AI
AI governance starts with a baseline someone signs

AI governance is the owners, limits and records showing who approved each AI system, what it may do and which number it should move. It runs on four signatures: a baseline Finance signs, an owner per system, approval limits, an audit trail. Future Works stakes a portion of its fee on what that baseline shows by week 12.
Most governance programs start at the top: a responsible AI policy, a council that meets monthly, a register of use cases in a shared folder. Each of those has a place, and the trouble starts below them. I wrote in August that policy is usually espoused clearly, with a steward named, a standard documented and an approval chain on a slide, while enforcement turns out to be manual. That gap, governance espoused versus governance enacted, applies to AI systems as much as to the data they read.
What is operational AI governance?
Gartner put a name to the gap on September 15, 2026, in its strategic predictions for 2027 and beyond. Organizations, it said, are "shifting from policy-based AI governance to operational governance that embeds controls directly into AI systems and workflows." The same release predicts that by 2030, 80% of Global 500 companies will contractually make their CIO or Chief AI Officer the "Evidence Custodian" for AI accountability. Evidence is the word I would underline. A committee produces minutes, and an AI system in production has to produce evidence every day that a named person is in charge of it and that it does what it was approved to do.
| What governance covers | On paper | In the run |
|---|---|---|
| Ownership | A committee or a center of excellence owns "AI" | Each AI system has one named owner who can pause it |
| Value | A business case slide with a projected benefit | A baseline Finance signed before the build, re-measured at a set date |
| Limits | An acceptable-use policy | Written approval limits: what the system decides alone, what goes to a person |
| Evidence | Quarterly minutes and a use-case register | A log of what the system did, on which version, and who approved each exception |
| Review | An annual policy refresh | A check every cycle, with agreed conditions for switching the system off |
The right column is more work, and it is less comfortable. A named owner can be blamed. An approval limit, once written down, will at some point stop an agent from finishing a job somebody wanted finished. A baseline signed by Finance means the business case can be checked later and found wrong.
What do NIST AI RMF and ISO/IEC 42001 ask for?
The public frameworks ask for much of the right column. NIST released its AI Risk Management Framework (AI RMF 1.0) on January 26, 2023. Under its Govern function, it asks that roles, responsibilities and lines of communication for managing AI risk are "documented and are clear to individuals and teams throughout the organization." It asks executive leadership to take responsibility "for decisions about risks associated with AI system development and deployment." It also asks for policies that define roles "for human-AI configurations and oversight of AI systems," which in an agentic workflow means writing down what the agent may do on its own and where a person decides. NIST notes on its site that the framework is being revised under the White House AI Action Plan, so check the current text before you map controls to it.
ISO/IEC 42001, published in December 2023, sets requirements for establishing, implementing, maintaining and continually improving an AI management system. ISO describes it as the world's first AI management system standard and lists traceability and transparency among its benefits. Both ask for defined roles and records that can be inspected, which a committee can oversee but cannot produce. The principles side is covered separately.
How did the AI Omnibus change the EU AI Act deadlines?
The EU AI Act has been in force since 1 August 2024, and its prohibited practices have applied since 2 February 2025, according to the European Commission. This year the AI Omnibus, in force since 27 July 2026, moved the high-risk dates: obligations for high-risk systems in the Act's Annex III areas now apply from 2 December 2027, and high-risk AI embedded in regulated products has until 2 August 2028. If you sell into the EU or operate there, check which of your systems fall in the Annex III areas, because those are the ones the December 2027 date covers. The regulated-industry detail is in a separate guide.
The extra time is easy to spend on another round of policy drafting. The better use is getting the right column working on the systems you already run, because an audit trail builds up as a system runs and cannot be reconstructed after the fact. Agent deployment is moving faster than governance. At its IT Symposium keynote on October 19, 2026, Gartner reported a Q2 2026 survey of more than 2,000 CIOs and technology executives: 71% of organizations expect to deploy agentic AI within the next 12 months, and only 21% have agentic AI governance capabilities in place today.
Who signs off on AI governance?
Four signatures do most of the work, and each one belongs to a person rather than a forum. The first is the value baseline. Before anything is built, Finance signs the metric, its current number from a trusted system, the method for calculating it, and the date it will be measured again. I have written before that this belongs in week two, before the demo impresses anyone. It is a governance document as much as a commercial one, because without it nobody can say later whether the system has earned the right to keep running.
The second is the owner: one named person per AI system, close enough to the operation to know when the output is wrong and senior enough to pause it. In the stalled pilots I wrote about in September, the vendor's team rotates off after the demo and the sponsor changes roles a quarter later, and a live AI system with no named owner goes the same way.
The third is the approval limits. The owner and whoever carries the risk (the controller for payments, the plant manager for a schedule change) agree in writing what the system decides alone and what it drafts for a person to approve. Our agentic AI architecture guide sets the thresholds on three axes, financial impact, data sensitivity and reversibility, and anything above them stops and goes to a person.
The fourth is the audit trail: a record of what the system did, which model version did it, and who approved each exception, reviewed on a date someone owns. This is the evidence Gartner's "Evidence Custodian" would be asked for.
The committee still has a job. It decides which systems get built, reads the evidence each cycle and switches off the ones that miss, the kind of executive decision NIST's Govern function asks leadership to own. It needs the four signatures underneath it to have anything to read.
How does Future Works build in AI governance?
Auditable by Design is written into our contracts: every request, decision and result on the ledger, governed, traceable execution the client can hand to audit. The first of the four signatures comes early in our 12-week cycle, when the client's Finance team signs the Approved Value Baseline in week 2. The work goes live from week 4, Finance verifies the value by week 12, and a portion of our fee is staked on a value target the client's Finance team validates (how the stake works). The first cycle is where the four records are first produced, on a workflow running in the client's operation.
The senior experts who run the work, each directing AI agents on our proprietary AI-native operating system, appear by name in the SOW and stay unless the client agrees to a change, so the people who built a system are still on it when the number gets checked. The instrumentation and guardrails live in that operating layer, and the compare page has a section on governance depth versus shipped outcomes.
A quick test for your own program: choose one AI system in production and ask its team for the four signatures. How long they take to assemble tells you which column your governance sits in.



