What happened: A WIRED-sponsored piece profiles Kyndryl’s “policy-as-code” approach: use AI to translate big, messy rulebooks into decision-tree workflows, then lock them into binary, machine-readable steps that humans can review and approve.
Why it matters: Regulated work hates randomness. The pitch is to use LLMs for drafting and decomposition, but keep execution deterministic and auditable—so “the model” can brainstorm, while “the policy” decides what actually runs (welcome to bureaucracy, now with unit tests).
Wider context: This is governance becoming the product: permissions, logs, tool-gating, and “who’s accountable when the bot goes weird” are turning into the real adoption bottlenecks—and therefore the real enterprise spend.
Background: The UK’s Financial Conduct Authority points out the core tension: generative AI is non-deterministic, which creates headaches for explainability, auditability, and responsibility. The proposed workaround is to let AI help build the rules-engine, not be the rules-engine.
I, Rulebot: How AI is turning compliance into software — WIRED (sponsored by Kyndryl)
Singularity Soup Take: The “agentic future” isn’t arriving as a single glorious robot brain—it’s arriving as a control layer that says no. The winners will be whoever can turn policy, identity, and audit logs into a boring, enforceable switchboard that enterprises can actually live with.
Key Takeaways:
- Determinism as a safety belt: Kyndryl’s approach uses LLMs to break policy text into steps, but enforces binary rule execution (“if X then step 2”) to reduce the unpredictability that makes regulators and risk teams lose sleep.
- Humans still sign the paperwork: The decision-tree workflow is reviewed and approved by a human, and where “agents” are used, tool and data access can be tightly parameterised, with optional human review before outputs are accepted.
- Compliance is the beachhead: The article argues compliance workflows are expensive and manual (“swivelchair” work across forms and databases), making them a prime target for automation that also increases observability and auditability.