Subcontractor Risk

AI subcontractor risk scoring: what actually gets measured, and what a score misses

The real inputs behind a subcontractor risk score, which ones automate, which ones mislead, and the signal almost nobody tracks: how a sub handles paperwork before the job starts.

On this page
  1. What is subcontractor risk scoring?
  2. Who is responsible for deciding whether a subcontractor is qualified?
  3. How do you score financial capacity when the subcontractor is private?
  4. Is bonding capacity a reliable proxy for financial strength?
  5. How do you tell whether a subcontractor’s insurance actually matches the contract?
  6. What does a subcontractor’s safety record actually prove?
  7. How do you verify licensing status and disciplinary history?
  8. Why is there no reliable past-performance record for subcontractors?
  9. How do workload concentration and backlog change the risk?
  10. What is the most predictive risk signal that almost nobody scores?
  11. Which subcontractor risk inputs can actually be automated?
  12. Where does a subcontractor risk score mislead you?
  13. How does Contracts Connected handle subcontractor risk inputs?
  14. Frequently asked questions

Automation can gather and check most of these inputs. It cannot supply the judgment that makes the number mean anything.

What is subcontractor risk scoring?

Subcontractor risk scoring is the practice of converting a subcontractor’s financial capacity, insurance adequacy, safety record, licensing status, past performance, and current workload into a comparable rating used to decide whether to award work and on what terms. Automation can gather and check most of these inputs. It cannot supply the judgment that makes the number mean anything.

The useful version of scoring is not really a number at all. It is a structured way of asking the same questions about every subcontractor, in the same order, against the same evidence standard, so that the answers are comparable across a bid list. The rating at the end is a summary of that work. It is not a replacement for it.

That distinction is where most scoring programs go wrong. A score nobody can decompose back into its inputs is worse than no score, because it carries the authority of a measurement without the traceability of one. When a project team asks why a subcontractor came back at 72 out of 100 and the honest answer is “the model said so,” the score has stopped being a risk tool and started being a liability.

Who is responsible for deciding whether a subcontractor is qualified?

On federal work, the prime contractor is, explicitly and in writing. FAR 9.104-4(a) states that “Prospective prime contractors are responsible for determining the responsibility of their prospective subcontractors.” The government does not perform this determination for you, and “they were already on our approved list” has never been an answer to it.

The Federal Acquisition Regulation also supplies something more useful than an obligation. FAR 9.104-1 sets out the general standards a responsible contractor has to meet, and it reads almost exactly like a scoring rubric written before anyone thought to put the word AI in front of one. To be determined responsible, a prospective contractor must:

Have adequate financial resources to perform the contract, or the ability to obtain them; Be able to comply with the required or proposed delivery or performance schedule, taking into consideration all existing commercial and governmental business commitments; Have a satisfactory performance record; Have a satisfactory record of integrity and business ethics; Have the necessary organization, experience, accounting and operational controls, and technical skills, or the ability to obtain them; Have the necessary production, construction, and technical equipment and facilities, or the ability to obtain them; and Be otherwise qualified and eligible to receive an award under applicable laws and regulations.

FAR 9.104-1, General standards

Read that as a risk model rather than as regulatory text: financial resources, schedule capacity measured against existing commitments, performance record, integrity, organizational and accounting controls, equipment, and eligibility. It holds up better than most commercial scoring schemes for one reason. It is written in terms of what a contractor must be able to do, not in terms of whatever data happens to be purchasable about them.

One federal requirement deserves separate mention because it is the single cleanest automated check in the entire model. Under FAR 9.405-2(b), a contractor is prohibited from entering into any subcontract in excess of $45,000, other than a subcontract for a commercially available off-the-shelf item, with a party that has been debarred, suspended, proposed for debarment, or voluntarily excluded, unless there is a compelling reason to do so. FAR 52.209-6 carries that duty into the subcontract clause itself. The exclusion records live in the System for Award Management, FAR 9.404 confirms they are publicly accessible, and checking them is fast, cheap, and unambiguous. This is the one input where a machine gives you a definitive answer.

How do you score financial capacity when the subcontractor is private?

Mostly you cannot score it directly, which is why nearly every model quietly substitutes proxies for it. The large majority of construction subcontractors are privately held, file nothing publicly, and have no obligation to hand anyone a balance sheet. What you receive is what they choose to give you, usually a company-prepared or reviewed statement rather than an audited one, and usually several months stale by the time it arrives.

The population-level data explains why this input matters anyway. According to the Bureau of Labor Statistics Business Employment Dynamics program, of private-sector establishments born in March 2013, 76.0 percent of construction establishments were still operating one year later, 53.9 percent at five years, and 40.1 percent at ten years. Construction had the sharpest first-year decline of any large industry group, a fall of 24.0 percentage points.

Those numbers say something counterintuitive that changes how you should weight age. Construction’s ten-year survival rate of 40.1 percent is actually higher than the 34.7 percent rate across all private-sector establishments. Construction risk is not spread evenly across a company’s life. It is concentrated violently at the front. A three-year-old subcontractor and a fifteen-year-old subcontractor are not two points on a smooth curve, they are different risk classes, and a model that scores years in business linearly is mispricing both.

What to actually collect: a current work-in-progress schedule, accounts payable aging, the identity of the bank providing the line of credit and whether it has changed recently, and supplier references, which are frequently more candid than bank references because suppliers get paid last and notice first.

Where a model misleads here: subcontractor failures rarely look like insolvency on paper. They look like a cash squeeze caused by a completely different project, one you have no visibility into, arriving three months after the financial statement you scored.

Is bonding capacity a reliable proxy for financial strength?

Bonding capacity is the strongest single financial proxy available, because it is the only input where an independent third party puts its own money behind an opinion about the subcontractor. A surety underwrites the balance sheet, the work-in-progress schedule, and the management team before agreeing to guarantee performance, and it pays if the contractor fails. That is a materially different quality of signal than a self-reported number.

The federal framing is worth knowing even when you are scoring subs on private work. FAR 28.102-1 implements the Miller Act, which requires performance and payment bonds for any construction contract exceeding $150,000. On the surety side, the Treasury Department publishes Department Circular 570, described by the Bureau of the Fiscal Service as “a complete list of companies that write or reinsure federal bonds,” most recently updated July 1, 2024. Each listed company carries a published per-bond underwriting limitation, and amounts above it must be covered by co-insurance or reinsurance.

Two caveats a scoring model will miss:

A bonding letter is not a commitment. It states what a surety believes a contractor could support under current conditions. It does not obligate the surety to issue any specific bond, and conditions change. Scoring a letter as though it were a guarantee overstates it.

Some capacity is government-supported. The Small Business Administration guarantees bid, performance, and payment surety bonds issued by certain surety companies, covering contracts up to $9 million for non-federal work and up to $14 million for federal work. A bond written under that guarantee is a real bond and the program exists for good reason. But a model that reads bonding capacity as pure private-market underwriting will misread partially guaranteed capacity as a stronger independent signal than it is. If bonding carries weight in your score, you need to know which kind you are looking at.

How do you tell whether a subcontractor’s insurance actually matches the contract?

You read the endorsements, not the certificate. A certificate of insurance is a summary document issued for information. It establishes that a policy existed on the day the certificate was produced. It does not prove that additional insured status was actually endorsed onto the policy, that a waiver of subrogation reaches your project, or that the coverage form is the one your subcontract requires.

On federal installations the obligation to hold that proof sits with the prime. FAR 52.228-5 requires that “The Contractor shall maintain a copy of all subcontractors’ proofs of required insurance, and shall make copies available to the Contracting Officer upon request.” Holding the certificate is not optional and neither is producing it.

This is the input where automation earns the most, because the underlying work is high-volume comparison against a known requirement: limits against required limits, policy dates against the period of performance, the named insured against the legal entity that signed the subcontract, and endorsement forms present against endorsement forms required. All of that is checkable, repeatable, and enormously tedious for a person to do across two hundred subcontractors. It is genuinely the strongest case for automation on this page.

Where it still misleads: a fully compliant certificate with every endorsement attached tells you nothing about whether the policy contains an exclusion that guts coverage for this particular scope. Residential work exclusions, height and story limitations, and specific means-and-methods exclusions all live inside the policy, not on the certificate. Aggregate limits are the quiet one. A general aggregate that looks comfortable against your $2 million scope is being consumed simultaneously by every other job that subcontractor is running, and nothing on the certificate tells you how much is left.

What does a subcontractor’s safety record actually prove?

Less than most scoring models assume, and for a structural reason rather than a quality one: for a large share of subcontractors, the data simply does not exist.

The recordkeeping rules explain the gap precisely. Under 29 CFR 1904.1, “If your company had 10 or fewer employees at all times during the last calendar year, you do not need to keep OSHA injury and illness records.” Above that, 29 CFR 1904.41 governs what actually gets submitted to OSHA electronically. Construction establishments with 20 to 249 employees submit information from Form 300A, the annual summary, because NAICS 23 Construction appears in Appendix A to Subpart E. Establishments with 250 or more employees also submit the 300A. Detailed case-level data from Forms 300 and 301 is required only from establishments with 100 or more employees in industries listed in Appendix B, and the only construction industry in that appendix is NAICS 2381, Foundation, Structure, and Building Exterior Contractors. Submissions are due by March 2 of the following year.

Put concretely: an eight-person specialty subcontractor keeps no OSHA records at all. A forty-person mechanical subcontractor keeps a log and submits only a one-page annual summary. Detailed incident data exists for a narrow slice of larger firms in one construction subsector. Any model that scores subcontractors on OSHA data is scoring the large ones on thin data and the small ones on nothing, then reporting the difference as risk.

OSHA is unusually direct about this on its own establishment-specific data page:

OSHA does not validate the employee or injury and illness counts reported by establishments. Concluding that establishments are the “most dangerous” or the “least dangerous” solely based on whether they have the highest or lowest rates from these data would be inappropriate.

Occupational Safety and Health Administration, Establishment Specific Injury and Illness Data

That is the agency that collects the data instructing you not to rank with it. A product that ranks subcontractors on scraped injury rates is doing the exact thing the source of the data says is inappropriate, and it is worth asking any vendor how they handle that.

Industry-level context also cuts against the intuition that construction is uniformly the dangerous trade. In the Bureau of Labor Statistics 2024 Survey of Occupational Injuries and Illnesses, Construction (NAICS 23) recorded 2.2 total recordable cases per 100 full-time equivalent workers, with 1.3 cases involving days away from work, job restriction, or transfer. The private-industry rate overall was 2.3 total recordable cases per 100 full-time equivalent workers. Construction’s recordable rate sits slightly below the all-industry figure. What differs in construction is severity, and a recordable-case rate does not measure severity at all.

The other common safety input is the experience modification rate, a workers compensation rating factor. An employer whose losses match the expected losses for its classification and payroll rates at 1.00, better-than-expected experience produces a factor below 1.00, and worse experience produces a factor above it. It is more useful than a raw injury count because it is normalized against expectations for that trade and size. It also has four limits worth naming: it lags, because it is built from an experience period that closes well before the current year; it measures claim cost rather than safety, so a subcontractor that manages claims aggressively can carry a good factor and a poor safety culture; rating systems differ by state, so factors are not cleanly comparable across state lines; and very small employers may not be experience-rated at all, which returns you to the same coverage hole.

How do you verify licensing status and disciplinary history?

Through the state board that issued the license, one state at a time. There is no national contractor licensing database, because there is no national contractor license. Licensing is administered state by state, and in a number of states by county or municipality on top of that, with different classifications, different renewal cycles, different disciplinary disclosure practices, and different lookup tools.

Disclosure depth varies widely. California’s Contractors State License Board, for example, offers a public license lookup that lets users verify license information including complaint disclosure. Other state boards publish status and expiration date and little else. If your scoring model treats “licensed: yes” as a uniform field across fifty states, it is flattening genuinely different amounts of information into one box.

What automates well: status, expiration date, classification, and the bond and workers compensation fields where a board publishes them. These are structured, public, and checkable on a schedule rather than once at onboarding, which matters because a license valid at prequalification can lapse mid-project.

What it misses, and this is the expensive one: whether the classification actually authorizes the scope you are buying. A subcontractor can hold a current, unblemished license in a classification that does not cover the work described in your subcontract. That is a real legal exposure, in some states an unlicensed-work exposure that can affect the subcontractor’s ability to enforce payment, and it reads as a perfectly clean score. Classification-to-scope matching is a judgment a person has to make against the actual scope language.

Why is there no reliable past-performance record for subcontractors?

Because the federal past-performance system does not cover subcontractors, and even for prime contractors it is closed to you. FAR 42.1502 requires that “Past performance evaluations shall be prepared for each construction contract of $900,000 or more, and for each construction contract terminated for default regardless of contract value.” Those evaluations assess the entity that held the contract. Subcontractors are not evaluated in their own right.

Access is restricted on top of that. FAR 42.1503 provides that during the period the information may support future award decisions, “the completed evaluation shall not be released to other than Government personnel and the contractor whose performance is being evaluated.” The regulation contemplates using past performance data within three years of contract completion, six years for construction and architect-engineer contracts. None of it is a resource a general contractor can query when deciding whether to award a subcontract.

The practical consequence is the most important sentence on this page: the only subcontractor past-performance data you will ever have is the data your own projects generated. That makes your internal project history the single highest-value input in the entire model, and it is the input most firms fail to capture in any structured way. Change order behavior and whether the changes were legitimate. Schedule adherence against the actual sequence, not the baseline. RFI quality. Whether they staffed the job the way they said they would at the buyout meeting. Whether the superintendent would take them again, which is frequently the most accurate single data point available and almost never written down anywhere retrievable.

If you are evaluating tooling for this, the relevant comparison is covered separately in best subcontractor prequalification software→. This page is about the method, not the market.

How do workload concentration and backlog change the risk?

They change it faster than any balance sheet does, and the federal standard is explicit that capacity must be judged against commitments already made. FAR 9.104-1(b) requires a responsible contractor to be able to comply with the required or proposed delivery or performance schedule, “taking into consideration all existing commercial and governmental business commitments.” A financially sound subcontractor with a fully committed backlog is a higher schedule risk than a thinner one with real capacity, and only one of those two facts shows up on a financial statement.

Two ratios do most of the work here, and they cut in opposite directions.

Your share of their revenue. Too small and you are not a priority when crews get pulled. Too large and their survival becomes tied to your payment behavior, which sounds like control and is actually concentrated exposure. If you are 60 percent of a subcontractor’s book and your project goes sideways, you have manufactured the counterparty failure you were trying to score for.

Their concentration with someone else. A subcontractor whose capacity is dominated by one other customer inherits that customer’s problems. You are underwriting a relationship you cannot see.

This input barely automates. Backlog is self-reported, there is no registry, and the numbers arrive from the subcontractor’s own work-in-progress schedule if they are willing to share one. A surety relationship is often the better path, because a surety sees the backlog it is bonding. Treat any workload figure in a scoring model as an assertion with a date on it, not a measurement.

What is the most predictive risk signal that almost nobody scores?

How the subcontractor handles paperwork before the job starts. It is the earliest behavioral sample you can get, and it costs nothing to collect because you are already collecting it.

The evidence is already sitting in your intake process. How many days from request to a complete prequalification package. How many rounds of correction before a certificate of insurance actually matched the requirement. Whether the W-9 came back with a legal name matching the entity on the subcontract or with the DBA someone typed from memory. Whether they read the insurance requirement or sent their standard certificate and hoped it cleared. Whether they asked one substantive question about the scope or signed without reading it.

This is not a soft signal dressed up as a hard one. FAR 9.104-1(e) requires a responsible contractor to have “the necessary organization, experience, accounting and operational controls, and technical skills.” Pre-award document behavior is the cheapest observable proxy for exactly that standard. A subcontractor whose office cannot produce a correct certificate of insurance in three weeks is showing you, at no cost and before you are committed, how that same office will handle pay applications, lien waivers, certified payroll, and closeout documentation for the next eighteen months. The office that struggles with a one-page form in July does not become the office that submits clean certified payroll every week in November.

It is also almost entirely automatable, and it is measured by nearly nobody, for a simple reason: it cannot be purchased from a data provider. It exists only if your own intake system records timestamps, correction cycles, and completeness at the point the documents arrive. Firms that instrument their intake get a proprietary predictive signal on every subcontractor they touch. Firms that do not are left scoring the same purchasable fields everyone else scores.

The subcontractor risk tier scoring rubric→ is the working tool that turns these inputs into tiers, and the federal subcontractor prequalification questionnaire→ is the intake instrument that generates the behavioral data in the first place.

Which subcontractor risk inputs can actually be automated?

The matrix below is the honest answer, input by input. “Automatable” means a system can collect and check it without a person reading a document. Reliability describes how much weight the signal can carry on its own.

SignalWhere the data comes fromAutomatableReliabilityWhat it misses
Financial capacitySubcontractor-supplied statements, WIP schedule, AP aging, supplier and bank referencesPartly. Intake and tracking automate; interpretation does notModerate. Usually unaudited and months staleA cash squeeze originating on a project you cannot see. Failures rarely look like insolvency on paper
Bonding capacitySurety letter from the subcontractor's surety; Treasury Circular 570 for the surety's own standingPartly. The letter is a document; the underwriting behind it is opaqueHigh for what it covers. A third party has money at riskThat a letter states capacity, not commitment. Also whether capacity is supported by the SBA guarantee rather than pure private underwriting
Insurance adequacy and endorsementsCertificates, endorsement forms, and the policy itselfYes, strongly. Limits, dates, named insured, and endorsement presence all check mechanicallyHigh when endorsements are actually reviewed, low when only the certificate isPolicy exclusions that gut coverage for the scope, and aggregate limits already consumed by the sub's other jobs
Safety record (OSHA data)OSHA establishment-specific 300A data; 300 and 301 case detail for a narrow subsetYes where data existsLow to moderate. OSHA states it does not validate the counts and warns against rankingEvery subcontractor under the reporting thresholds. Firms with 10 or fewer employees keep no records at all
Experience modification rateSubcontractor-supplied rating worksheet from its workers compensation carrier or state fundPartly. The number is easy to capture, the context is notModerate. Normalized to trade and payroll, which raw injury counts are notIt lags the current year, measures claim cost rather than safety, is not comparable across states, and may not exist for very small employers
Licensing status and disciplinary historyState licensing board lookups, one state at a timeYes for status, expiration, and classification where the board publishes themHigh for the fact of licensure, variable for disciplineWhether the license classification actually authorizes your scope. A clean license in the wrong classification scores perfectly
Federal exclusions and debarmentSystem for Award Management exclusion records, public per FAR 9.404Yes, completely. The cleanest check availableVery high. It is a definitive recordNothing about capability. Absence of exclusion is a floor, not a qualification
Workload concentration and backlogSubcontractor's WIP schedule and self-reported backlog; surety insight where availableBarely. No registry existsLow to moderate. Self-reported and datedConcentration with another customer whose problems the sub will inherit, and your own share of their book
Pre-award document behaviorYour own intake system: response times, correction cycles, completeness on first submissionYes, fully, if your intake records itHigh in practice, and proprietary to youNothing about financial depth. It measures the office, which is often the thing that actually fails

Read down the automatable column and the pattern is clear. The inputs that automate cleanly are the ones with a public register or a fixed requirement to compare against. The inputs that resist automation are the ones that depend on what the subcontractor chooses to tell you. A model weighted toward the first group is measuring what is easy to measure, which is not the same as measuring what matters.

Where does a subcontractor risk score mislead you?

A score misleads most when it converts a judgment into a number and then hides the judgment inside it. Every weighting decision in a scoring model is an opinion about what matters. Deciding that safety is worth 20 points and financial capacity is worth 30 is a defensible opinion, but it is still an opinion, and arithmetic performed on it afterward does not make it evidence. The number that comes out the far end looks objective in a way its inputs never were.

Five specific failure modes are worth naming plainly.

Scope blindness. A score is calculated about a company. Work is awarded for a scope. The subcontractor who tops your bid list may be excellent at what they usually do and wrong for this sequence, this site, or this delivery method. No general score encodes scope fit, and the most expensive subcontractor failures are usually competence mismatches rather than company failures.

Availability bias against small subcontractors. Larger firms generate more public data. They cross OSHA reporting thresholds, carry bonding relationships, and have the office capacity to return a complete package quickly. A model built on data availability will systematically rate small subcontractors as riskier for being small rather than for being risky. If your firm has small business or disadvantaged business participation goals, an unexamined scoring model works quietly against them.

Staleness. Financial statements are annual. Experience modification factors are built from closed experience periods. Backlog is a snapshot. A score computed in January describes a company that may not exist in the same shape in September, and the risk that actually shows up is usually the change, not the level.

Gaming. Any rubric you publish gets optimized against. Subcontractors who bid your work regularly will learn which fields drive the score and manage those fields. That is not dishonesty, it is a rational response, and it degrades the predictive value of exactly the inputs you weighted most heavily.

The failure you get is rarely the failure you scored. The subcontractor who fails you usually does so because a key foreman left, an unrelated project stopped paying them, or the estimator who priced your job misread the drawings. None of that is in any model.

What this page does not cover. This is a methodology reference, not legal or insurance advice. It does not interpret specific policy language, which requires a licensed broker or coverage counsel. It does not address SBA affiliation and joint venture rules, state-specific prequalification statutes, or the enforceability of any particular subcontract term. Federal requirements cited here are current as of the last updated date above and change. Confirm current requirements with qualified counsel and your contracting officer before relying on them for an award decision.

The right posture is to treat a score as a triage instrument. It tells you which subcontractors need a closer look and which questions to ask first. It does not tell you who to hire.

How does Contracts Connected handle subcontractor risk inputs?

Contracts Connected works on the half of this problem that is genuinely mechanical: getting the documents in, reading them, checking them against the actual requirement, and keeping them current. Certificates of insurance are checked at the endorsement level rather than the expiration date. Licenses are verified against the issuing board and rechecked on a schedule instead of once at onboarding. W-9s are validated against the legal entity on the subcontract. Every document is tagged to the subcontractor, the project, and the requirement it satisfies at the moment it arrives.

That produces the input almost no scoring model has: a complete, timestamped record of how each subcontractor actually behaves on paperwork before the work starts, generated as a by-product of collecting the documents you were going to collect anyway. The scoring judgment stays where it belongs, with the people accountable for the award. What changes is that the judgment gets made against a complete file instead of a partial one.

Our co-founder ran a large federal general contracting firm for 15 years, about 100,000 documents a year.

Frequently asked questions

What is subcontractor risk scoring?

Subcontractor risk scoring converts a subcontractor’s financial capacity, bonding, insurance adequacy, safety record, licensing status, past performance, and current workload into a comparable rating used to decide whether to award work and on what terms. The useful version is not a single number. It is a structured way of asking the same questions about every subcontractor with the same evidence standard, so the answers are comparable across a bid list.

Can AI automate subcontractor risk scoring?

Automation handles collection and checking well: verifying license status against a state board, checking federal exclusion records, comparing insurance limits and endorsement forms against the subcontract requirement, tracking expirations, and measuring how long a subcontractor takes to return a complete document package. It does not handle weighting, scope fit, or the judgment about whether a specific subcontractor can perform a specific scope. Those stay human.

Is OSHA injury data reliable for scoring subcontractors?

Only within limits OSHA states itself. OSHA does not validate the injury and illness counts establishments report, and it warns against concluding that establishments are the most or least dangerous based solely on those rates. Coverage is also thin. Companies with 10 or fewer employees keep no records at all under 29 CFR 1904.1, and most construction firms submit only the Form 300A summary rather than case detail.

Why is subcontractor past performance so hard to score?

Because no shared record exists. Federal past performance evaluations under FAR 42.1502 are prepared for construction contracts of $900,000 or more, and they assess the prime contractor rather than its subcontractors. FAR 42.1503 also restricts release of a completed evaluation to government personnel and the contractor whose performance is being evaluated. The only subcontractor performance data you will have is what your own projects generated.

Does bonding capacity prove a subcontractor is financially sound?

It is the strongest available proxy, because a surety puts its own money behind the opinion, but it is not proof. A bonding letter states capacity, it does not commit the surety to any specific bond. Capacity can also be supported by the SBA Surety Bond Guarantee program, which backs bonds up to $9 million on non-federal and $14 million on federal contracts, so the signal is not purely private underwriting.

What does a subcontractor risk score fail to catch?

Scope fit, mainly. A subcontractor can score well on every input and still be the wrong choice for a specific scope, sequence, or site. Scores also favor larger subcontractors, who simply generate more public data, and they lag reality because most inputs are annual or self-reported. A score is a summary of the questions you asked, not a prediction of how the work will actually go.

Contracts Connected is the Back-Office System (BOS) for construction.

The AI agent is in beta. We are accepting teams now.

SOC 2 Type II · U.S. hosted · AI optional
Join the agent beta