The assessment ends with a recommendation written like this one: the AI opportunities ranked, what each is worth, what it would cost, what data it touches, and what has to be true first. This sample is for a fictional contractor. The format is what you would receive.
Illustrative sample. The company, people and figures are fictional. The format is what an assessment client receives.
Decision memo
To: Owner and General Manager, Cedar Basin Electric (fictional)
From: Ashlee, Sandcastle
Date: 29 September 2026
Subject: Which AI opportunity to fund first
Background
Over three weeks we interviewed six people and inventoried nine systems. Eleven places where AI could save time came out of those interviews. Four are shown here: the top three, and one that was asked about and that we recommend against. Each was scored the same way, on value, feasibility, data readiness and risk.
The pattern behind most of them is the same. Crew hours, job costs and vendor invoices are each entered once and then re-keyed or checked by hand somewhere else, about 24 hours a week across two coordinators, accounting and the project managers.
Options considered
| Rank |
Opportunity |
Score (1 to 5) |
Rough annual value |
Budget range |
Data level |
What has to happen first |
| 1 |
Timesheet exception flags before payroll. Each week’s crew hours are checked against the schedule, the job and each person’s pattern, and the odd ones are listed for a coordinator to review. |
4.1 |
$11,000 to $17,000 a year (6 hours a week of manual checking, plus fewer payroll corrections) |
$8,000 to $20,000 |
Level 3, protected (employee hours and pay) |
None. The field time-entry app already has an export. |
| 2 |
Daily job report drafts. Foremen’s notes and photos become a first draft of the daily report, which the project manager edits and approves. |
3.4 |
$15,000 to $25,000 a year (5 project managers, about 2 hours a week each) |
$10,000 to $25,000 |
Level 2, business |
Two crews still take notes on paper. Notes have to be captured on the phone first. |
| 3 |
Vendor invoice matching. Incoming invoices are read, matched to purchase orders, and only the mismatches go to accounting. |
3.0 |
$10,000 to $15,000 a year (about 8 hours a week of re-entry and matching) |
$10,000 to $30,000 |
Level 2, business |
Blocked. The job-costing system has no export. Fix that first ($2,500 to $15,000, or the vendor’s cloud edition). |
| 4 |
A chatbot on the public website. |
1.5 |
Unclear. Nobody interviewed could name a question customers ask that the site does not already answer. |
$25,000 to $60,000 |
Level 1, open |
Not recommended. High cost, customer-facing risk, no measurable saving. |
Value is hours saved a week, times 48 weeks, times a loaded hourly cost ($65 for coordinators and accounting, $90 for project managers), then reduced by half or a quarter because a person still reviews the output. The arithmetic for each line is in the appendix of the full report.
Recommendation
Fund opportunity 1 now. It has the highest score, it needs nothing fixed first, a coordinator reviews every flag so nothing reaches payroll unchecked, and it pays for itself inside two years even at the top of the budget range and the bottom of the value range.
Before spending anything, check whether the field time-entry app’s own exception report covers part of this. If it covers half, the build gets smaller or unnecessary, and that is a good outcome.
Opportunity 2 is the second project, once the two paper crews are entering notes on the phone. Opportunity 3 has real value but waits on the job-costing export. We recommend against the website chatbot.
What data can go where
- Crew hours and pay are level 3, protected. They are processed only under a signed data agreement that excludes training on your data, with names replaced by employee numbers. They are never pasted into a public chat tool.
- Job reports and vendor invoices are level 2, business. They can go to an AI service under business terms, with the minimum fields needed.
- Nothing in these four opportunities touches card numbers or customer contracts. If that changes, the rule is to stop and ask.
What has to be true for this to work
- The field app’s vendor confirms in writing that weekly hours can be exported automatically.
- One coordinator owns the weekly exception list and has the authority to correct hours before payroll runs.
- The first month runs alongside the current manual check, so the flags can be compared with what people catch today.
Risks and how we would handle them
- Too many false flags and the list gets ignored: start with three rules that the coordinators already apply by eye, measure, then add.
- A real error is missed: the manual check stays in place for the first month, and nothing is removed from the process until the numbers agree.
- Staff read it as surveillance: the list goes to the coordinator who already checks hours, nobody new sees it, and the crews are told what it does.
Next steps
- Owner chooses: opportunity 1 now, a smaller first step, or nothing yet.
- The coordinator checks the field app’s built-in exception report this week.
- If the build goes ahead, it is a four to six week project with a fixed quote, and half the assessment fee is credited against it.
Every recommendation includes options that do not involve Sandcastle. Here, the field app’s own report may cover part of opportunity 1, and opportunity 2 can start with a tool the company already licenses.
