§ 01 · The 97% problemThe dollar that doesn't know what it bought.
Software is 2-3% of every dollar spent in the asset-heavy economy. The other ninety-seven — materials, labor, machinery, working capital — runs on spreadsheets, tribal knowledge, and 30-day-late variance reports. That's the surface this report is for.
The collapse of that 97% into anything resembling a queryable, real-time decision surface is the largest software opportunity of the next decade. Not because operators want software. Because the math has finally outgrown the tooling.
"I had no idea this happened until three months too late." — CEO of an operator with $30M+ in annualized revenue, describing $150K+ of unflagged margin loss in a 90-day window. The systems were all in place. None of them spoke to each other.
This isn't a software-adoption problem. It's a system-of-record problem. CRMs document pipeline. ERPs record cost. FSMs coordinate trucks. Accounting closes the books. None of them, on their own, can answer the CFO question: which of our 300 customers is actually profitable?
§ 02 · The substrate gapThe unit nobody canonicalized.
Every asset-heavy operator runs the same six-to-twelve systems. Salesforce or HubSpot for CRM. NetSuite or SAP for ERP. ServiceTitan or Jobber for FSM. QuickBooks or Sage for the ledger. A dispatch board. A contract repository. Asset registries. Maps. Identity. Each system is competent on its own. None of them share an ontology.
The missing object is the address. Not the customer. Not the contract. Not the SKU. The address is the gravitational point where every commercial decision routes — a depot, a panel, a pole, a meter, a unit door, a chiller, a charger, a parcel locker. Bundle 600 of them across one operator and you have a unit-economic graph. Bundle them across a thousand operators and you have the substrate the next decade of operational tooling will be built on.
Canonicalization before prediction
Pretty charts over an unreconciled graph create false confidence. Most internal builds die in entity resolution — long before any ML question matters.
Score → score → act
Retrieval + classical + generative, in that order. Every output is machine-readable AND operator-inspectable. Black-box recommendations don't survive a CFO review.
Task-specific agents
Reprice this account. Escalate this renewal. Bundle this add-on. Not chatbots. Narrow agents that read from a canonical graph, dispatch from state changes.
Outcomes write back
Every closed job sharpens the next decision. Most operators run open loops — work done, outcomes never measured back. Closing the loop is what makes the system self-improving.
§ 03 · Seven verticals · underwriting laddersThe math per industry.
The substrate is universal. The math on top of it is industry-specific. Below: the canonical 5-6 levers that drive composite margin for each of seven asset-heavy verticals, derived from operator interviews and design-partner data. Each links to the detailed underwriting ladder.
| Vertical | Composite KPI | Top 3 levers | Ladder |
|---|---|---|---|
| HVAC service | Per-call gross margin · 38.4% | Labor hours · parts margin · truck-roll utilization | → Ladder |
| Fiber / Telecom | Project IRR · 14.0% | Build cost / HP · take rate · ARPU mix | → Ladder |
| EV charging | 5-yr site IRR · 12.4% | Build cost / stall · utilization · demand charge | → Ladder |
| Distributed solar | Per-project margin · 22.3% | $/W installed · take rate · 25-yr O&M | → Ladder |
| Construction | Project gross margin · 10.8% | Bid win rate · CO capture · sub-default risk | → Ladder |
| Industrial IoT | Composite deployment IRR · 38% | Sensor cost · uptime · predictive→reactive ratio | → Ladder |
| Logistics | Lane contribution margin · 22.4% | Cost / pallet-mile · empty miles · dock turn | → Ladder |
Three patterns emerge across the seven:
First: the composite KPI is always 5-6 levers deep. A single-number margin headline collapses real operational risk. Operators that compete only on the composite end up surprised by the levers underneath.
Second: the lowest-decile slice of any operator's book is what drags the composite. The opportunity isn't a 5-point lift across all sites — it's a 30-point lift on the bottom 10% by walking, repricing, or repositioning them.
Third: every vertical's math is more similar than different. The data sources differ (BLS for HVAC labor, EIA for diesel, FRED for capital, OSM for routes). The shape of the underwriting decision is identical. Which is why a single substrate generalizes.
§ 04 · The agent layerTask-specific. Not chatbot.
The 2024-2025 wave of "AI for ops" produced two patterns that operators bounced off of. The chatbot pattern — sidebar "ask the AI" interfaces — failed because operators don't have time to ask questions; they need answers dispatched into their existing workflow. The black-box prediction pattern — opaque scores with no rationale — failed because no CFO signs off on a recommendation they can't explain to the board.
The pattern that's working in 2026 is narrower and more boring. One agent. One job. Cited rationale. Dispatched from state changes. Examples that work in production today, derived from our design-partner deployments and partner interviews:
When cost drifts > threshold
Live cost engine detects 11% supplier drift over 90 days. Agent generates refreshed quote with margin delta and SOP citation. Sales reviews; signs; ships.
When churn signal > threshold
Customer health + contract calendar surface a renewal at risk 90 days early. Agent flags the owner with context + suggested play. Owner doesn't discover at week-zero.
When expansion signal fires
New addresses, service-pattern change, or asset-utilization signal. Agent surfaces a personalized expansion play with LTV math attached.
Pattern-match against closed-won
Graph recognizes this account looks like account X that closed via owner Y's pattern. Routes accordingly. Inboxes don't fill up.
The shape of the operator's productivity gain isn't "AI does my job." It's "the system makes the next decision before I have to look for it." Agents become the muscle memory of the operating layer.
§ 05 · Where capital is headingFrom CPQ to decision substrate.
The category formerly known as "CPQ" — configure, price, quote — was built for software sales orgs configuring line items with SKU rules. It does not fit asset-heavy operations. Yet most of the venture capital deployed in 2023-2024 went there: Salesforce CPQ, PROS, Vendavo, all targeting B2B distribution + manufacturing.
The next decade of capital will move toward what's actually different: address-level underwriting. Decision substrates that score the unit (the address, the asset, the job) before the line item. The early signal in 2026:
Mid-market is the frontier. F500 operators have PROS and Vendavo. Sub-$1M operators have spreadsheets. The $20M-$300M asset-heavy mid-market has neither — and that's where the substrate is being built, design partner by design partner.
Agentic surfaces become standard. Both vori and gaiia (single-vertical operating systems) shipped agent-builder UIs in 2025. By end-2026, "operators can build their own agents" will be table stakes for any new operational platform — not a differentiator.
The data graph is the moat. Per-operator canonicalization is fixed cost. Once a substrate covers a region (or a vertical), competitors can't replicate it from a cold start. The graph compounds with every operator onboarded — and the first to scale captures the network effect.
Capital underwriting collapses into operational underwriting. Lenders, insurers, and infrastructure funds increasingly want address-level signal. The same graph that prices a quote can underwrite a debt facility. By 2028, the line between "ops software" and "capital-allocation software" disappears in this segment.
§ 06 · Methodology + sourcesHow the numbers were derived.
This is v1.0 of the report. Numbers in the underwriting ladders are composite benchmarks derived from: (a) operator interviews with 14 mid-market asset-heavy operators across the seven verticals between Q4 2025 and Q2 2026, (b) anonymized aggregated data from our first signed design partner, (c) published public-data sources (EIA, BLS, FRED, Statistics Canada, DAT, US Census), and (d) standard industry-report benchmarks where independent operator data was not available.
Figures are directional and composite, not contractual. The same lever (e.g. "build cost per home passed") will vary by an order of magnitude between a rural overbuilder and an urban incumbent. The ladders shown here are calibrated to the mid-market median — the operator a $20M-$300M PE roll-up or a regional infrastructure fund is most likely to evaluate.
v2.0 of this report — Q4 2026 — will incorporate data from a full design-partner cohort (target: 10 operators) plus partner-supplied benchmarks from FSM, ERP, and pricing-tool vendors who choose to contribute under an MNDA. Methodology will be re-stated openly. Composite math will be reproducible.
Reports about operating economics are routinely wrong on the order of magnitude. This one will be wrong on some details. The goal is not perfect math — it's an honest shape of the mid-market, and a set of frameworks operators can argue with.