AI meets deterministic systems: rules, workflow, and compute
Evidence: Public benchmarks, surveys, standards, and vendor material. The framing is my view. I invest in Tracfox.
Business and regulators need results that are accurate, explainable, and repeatable. AI on its own falls short of that in three ways, and each shows up differently by business domain:
- Rules. It reads contracts and regulations well, and reasons over them poorly. Its own account of how it reached an answer isn't a reliable record. When reasoning models used a hint, they mentioned it 25% to 39% of the time (Anthropic, Apr 2025).
- Workflow. A procedure that's right in month one drifts by month six.
- Compute. Tools are solving the arithmetic. Building and navigating the financial model is the harder part.
What it means by domain
- Compliance. Which rules apply, which control meets each, and what an examiner can walk back through. The scarce asset is the mapped chain from rule to control to evidence. Buyers can ask for that lineage, rather than for answers.
- Syndicated lending. Covenant tests, amendments, and consents. Whoever structures the agreement data owns the workflow that runs on it. Amendments are the risk: the definition changes, and the test has to follow.
- Private credit and funds. Side letters, borrowing base eligibility, and investor reporting, each negotiated one deal at a time. Little shared structure exists, so firms build it deal by deal or go without.
- Capital markets: trading. Trade-by-trade compliance across jurisdictions. Shared standards are furthest along here (ISDA's Common Domain Model), so the advantage shifts from owning the structure to running it.
- Insurance. Claims against policy wording, run through several generations of systems. 75% of life and annuity insurers run 4 or more admin platforms for closed blocks (EY Closed Blocks Survey, 2023). The rules often live in code, so mapping them out of the code comes before automating.
- Office of the CFO. The close is a procedure problem more than a math problem. Planning engines are already held by incumbents adding agents.
Where the types meet
Real decisions cross types, and the risk collects at the handoffs. A rule changes and the procedure doesn't. A procedure runs and calls the wrong number. Each tool's record stops where the next tool starts, and few vendors cover more than one type.
Open question: in your domain, which type is the hard one, and who owns the handoff?
Email me to discuss.
Based on a longer analysis.
From the work on AI and deterministic systems.