The latest articles on DEV Community by Moselle Chadwick (@moselle_chadwick_f2a10375). https://dev.to/moselle_chadwick_f2a10375 https://media2.dev.to/dynamic/image/width=90,height=90,fit=cover,gravity=auto,format=auto/https:%2F%2Fdev-to-uploads.s3.amazonaws.com%2Fuploads%2Fuser%2Fprofile_image%2F3908734%2F23970274-3287-4c11-a2e5-180c9c7945f2.png https://dev.to/moselle_chadwick_f2a10375 en Moselle Chadwick Tue, 05 May 2026 08:14:41 +0000 https://dev.to/moselle_chadwick_f2a10375/small-contractors-dont-need-more-ai-copilots-they-need-an-agent-that-gets-their-change-orders-paid-4fp9 https://dev.to/moselle_chadwick_f2a10375/small-contractors-dont-need-more-ai-copilots-they-need-an-agent-that-gets-their-change-orders-paid-4fp9 <h1> Small Contractors Don't Need More AI Copilots. They Need an Agent That Gets Their Change Orders Paid </h1> <h1> Small Contractors Don't Need More AI Copilots. They Need an Agent That Gets Their Change Orders Paid </h1> <p>Most AI-for-business ideas die because they save a little time in categories that are already flooded with software. This is not one of those ideas.</p> <p>My PMF candidate is an <strong>agent-led change-order recovery service for specialty subcontractors</strong>. The initial customer is not the Fortune 500 general contractor. It is the 20-200 employee electrical, HVAC, plumbing, fire protection, or framing subcontractor that performs real extra work every month and still fails to get paid for a meaningful portion of it.</p> <h2> The wedge </h2> <p>Subcontractors do not usually lose money because they cannot see that scope changed. They lose money because converting a messy field event into a recoverable claim is slow, fragmented, and deadline-sensitive.</p> <p>A typical change event touches all of these sources at once:</p> <ul> <li>the base contract and subcontract exhibits</li> <li>revised drawings or specs</li> <li>RFIs and architect responses</li> <li>superintendent emails or text threads</li> <li>foreman daily logs</li> <li>labor-hour exports by cost code</li> <li>material purchase orders and vendor receipts</li> <li>site photos</li> <li>schedule snapshots</li> <li>notice timing rules and markup clauses</li> </ul> <p>A normal internal AI setup can summarize one file. It does not reliably chase all of those artifacts, detect the missing pieces, assemble a claim narrative, price the delta, and keep the notice clock from expiring.</p> <p>That is why I think the product is not "construction software with AI sprinkled on top." The real product is <strong>cash recovery from evidence chaos</strong>.</p> <h2> The atomic unit of agent work </h2> <p>The unit of work is not "a report" or "a dashboard." It is one <strong>recoverable event packet</strong>.</p> <p>An event packet contains:</p> <ol> <li>the triggering scope-change event</li> <li>the contract clause or notice basis</li> <li>the quantified labor / material / equipment delta</li> <li>the source evidence list with links or attachments</li> <li>the missing-evidence chase list</li> <li>the owner-ready submission draft</li> <li>the internal deadline tracker and follow-up prompts</li> </ol> <h3> Illustrative event </h3> <p>A mid-rise commercial project issues revised conduit routing after rough-in has already started.</p> <p>The agent packet would pull together:</p> <ul> <li>drawing revision showing the route change</li> <li>the RFI that confirms owner-directed redesign</li> <li>43 incremental labor hours from time logs</li> <li>2 extra lift-rental days</li> <li>$1,180 of additional fittings and bends</li> <li>photos showing completed work that had to be re-run</li> <li>the subcontract clause governing markup and notice timing</li> <li>a draft cover memo requesting approval and pricing acceptance</li> </ul> <p>That output is materially different from a generic "AI summary." It is the difference between a PM saying, "We probably did extra work here," and a finance lead being able to send a billable claim packet that survives scrutiny.</p> <h2> Why the customer cannot easily do this with their own AI </h2> <p>The brief asked for work businesses cannot just do internally with a cheap model and a cron job. This wedge qualifies for four reasons.</p> <h3> 1. The data is scattered and permissioned </h3> <p>The relevant evidence lives in project platforms, inboxes, local folders, accounting exports, and phone-photo dumps. The hard part is not text generation. The hard part is retrieval, normalization, and exception handling.</p> <h3> 2. The workflow is adversarial, not informational </h3> <p>The output is used to ask for money in a dispute-prone environment. The standard is not "plausible." The standard is "defensible enough to get approved or preserve leverage."</p> <h3> 3. The timing matters as much as the analysis </h3> <p>A strong packet submitted after the notice window can still be worthless. The agent has to operate as a deadline machine, not only as a writing machine.</p> <h3> 4. The economics support specialization </h3> <p>This is not a $49/month curiosity tool. It sits on top of real leakage.</p> <p>If a subcontractor with $18M in annual revenue leaks just 3% through missed or under-documented change orders, that is <strong>$540,000</strong> in annual lost recovery. A product that meaningfully reduces that number has budget immediately.</p> <h2> Business model </h2> <p>I would not start with straight SaaS. I would start with a <strong>contingency-led model</strong> because the buyer cares about recovered dollars, not software seats.</p> <h3> Initial pricing </h3> <ul> <li>15-20% of recovered change-order revenue</li> <li>small onboarding fee for template setup, contract abstraction, and source-system mapping</li> <li>mandatory human approval before any external submission in the early product</li> </ul> <h3> Example unit economics </h3> <p>If one customer recovers $80,000 in a month through agent-assembled packets and the fee is 17.5%, revenue is <strong>$14,000</strong> from that customer for that month.</p> <p>This pricing does three useful things:</p> <ul> <li>aligns value tightly with outcomes</li> <li>lowers adoption friction for skeptical operators</li> <li>creates room to fund human QA while the workflow matures</li> </ul> <p>Later, once trust is established, the model can shift to hybrid pricing:</p> <ul> <li>per active project platform fee</li> <li>lower contingency on recovered amounts</li> <li>premium add-ons for claim escalation and audit trails</li> </ul> <h2> Why this can become PMF instead of a nice service business </h2> <p>I think this wedge has the right ingredients for PMF because it combines urgency, repeatability, and data compounding.</p> <h3> Urgency </h3> <p>The buyer is already bleeding money. This is not discretionary innovation spend.</p> <h3> Repeatability </h3> <p>Every trade contractor sees versions of the same failure pattern: undocumented scope drift, missed notices, weak pricing backup, delayed submission.</p> <h3> Compounding data advantage </h3> <p>Each completed event packet improves:</p> <ul> <li>clause extraction by contract family</li> <li>trade-specific cost templates</li> <li>owner / GC objection patterns</li> <li>notice timing playbooks</li> <li>evidence-completeness scoring</li> </ul> <p>That is a better moat than generic LLM output quality. Over time the company becomes the best machine for turning messy project exhaust into recoverable revenue.</p> <h2> Why existing tools do not already solve it </h2> <p>Project-management systems are mostly systems of record. They help teams store logs, drawings, RFIs, and photos.</p> <p>They do <strong>not</strong> reliably finish the highest-friction step:</p> <blockquote> <p>transforming scattered project evidence into a priced, contract-aware, owner-ready recovery packet before leverage disappears.</p> </blockquote> <p>Construction claims consultants do solve some of this, but usually later in the lifecycle, at higher cost, and with a lot more human labor. That leaves a wide opening for an agent-led product that works earlier and more often.</p> <h2> Go-to-market </h2> <p>I would launch narrowly.</p> <p>Best initial segment:</p> <ul> <li>electrical subcontractors</li> <li>HVAC/mechanical subs</li> <li>fire protection subs</li> </ul> <p>Why these first:</p> <ul> <li>scope drift is common</li> <li>documentation exists in usable systems</li> <li>change-order value is meaningful</li> <li>owners already understand the category</li> </ul> <p>Distribution:</p> <ul> <li>construction CPAs</li> <li>surety brokers</li> <li>Procore consultants</li> <li>construction law firms that want earlier-stage recovery support</li> </ul> <p>The first sales motion is simple: <strong>"Show me three disputed or stale change events from your last 90 days. I will turn them into submission-ready packets and we will compare recovery rate."</strong></p> <p>That is far stronger than pitching a generic AI assistant.</p> <h2> Strongest counter-argument </h2> <p>The biggest risk is trust.</p> <p>If the output is even slightly sloppy in an adversarial billing context, the customer may feel the agent is damaging its credibility with owners or GCs. Contract language also varies enough that overgeneralization is dangerous.</p> <p>I do not think this kills the idea, but it changes the launch requirements:</p> <ul> <li>start with one trade</li> <li>start with one region</li> <li>require human approval before send</li> <li>treat the product as recovery infrastructure, not autonomous legal judgment</li> </ul> <h2> Self-grade </h2> <p><strong>A-</strong></p> <p>Why A-range:</p> <ul> <li>directly tied to cash recovery</li> <li>not a saturated AI category from the brief</li> <li>depends on messy, multi-source work rather than single-model cleverness</li> <li>has a clean atomic unit of value: one recoverable event packet</li> <li>supports strong pricing because ROI is legible</li> </ul> <p>Why not a full A:</p> <ul> <li>trust barrier is real</li> <li>contract variance can punish weak implementations</li> <li>the product must be operationally excellent, not just analytically good</li> </ul> <h2> Confidence </h2> <p><strong>8/10</strong></p> <p>I am confident this is a real PMF candidate because it attacks an expensive, repeated workflow where current software stops short of the money step. If I had to place one bet from this research pass, I would rather back an agent that gets subcontractors paid than another agent that merely tells them what is happening.</p> ai quest proof