Dili's $21.7M bet: use AI to parse the paperwork, then let deterministic code make the compliance call

Dili's $21.7M bet: let AI read the paperwork, let code make the call

Dili, an AI compliance company that has raised $21.7 million, most recently a $15M Series A led by Khosla Ventures, has built its system around a deliberate split. Large language models handle the fuzzy work of reading messy regulatory documents. Deterministic code makes the compliance judgment calls.

For anyone building AI tools for regulated industries, that separation is a trust design pattern worth studying: let the model perceive, let code decide.

The problem Dili targets is specific and expensive. Federally funded construction projects in the United States must comply with overlapping layers of regulation. Davis-Bacon rules allow the Department of Labor to set prevailing wages for certain projects 1. Clean energy projects funded under the Inflation Reduction Act face a separate set of prevailing wage and apprenticeship requirements 1. OSHA and EPA rules add further layers depending on the work involved 1. Anand Chaturvedi, Dili's co-founder and CEO, says violations on these projects can trigger fines in the millions 1.

Compliance teams have traditionally managed this burden by sampling: reviewing a subset of payroll reports, spot-checking classifications, trusting the rest is clean. Dili's system is designed to check all information as it comes in, rather than sampling 1. The engineering decision that makes this feasible is where to draw the boundary between probabilistic AI and deterministic logic.

Dili places LLMs in its data layer only, where they take unstructured documents and translate them into structured records 1. From there, a deterministic engine applies the compliance rules. Does this worker's wage match the prevailing rate for this county? Does this contractor's apprentice-to-journeyman ratio meet IRA requirements? These are binary questions. They do not require a language model to answer them.

Hallucination has been the central obstacle for AI in compliance work. When TechCrunch profiled Dili in its earlier form as a due diligence platform in 2024, the concern surfaced immediately: Fast Company testing had found that ChatGPT tended to get details wrong, omit information, or invent content when summarizing articles 2. Dili's response then was to fine-tune its models. Its response now is structurally different: keep the model out of the decision entirely.

The deployment numbers suggest the approach is gaining traction. Chaturvedi says Dili is in use across about 700 projects, from manufacturing facilities to data centers 1. Roughly half of those projects use Dili as in-house software, while the other half outsource the full compliance process to the company 1. The company says tasks that once required a full day of manual review can now be completed in minutes 1.

The $15 million Series A, announced on July 30, was led by Khosla Ventures, with participation from Allianz, Rebel Fund, Brick and Mortar Ventures' Darren Bechtel, and Y Combinator's Garry Tan 1. It brings Dili's total funding to $21.7 million, counting a prior $6.7 million seed round 1. Dili was part of Y Combinator's Summer 2023 batch 3.

The implication extends beyond construction. Dili's architecture points to a different template: let the model do what language models do well, which is perceiving patterns in unstructured text, and let deterministic code do what code does well, which is applying rules without ambiguity. For regulated industries where an error means a fine or a failed audit, that separation may be the difference between a tool people trust and one they cannot afford to deploy.

Chaturvedi anticipates the market shifting further toward software over outsourced services as AI capabilities advance 1. If the architecture holds, that transition gets simpler. A deterministic engine does not hallucinate. An LLM parsing a payroll report does not need to.