed per technician per week.
Then we measure
A measured reduction in drafting time with technician approval on every report — targets set against the baseline in discovery.
Knowledge assistant
Baseline first
Time staff spend locating procedures, specs, and policy answers; repeat questions reaching senior staff.
Then we measure
Cited answers in seconds, measured answer accuracy on a review sample, and fewer interruptions to senior staff.
Quote workflow
Baseline first
Quote turnaround time and rework rate today.
Then we measure
Faster first drafts with pricing judgment kept fully human — no quote leaves without review.
Exit criteria are agreed in writing before each pilot starts: adoption by the pilot team, measured accuracy on a human-reviewed sample, and a clear scale / adjust / stop decision at the end. If a pilot misses its bar, we recommend stopping it — and say so.
Risk register — named up front
| Risk | What it looks like | Mitigation in this plan |
|---|---|---|
| Adoption stalls | Tools get licensed, dashboards get built, nobody changes how they work. | Named champions per department, training in Phase 1, usage reviewed at every checkpoint — and pilots start where the pain is highest. |
| Wrong or invented answers | An assistant states something confidently that is not true. | Retrieval limited to approved sources, visible citations, refusal when sources are missing, human approval on anything that leaves the building. |
| Data leakage | Sensitive material ends up in a public AI service. | Four data classes with an approved-platform matrix; restricted content only ever on local models; audit logs on administrative and AI activity. |
| Vendor price or behavior changes | A license doubles in price, or a model update changes output quality. | Multi-platform strategy — no single-vendor dependence; local models as a floor; sectional plan lets you re-route without sunk-cost pressure. |
| Scope creep | A focused pilot grows into an unbounded platform project. | Fixed-scope sections priced separately; anything new is a new section you approve — or decline. |
| Over-reliance | Staff stop checking AI output; skills atrophy in critical roles. | Human-in-the-loop is a design rule, not a preference: approval steps stay mandatory for reports, quotes, and quality documents. |
What we need from Vulcan Industrial
- An executive sponsor with authority to unblock decisions
- A pilot team of 3–5 people who actually do the work being piloted
- Access windows to the documents and systems in scope, under the data classification rules
- Thirty minutes a week for a standing checkpoint — decisions, not status theater
That is the whole ask. No steering committee, no six-month requirements phase.