Surety demand has outrun the desks that underwrite it
Two 2026 developments have pushed Indian surety volumes past the point where manual credit judgement can keep up. IRDAI's 2026 surety operational framework tidied up issuance mechanics and encouraged insurers to treat surety as a standing line rather than an experiment, building on the earlier removal of the solvency margin loading and the 30% aggregate exposure cap. In April 2026 the Ministry of Power allowed contractors to submit surety bonds in place of bank guarantees for bid and performance security across its projects, a signal that flows into transmission, thermal and renewable tenders where the ticket sizes are large and the timelines are tight.
The demand side is straightforward. Contractors want their bank guarantee limits back so working capital is not locked as margin money against non-fund-based lines. Employers, from NHAI and the Railways to state discoms, are increasingly willing to accept surety. What has lagged is the insurer's ability to say yes quickly and at a sensible price.
Surety is closer to credit underwriting than to property or engineering underwriting. The insurer is not pricing a physical peril, it is estimating whether a contractor will complete a contract and, if not, what it will cost to make the obligee whole. That requires a view on the contractor's balance sheet, order-book concentration and execution history, assembled fast enough to meet a tender deadline. Doing this by hand for a widening pipeline of mid-market contractors is where desks stall. Machine-learning default prediction is entering exactly this gap, not to replace the underwriter's sign-off but to give the desk a ranked, evidenced starting point.
What a contractor default-prediction model actually scores
A surety default model estimates a probability of default over the bond tenor and, alongside it, a loss-given-default if the bond is invoked. The two together drive whether capacity is offered, how much, and at what rate. This is a different question from a lender's one-year PD, because a performance bond can run for the construction period plus a maintenance period, so survival models that account for time to default tend to fit better than a single-point score.
The features are contractor-specific rather than project-specific. Financial-statement ratios do the heavy lifting: current ratio and working-capital adequacy, debt-to-equity, interest coverage, receivable days and the trend in operating margin. On top of those sit execution signals, the ratio of order book to annual turnover (an over-stretched book is a classic default precursor), sector and geographic concentration, the share of arbitration-locked receivables, and the promoter's track record across earlier entities. Vendors such as Underwrite.ai have shown that gradient-boosted and ensemble methods on this kind of data can separate defaulters from performers well ahead of a filed default.
The output that reaches the underwriter is not a black-box number. A usable model returns a score band, the top features pushing the score up or down, and a peer comparison against similar contractors in the same trade. A civil contractor with a three-times order-book-to-turnover ratio and slipping interest coverage reads very differently from a mechanical contractor with a conservative book and clean coverage, even at the same headline turnover. The model makes that contrast explicit and repeatable, which is what a growing surety line needs from its first-pass triage.
The Indian data trail a surety model can actually read
A default model is only as good as the data an Indian contractor leaves behind, and that trail is now rich enough to be useful if it is stitched together properly. GST returns (GSTR-1 and GSTR-3B) give a near-real-time read on turnover and its trajectory, harder to dress up than annual accounts. MCA and ROC filings supply audited financials, the register of charges (how much is already pledged to lenders), and any strike-off or default flags on the entity and its directors.
Project and payment history is the second layer. GeM order records, and award and progress data from NHAI, CPWD and the Railways, evidence what the contractor has actually executed and whether it finished on time. Bank-statement analysis pulled with consent through the Account Aggregator framework shows real cash-flow behaviour, cheque returns and the working-capital cycle, which financial statements smooth over. Commercial credit bureau data from CIBIL Commercial or CRIF adds repayment conduct on existing facilities.
The risk layer completes the picture: NCLT and IBC filings, GST cancellations, and litigation databases flag the contractor moving toward distress before an audited loss appears. The engineering challenge is not access, it is reconciliation, matching a GST number to a PAN to a CIN to a bureau record so the same contractor is not scored twice or missed entirely.
Turning the score into faster bid, performance and advance-payment decisions
Speed is the commercial point of AI in surety, because bond demand is deadline-driven. A contractor bidding a tender needs a bid bond within days, and if the insurer cannot respond in that window the business simply goes to a bank guarantee. A default model lets the desk run a first-pass triage the moment a request lands: a clean high-score contractor within existing capacity routes to near-automatic issuance, a borderline case goes to an underwriter with the red flags already surfaced, and a weak case is declined or referred with reasons attached.
The three main bond types carry different risk and should be priced differently, which a model makes tractable at volume. A bid bond is low-exposure and short, calling only if the contractor wins and then refuses to sign. A performance bond, typically 5% to 10% of contract value, runs the full construction period and is where default risk concentrates. An advance-payment bond is the sharpest, because the obligee has already paid cash out and the bond has to return it, so it warrants the tightest score threshold and the strongest counter-indemnity.
The score feeds the commercial structure, not just the yes or no. It informs the premium rate, whether a sub-limit or reduced bond percentage is warranted, how much counter-indemnity or collateral to require, and whether to cap tenor. Used this way, a desk can widen its contractor base without loosening its risk appetite, saying yes faster to the strong names and pricing the marginal ones honestly rather than turning them away for lack of time to assess.
Guardrails: IRDAI rules, the Contract Act, and defensible declines
AI in surety underwriting sits inside two bodies of law that a model cannot override. The insurance side runs through IRDAI's surety guidelines, first issued in 2022 and progressively eased so that the solvency loading and the aggregate exposure cap no longer choke capacity, and then organised by the 2026 operational framework. IRDAI's broader position on AI in underwriting expects insurers to keep a human accountable for the decision and to be able to explain it, so a default score is an input to underwriting, not the underwriting itself.
The contractual side runs through the Indian Contract Act, 1872. A surety bond is a contract of guarantee under Section 126, and the surety's rights on payment, including subrogation against the principal and the counter-indemnity that lets the insurer recover, flow from Sections 140 and 141. A model that scores default probability does nothing to change these mechanics, which is why the counter-indemnity and recourse structure still has to be underwritten deliberately even when the score is strong.
Explainability is the practical guardrail. When a contractor is declined or priced up, the insurer should be able to point to the features that drove it, audited financials, order-book stretch, a charge on assets, rather than an opaque number, both to defend the decision and to avoid quietly locking MSME contractors out of the market the framework was meant to open. Model-risk governance, periodic back-testing against realised defaults, and monitoring for drift as the contractor population shifts are the same disciplines a lender applies to a credit model, and surety desks adopting AI need them from day one rather than after the first invoked bond.
Where policy-wordings intelligence sits next to the score
A default score decides whether to issue a bond and at what price. It says nothing about what the bond actually promises once it is issued, and that is where more surety disputes are lost than on credit alone. Whether a bond is on-demand or conditional, what documents an obligee must produce to invoke it, how the maintenance-period obligation is worded, and how the insurer's recourse against the contractor is framed all live in the bond wording, and they vary materially between insurers.
A strong contractor on a badly worded on-demand bond can still generate a wrongful-invocation fight, while a weaker contractor on a tightly conditioned bond with clean counter-indemnity may be the safer exposure. The score and the wording have to be read together, and a broker placing surety needs both views before the bond is issued rather than after it is called.
This is also why comparing wordings across insurers matters as surety capacity widens. The market is still young in India, and the bond forms are not standardised, so two insurers quoting the same contractor can differ sharply on invocation conditions, notice periods and the extent of the contractor's release on completion. A default score narrows the field of acceptable contractors, but the wording decides which insurer's bond a broker should actually recommend for a given tender.
Sarvada gives commercial insurance brokers structured, searchable access to insurer surety and bond wordings and the intelligence around them, so the contractors a default model clears can be matched to the bond terms that will actually respond when a project slips. Request Access to place the credit view and the wording view side by side before the bond is issued.