FieldOps AI

Turning contractor data into earlier, evidence-backed decisions
AI operations intelligence for commercial contractors

Concept Project — Self-Directed

Role

Product Strategy · UX Research · AI Experience Design · Interaction Design · Prototyping · Usability Testing

Product

0→1 enterprise product concept

Primary User

Director of Operations / Operations Manager

Industry

Commercial contracting / specialty trades

The Problem

Commercial contractors generate enormous amounts of project data. Daily reports document field conditions. Timecards capture labor. Estimates define planned scope. Accounting systems track cost. Project-management platforms record changes, schedules, and documentation.

The problem isn’t the absence of data. The information often remains disconnected until someone manually realizes something has gone wrong.

“Commercial contractors don’t need another dashboard. They need earlier visibility into operational signals that put project profitability at risk.”

Why This Problem

This project didn’t start with market research — it started with five years running paint crews.

Before I was a designer, I put myself through college by painting dorms on campus. When the contractor I was working for stopped showing up, I kept the job going myself — and ended up becoming the contractor. That turned into a full commercial painting operation: five four-man crews and a handful of partners, tracking labor, change orders, scope creep, and daily site conditions the hard way, on paper and in my head, for four years before I sold my share and went back to school for design.

Every detail in this case study — the labor variance, the skim coating, the adhesive removal, the daily reports that don’t talk to the accounting system — isn’t researched. It’s remembered. FieldOps is the tool I wish I’d had running those crews: something that would have caught the Riverside-style scope creep before it ate the margin, instead of after.

Product Opportunity

The product question became: How might we help contractor operations leaders recognize project risk while there is still time to act?

FieldOps would not become another ERP, project-management system, estimating platform, scheduling tool, or accounting application. Instead, it would sit above those systems and connect their signals.

Existing Workflow
Field event

Data

Delay

Cost variance

Investigation

Action
Future Workflow

 

Field event

AI signal

Evidence

Human decision

Action

Designing AI With Clear Boundaries

The key design challenge was deciding what authority the AI should have.

  1. Inform — Labor is currently 18% above budget.
  2. Explain — Most of the variance occurred during additional surface preparation.
  3. Recommend — Review this work for possible change-order eligibility.
  4. Prepare — Draft a Potential Change Order using available project evidence.
  5. Act — Only after explicit human authorization.

“The challenge wasn’t giving AI access to more information. It was determining when the system should inform, recommend, prepare, or act.”

1
Inform
Labor is 18% above budget
2
Explain
Variance traced to surface prep
3
Recommend
Review for change-order eligibility
4
Prepare
Draft a potential change order
5
Act
Only after human approval

The Pivot: From Full System to Intelligence Layer

FieldOps didn’t start as an intelligence layer — it started as a full operations platform: hour tracking, proposals, change orders, and daily site updates for every job, all built from scratch.

Research into the existing landscape changed that. Every major contractor platform already has hour tracking, proposals, and change-order workflows — and every one of them was racing to bolt on an AI chat feature. But watching how those chat features actually got used told a different story: ops leaders and PMs were spending hours in conversations with AI, still having to manually pull the information out themselves. The chat was another interface to operate, not intelligence that came to them.

That reframed the product entirely. FieldOps didn’t need to rebuild hour tracking, proposals, or change orders — those already existed. What was missing was a layer that sat above all of it, watched for patterns across systems that don’t talk to each other, and surfaced risk before anyone had to ask. I cut the full-system scope down to exactly that: an intelligence layer, not another platform to log into.

The Scenario

To make the concept realistic, I designed the product around Summit Commercial Coatings, a fictional commercial painting contractor.

30+

Employees

5

Field Crews

$7M

Annual Revenue

12–18

Active Projects

The primary persona was Alex Morgan, Director of Operations.

“Which jobs need my attention today?”

The Riverside Risk

The primary prototype follows Riverside Medical Center ($286K contract), where unexpected surface preparation caused labor to accumulate faster than planned.

Original bid scope


Standard surface preparation

Minor repairs

Primer

Two finish coats

Observed field condition


Mechanical adhesive removal

Additional wall repair

Skim coating

Additional primer prep

Daily reports described adhesive removal, damaged drywall, skim coating, and additional preparation. Labor records showed 142 excess hours accumulating around the same work. Individually, each system contained only part of the story. FieldOps connected them.

29.0%

Target Margin

21.8%

Forecast Margin

$20.6K

Margin at Risk

Designing for Prioritization, Not Another Dashboard

The Overview does not lead with generic KPI cards. It answers the operations question first: What requires my attention?

Riverside appears as the highest-risk project, with the $20.6K margin-at-risk signal, the margin decline from 29.0% to 21.8%, 142 excess labor hours, and the AI recommendation presented before secondary reporting metrics.

Making AI Conclusions Inspectable

The next design question was not simply whether the AI was correct, but whether the user could inspect the reasoning.

Supporting evidence included five daily reports, fourteen site photographs, crew time records, the original scope, and project specifications.

“Success is not how often someone talks to the AI. It is whether the right information reaches the right decision-maker early enough to change the outcome.”

“Instead of asking ‘Do you trust the AI?’, the interface asks: Does the evidence support this conclusion?”

From Risk to Action

Once the user determines the condition may be outside the intended scope, FieldOps prepares an internal Potential Change Order draft. The AI assembles labor, materials, mobilization, markup, schedule impact, and supporting documentation for human review.

$8,946

Labor

$1,280

Materials

$740

Other

$13,159

Suggested PCO

The interface deliberately distinguishes cost exposure, PCO value, approved recovery, and collected revenue. The AI prepares the work; the human approves it.

Approval requires an explicit human action — nothing here happens automatically

$20,600

Detected Margin Risk

$13,159

Potential Recovery

$7,441

Remaining Exposure

Showing the Outcome Without Overclaiming

The product does not claim that the $13,159 has been recovered. It clearly states that recovery remains potential until the contractual change process advances.

Ask Copilot

Proactive alerts handle known risks, but operations leaders also need to investigate questions the system has not surfaced. Ask Copilot supports portfolio-level questions and returns structured operational intelligence rather than long chat responses.

“Which active projects are most likely to finish below our 27% margin target?”

The answer surfaces Riverside, Dominion Office Renovation, and Fairfax Distribution Center with projected margin, primary risk driver, confidence level, and grounded project sources. Copilot became an investigation tool — not the homepage.

Prototype Testing

I tested the clickable prototype with four participants across five tasks: identify the highest-risk project, understand Riverside’s risk, validate the evidence, prepare and internally approve a PCO, and use Copilot to investigate portfolio-level risk.

4

Participants

3 of 4

Full-Flow Success

20

Task Attempts

17/20

Completed

Three participants completed all five tasks and described the workflow as very easy. One retired operations manager with lower familiarity with modern digital interfaces completed two tasks, made wrong turns in the middle of the flow, and could not complete the Copilot task.

Because the sample was small, I treated the 85% task-completion rate as directional qualitative evidence — not a statistically meaningful performance metric.

What Testing Changed

  • Evidence Review — Strengthened human-decision wording and clarified the transition to preparing a PCO draft.
  • Action Review — Changed the approval language to “Approve Internal PCO” so users would not confuse internal review with contractual approval.
  • Copilot — Changed the navigation label from “Copilot” to “Ask Copilot” to make the feature’s purpose more explicit.
Participant Identify
risk
Understand
risk
Validate
evidence
Prepare
PCO
Ask
Copilot
P1
P2
P3
P4
P4 was a retired operations manager with lower familiarity with modern digital interfaces.

Prototype QA & Iteration

Participant research and prototype QA were tracked separately. Working AI-assisted meant catching a different category of issue than typical design QA — problems that came from the AI-generated output itself, not just standard iteration.

Two examples stood out. The AI Recommendation alert card — one of the most important elements on the page, since it’s where the system surfaces its conclusion to the user — came out of AI generation with no internal padding at all, text sitting flush against the card edges. It read as unfinished and undermined the exact moment the interface most needed to feel trustworthy. I added the padding manually in Figma. On the Overview dashboard, the main scrolling region had a weak, abrupt ending — content just stopped rather than resolving into a proper close to the page. Both are the kind of details an inexperienced eye — human or AI — misses, but that quietly damage how polished and trustworthy an interface feels.

Beyond these two, QA also identified collapsed financial cards, numeric misalignment, inconsistent result-link alignment, sidebar movement, navigation geometry differences, and hidden text-overflow issues — each documented as Finding → Impact → Change → Result.

Visual Direction

The final interface uses an Industrial Intelligence visual system inspired by the environments FieldOps supports — work sites, safety equipment, and industrial materials, not a typical SaaS dashboard.

That direction wasn’t the starting point. Left to its own defaults, AI generation pulled toward purple and blue — a palette that reads as generic tech SaaS, the same look as a hundred other dashboards, with none of the grounded, on-the-job feel this product needed. Commercial contractors live in concrete, steel, and safety orange, not gradient-purple software chrome. I moved the palette to 60% concrete neutrals, 30% steel/charcoal, and 10% functional orange for actions — with semantic red/amber/green layered on top for risk states — so the interface would feel like it belonged in the world it was actually built for.

60% Concrete neutrals
30% Steel / charcoal
10% Functional orange
High risk
Watch
Healthy

Outcome

FieldOps AI began as a question about AI dashboards. It evolved into a narrower and stronger product idea: an operational intelligence layer that connects field activity to financial consequences before those consequences become expensive surprises.

  • AI should surface evidence, not simply conclusions.
  • Automation authority should increase gradually as risk decreases and user trust increases.
  • In high-stakes workflows, AI should prepare decisions — not quietly make them.

What I Would Measure Next

  • Time to risk detection — Can the system reduce risk discovery from several days or weeks to less than 24 hours?
  • Change-order capture rate — Are more legitimate scope deviations documented before the opportunity is lost?
  • Investigation time — How much PM/operations time is required to assemble scope, labor, reports, photos, and cost evidence?
  • Early margin visibility — How often are significant project variances discovered before project completion?
  • AI recommendation quality — How often are recommendations accepted, modified, rejected, or investigated further?
"Success is not how often someone talks to the AI. It is whether the right information reaches the right decision-maker early enough to change the outcome."