AI is becoming a control system

Enterprise AI is acquiring an audit trail, and the trail is starting to look like the product.

Most enterprise AI pitches we see still open with capability: what the model reads, writes or decides. But the buyer’s question has moved on. They want to know who gave the system permission to act, which source it used, what it did when it wasn’t sure, and whether the answer can be reconstructed after the model, the prompt and the underlying rule have all changed.

Seven announcements between 19 and 22 October point the same way. A regulator asked firms to make AI behaviour observable. An industry body attached provenance to executable rules. A startup put compliance checks in the path of each agent action. A data provider followed its users into a general-purpose assistant. Two funding stories rewarded governed workflows over free-form chat. Most of the figures below come from the companies themselves and haven’t been independently verified; I’ll say that once here rather than after every number.

The FCA wants evidence

On 19 October the Financial Conduct Authority published examples of good and poor AI practice. There is no new AI rulebook in it. It takes existing expectations and applies them to systems that drift, change and act with less supervision. The practical asks are specific: define where a system must abstain or escalate, monitor live outcomes, retain prompts and outputs, record model versions and filters, and demonstrate that the controls work in practice.[1]

A policy document can’t satisfy that. The evidence has to come out of the operating system around the model.

In the accompanying speech the FCA also said it is building agentic supervisory tools for roughly 9,000 small wholesale-market transaction-reporting firms, plus machine-readable data-collection APIs on top of its Handbook API.[2] A speech is a statement of intent, not a working system. But if supervision becomes continuous and machine-readable, firms will need controls that answer in the same language — structured permissions, event logs, outcome monitoring, exception histories.

For a founder selling into a regulated firm, “we use a better model” is a fragile line. “We can show what the model did, under whose authority, and what happened next” maps onto a budget that survives model upgrades.

Rules with the reasoning attached

ISDA’s Digital Regulatory Reporting project already turns agreed interpretations of reporting rules into executable code. On 20 October it added a Tracer Agent, built with Gentek AI, that lets users search the history behind each coding decision and trace it back to the regulatory text.[3]

Code and the reasoning behind it are different assets. Rules change, interpretations get contested, and an implementation can be correct yet indefensible because the memory of why it was done that way sits in old meeting minutes or with people who have left. ISDA is trying to keep that chain intact.

A day later Solcis, a 6clicks spin-out, launched a self-hosted Compliance Studio and Policy Engine that converts obligations into policies, checks agent actions inline, returns Allow / Deny / Escalate, and stores tamper-evident evidence.[4] The architecture is the interesting part: compliance stops being a review after the workflow and becomes a decision inside it.

A runtime layer sitting between agents and tools can outlive any particular model. Whether it’s a moat depends on latency under load, false-escalation rates, how accurately obligations get authored into policies, integration cost, and who carries the liability when the engine approves the wrong action. None of that is public yet.

Content follows the user

Lexology shipped an MCP connector for Claude on 22 October. Subscribers can query regulatory updates and analysis inside the assistant, with links back to source. The broader PRO product claims 1,200+ sources across 170 jurisdictions; the connector exposes only updates and analysis for now.[5]

This is a distribution decision. Users want specialist data in the window where they draft and decide, so the provider’s own interface matters less and entitlement, freshness, provenance and coverage matter more.

It is also bad news for startups whose product is a polished retrieval chat on top of someone else’s corpus. Once the corpus is reachable through an API or MCP connector, the chat layer has a short shelf life. What’s left to defend is the maintained obligation graph, customer-specific applicability logic, source rights, correction history and write-back into systems of record.

Capital is backing governed workflows

Pigment said on 21 October that it raised $250m at a $2.3bn valuation, having passed $100m ARR, and that agents now create 85% of new metrics on its planning platform.[7] Investors aren’t paying for a finance chatbot. The pitch is a governed data model, deterministic computation and traceable actions.

Terminal X is the early-stage version of the same bet, aimed at investment teams. An OpenAI case study from 20 October describes agents built around each firm’s own process, every output graded against a 54-criterion rubric, 88.6% agreement with senior reviewers over 500 scored judgments, and $2.4m annualised savings at one customer.[6]

Behind the case-study numbers the mechanism is simple: a firm-specific evaluation layer turns judgment into tests, corrections into reusable memory, and each answer into an auditable event. That is hard to swap out. It also appears to take weeks of forward-deployed engineering per customer, which raises the question of whether the result is software, consulting, or a mix that can stay margin-positive.

What we’re looking for

We want products where the evidence gets better with use instead of piling up as dead logs: the system learns from overrides, ties corrections to specific policies, replays old decisions against a new model and shows whether outcomes improved.

For us it means more than guardrails and constant model prompt updates, more than agent harnesses fixing themselves, updating their skills and tooling. This means memory that mutates where it needs to adapt, yet stays solid where it needs to remember things clearly. It means provable causality and traceable reasoning, data and reasoning provenance at scale.

The questions we now weight most heavily in diligence:

• Does the product control an action, or only describe what should happen?
• Can a customer reconstruct source, permission, model version and decision path for any given output?
• Do corrections become tests that stop the same failure recurring?
• Does the control layer survive a change of model, agent framework or system of record?
• How many weeks of human implementation before the software is useful on its own?

Models will keep getting cheaper and better. That doesn’t make application companies redundant; it moves the durable part of the application away from generated text and towards permissions, evaluation, provenance and recovery.

Source notes

[1] https://www.fca.org.uk/publications/good-poor-practice/ai-financial-services — AI in financial services: good practice and areas for improvement
[2] https://www.fca.org.uk/news/speeches/building-future-finance-age-ai — Building the future of finance in the age of AI
[3] https://www.isda.org/2026/10/20/isda-launches-ai-driven-traceability-tool-for-its-digital-regulatory-reporting-solution — ISDA launches AI-driven traceability tool for DRR
[4] https://solcis.ai/blog/introducing-solcis — Introducing Solcis: the open compliance layer for AI agents
[5] https://www.centellic.com/news/lexology-brings-regulatory-intelligence-to-claude-with-lexology-pro-mcp — Lexology brings regulatory intelligence to Claude with Lexology PRO MCP
[6] https://openai.com/index/terminal-x — Terminal X matches senior investor judgment 88.6% with GPT-6.1
[7] https://www.prnewswire.com/news-releases/pigment-raises-250m-at-2-3b-valuation-302914892.html — Pigment raises $250M at $2.3B valuation