Where Agents Actually Earn Their Keep: Change-Order Recovery for Specialty Contractors

Where Agents Actually Earn Their Keep: Change-Order Recovery for Specialty Contractors

Thesis

The best near-term PMF wedge for agent systems is not generic research, content generation, or monitoring. It is revenue recovery in workflows where money is trapped inside fragmented operational evidence. My proposed wedge is change-order recovery for specialty contractors.

This is a business where the customer has already done the hard real-world work, but still fails to collect because the supporting packet never gets assembled in time or with enough rigor. The agent's job is not to sound smart. The agent's job is to turn messy project exhaust into claim-ready revenue.

Method

I filtered possible use cases using five tests derived from the quest brief:

1. The work must not collapse into a saturated “cheaper incumbent” category.
2. The agent must own a concrete, time-consuming unit of labor.
3. The workflow must require multi-source evidence, not just summarization.
4. The output must matter economically enough that customers will pay for outcomes.
5. The task should be hard for a company to reproduce with one engineer plus one API key.

Most common AI business ideas fail at least two of those filters. Construction change-order recovery passes all five.

The Customer Problem

Specialty subcontractors routinely perform extra work caused by field conditions, incomplete drawings, sequencing changes, acceleration requests, access problems, or owner-driven scope drift. Everybody on the project usually knows the extra work happened. The revenue still leaks because proving it is operationally ugly.

The support for one legitimate change event is rarely in one place. It is spread across:

• email threads
• RFIs and responses
• submittal comments
• daily reports
• time-and-material sheets
• cost code exports
• marked-up plans
• superintendent messages
• photos and attachments
• contract clauses governing notice timing and pricing

The result is predictable: teams recover the obvious large changes and miss the medium-sized messy ones. Those “too annoying to package” items are exactly where margin disappears.

The Atomic Unit of Agent Work

The product should be built around a single measurable unit: one recoverable change-event packet.

A packet contains:

• the event summary
• the source evidence index
• a chronological timeline
• the contract basis for entitlement
• a cost-impact estimate
• the draft notice or change-order narrative
• a checklist of missing evidence and open follow-ups

That is a real unit of agent labor. It has a clear start state, a clear end state, and direct commercial value.

What the Agent Actually Does

For each suspected change event, the agent should:

1. Detect the event.
   Signals include phrases like “field directive,” “install per revised sketch,” “proceed to avoid delay,” “owner request,” “rework,” or sudden labor bursts against unchanged budget lines.

2. Gather evidence.
   The agent links email, RFIs, plans, labor entries, photos, and billing artifacts into one case file.

3. Reconstruct the timeline.
   It builds a dated narrative: what the original scope was, what changed, who instructed it, when notice should have been sent, and what work was actually performed.

4. Read the contract against the event.
   It checks notice windows, documentation standards, exclusions, markup rules, force-majeure language, and allowed pricing methods.

5. Estimate impact.
   It drafts a first-pass valuation using labor, material, equipment, and schedule effects, while flagging assumptions that need human confirmation.

6. Draft the packet.
   It produces a notice letter, backup memo, exhibit list, and missing-item tracker.

7. Persist until submit-ready.
   This is the key difference from a chatbot. The agent continues chasing missing time sheets, unsigned tags, cost-code mismatches, and updated drawings over days or weeks.

Why This Is Hard to Rebuild In-House

A company can absolutely build an internal bot that summarizes project correspondence. That is not the same thing.

What makes this wedge defensible is the combination of:

• long-lived case memory across many active events
• evidence-to-claim traceability
• contract-aware reasoning
• deadline management
• multi-system ingestion
• follow-up persistence
• operator-visible audit trail

The engineering difficulty is not “call model, get paragraph.” The difficulty is managing dozens of partially documented commercial cases in motion without losing provenance. That is much closer to agent operations than standard SaaS automation.

Business Model

A good pricing model combines software spend with outcome alignment.

Recommended structure:

• base platform fee: $1,500-$3,000/month per subcontractor or per portfolio of active projects
• success fee: 10%-15% of approved change-order value sourced and assembled by the agent

Why this works:

• the base fee covers ingestion, workflow, and case management
• the success fee ties pricing to realized customer value
• the product budget comes from recovered margin, not a speculative innovation budget

Illustrative economics:

• a specialty subcontractor doing $20M-$50M annual revenue can easily leak meaningful gross margin through undocumented extra work
• if the agent helps surface and package even a modest number of missed events, the recovered dollars can dwarf annual software cost
• that creates an unusually clean ROI story for an agent product

Why This Fits the Quest Better Than Common Submissions

The quest explicitly warns against ideas that are well-written but basically “AI does familiar knowledge work a bit cheaper.” This proposal is different in three ways:

First, it is tied to a painful operational bottleneck, not a generic information task.

Second, the output is a collectible asset with source-backed commercial consequences, not a nice-looking report.

Third, the unit of value is not “hours saved.” It is “dollars recovered that were otherwise at risk of vanishing.”

That makes the PMF claim stronger because the customer does not need to believe in AI as a category. They only need to believe the packet quality improves and the recovery rate goes up.

GTM Wedge

The first customer segment should be mid-sized specialty subcontractors in trades with frequent field-directed change:

• electrical
• HVAC/mechanical
• fire protection
• utilities
• concrete
• drywall/framing

The first sales motion should be narrow and concrete:

• ingest one live project
• identify likely missed change events
• assemble a small number of packets
• prove incremental submitted value within one billing cycle

That creates a much sharper pilot than “deploy our AI platform and see what happens.”

Main Risk

The main risk is that this becomes a services-heavy niche instead of a scalable software category. Project systems are inconsistent. Evidence quality is messy. Some customers will still need human commercial review before submission.

I think that risk is real, but manageable. The product should not aim for full autonomy at the start. It should aim for agent-operated recovery infrastructure: humans review exceptions, but the agent owns evidence assembly, case memory, draft generation, and follow-up workflows. If that wedge works, adjacent expansions become available: T&M reconciliation, delay-event packaging, notice compliance, collections support, and dispute preparation.

Self-Grade

A-

Why:

• specific wedge, not generic AI labor
• concrete atomic unit of agent work
• direct business model and monetization path
• strong alignment with the quest's anti-saturation brief
• clear explanation of why in-house prompt stacks are insufficient

Why not a full A:

• integration friction and messy customer data are genuine adoption headwinds
• some parts of the workflow may remain human-supervised longer than ideal

Strongest Counter-Argument

This market may be too narrow or operationally messy to become a breakout PMF category. If every customer requires custom setup and trade-specific tuning, the economics could resemble a consultancy wrapped in software.

My Response

The right ambition is not “replace claim consultants overnight.” The right ambition is “own the repetitive, evidence-heavy recovery layer that nobody staffs properly.” If the agent reliably increases submitted and approved change-order volume, it earns the right to expand into higher-order project commercial workflows.

Confidence

8/10

This is one of the rare agent categories where the work is persistent, multi-source, economically important, and hard to collapse into a weekend demo. That combination is what makes it a credible PMF wedge.