Blog posts
Events
Cloudgov University
Walk through a typical enterprise and you will find three generations of AI living under three different regimes. The fraud model from 2019 answers to a model risk committee. The GenAI writing tools answer to an acceptable use policy from 2023. The agents arriving this year answer, so far, to nobody. Three vocabularies, three owners, three levels of maturity, and attackers, auditors and invoices that treat it all as one thing.
The short answer
ML models, GenAI tools and autonomous agents differ in what they risk, but the governance job is one job with five parts:
- Inventory: every model, tool and agent, with an owner
- Identity: every AI actor attributable, human or machine
- Policy: what each may access, spend and do, enforced at use
- Oversight: human accountability scaled to consequence
- Evidence: records that reconstruct any event
Run the five parts once, on one plane, and each generation becomes a profile within the blueprint rather than a program of its own.
Three generations, three risk shapes
- Predictive ML: the quiet incumbents. Scoring, forecasting and classification models embedded in decisions for a decade. Their risks are statistical: drift, bias, silent degradation. Their governance, model risk management, is the most mature discipline in the portfolio, and also the slowest, built for a world where models changed quarterly.
- GenAI: the workforce layer. Chat, coding assistants and content tools in every employee’s hands. The risks moved from statistics to usage: data leaving the organization in prompts, ungoverned tool sprawl, unattributed token spend, and output quality riding on whoever wrote the prompt. Governance here, where it exists, is usually a policy document and an approved tool list, both aging every six weeks.
- Agents: the acting layer. Software that decides and does: touching systems, chaining tools, spending tokens autonomously. The risks are operational: blast radius, irreversibility, runaway spend, accountability gaps. Governance here is being invented in real time, and often only after the first incident.
The instinct is to treat these as three programs, and it fails in a predictable way: the boundaries blur faster than the programs can form. The GenAI coding assistant becomes an agent the day it can commit. The ML model gets wrapped in an agent that acts on its scores. A portfolio governed by generation fragments exactly where the risk concentrates: at the seams.
The blueprint: five parts, three profiles
Part 1: One inventory
A single register of every AI actor: trained models, GenAI tools, embedded assistants, agents, with an owner, a purpose and a data scope for each. The register is live, fed by the control plane rather than by annual surveys, because a survey based inventory is fiction by the second quarter. Shadow entries, discovered usage without owners, get owners or get retired.
Part 2: Identity for every actor
Humans have identity; AI actors need it equally: the model version behind a decision, the tool behind a generation, the agent behind an action. Identity is what makes the rest of the blueprint enforceable, budgets attach to it, policies scope by it, evidence traces through it. The test: for any AI touched outcome in your organization, can you name the acting identity in one query?
Part 3: Policy enforced at the point of use
Each generation gets a policy profile, enforced by the same plane:
- ML profile: deployment gates, drift monitors, revalidation schedules
- GenAI profile: model allowlists, data boundaries, per team and per employee token budgets, routing
- Agent profile: tool and data scoping, per agent budgets, approval gates with rollback, instant pause
The profiles differ; the enforcement point does not. One plane, through which the activity actually flows, applies whichever profile the actor’s identity carries. This is what keeps policy real when the landscape shifts every six weeks: update the profile centrally, and it is enforced everywhere the same day.
Part 4: Oversight scaled to consequence
The human role is graded, not uniform: statistical review boards for high stakes ML, usage visibility and budget accountability for GenAI, approval queues and earned autonomy for agents. One principle covers all three: the more consequential and less reversible the action, the closer the human. Writing that principle once, and implementing it as approval gates and autonomy levels, replaces three separate oversight philosophies that never agreed.
Part 5: Evidence as a by product
Every interaction and action, logged against its identity, reconstructable end to end. Not as a compliance project: as the exhaust of the control plane doing its job. This is the part auditors, boards, ISO 42001 assessors and incident responders all consume, each through their own view, and it is either nearly free (control plane) or nearly impossible (assembled retroactively across three programs).
The migration path from three regimes to one
- Days 1 to 30: stand up the plane where the newest risk is: GenAI and agent traffic. Instrument, attribute, baseline. The ML portfolio keeps its committee for now.
- Days 31 to 60: apply the GenAI and agent profiles: budgets, allowlists, scoping, approval gates. Register the ML models in the same inventory, mapping their existing controls onto the blueprint’s vocabulary.
- Days 61 to 90: unify reporting: one register, one evidence stream, one scorecard to leadership, with the model risk committee consuming the same records through its own lens. From here, every new AI actor of any generation is born into the blueprint.
What this looks like by role
For the CISO
The blueprint ends the coverage argument: no more deciding whether the coding assistant is “GenAI” or “an agent” before knowing whose controls apply. Identity plus profile answers it structurally, and the seams, where generations blur, stop being gaps.
For the CIO
One program, one budget line, one leadership story: AI of every generation, inventoried, governed and evidenced on a plane that absorbs the next generation too. The alternative, a new program per AI wave, has a predictable cost curve and it points the wrong way.
For the chief data or AI officer
The blueprint is adoption infrastructure: teams ship faster when the governance question has a standing answer. Your role shifts from adjudicating each case to owning the profiles, which is a better job.
For the model risk leader
Nothing here demotes model risk management; it federates it. The committee’s controls become the ML profile, its records join the shared evidence stream, and its discipline, the most mature in the portfolio, becomes the template the newer generations inherit.
How Cloudgov.ai implements the blueprint
- One inventory and identity plane: every model interaction, tool and agent attributed via TokenShield and AgentShield.
- Profiles enforced at use: allowlists, budgets and scoping in real time; approval gates and rollback through ActionShield for the acting layer.
- Oversight scaled to consequence: autonomy levels earned per action class, with the trust telemetry to prove it.
- Evidence by construction: one attributed record stream serving audits, boards and frameworks, with MulticloudShield holding the environmental ground truth underneath.
The scorecard across all three generations is the Governance ROI score: value per unit of governed AI, whatever its vintage.
Frequently asked questions
Should ML, GenAI and agents be governed by one program or three?
One program with three profiles. The generations differ in risk shape but share the same five governance parts: inventory, identity, enforced policy, scaled oversight and evidence. Separate programs fragment exactly where generations blur, which is where risk concentrates.
How is agent governance different from GenAI governance?
GenAI governance manages usage: data boundaries, tool allowlists, token budgets. Agent governance manages action: scoping, approval gates, rollback, blast radius and pause. Both run as profiles on the same enforcement plane.
What happens to our model risk management committee?
It federates rather than dissolves: its controls become the ML profile, its records join the shared evidence stream, and its maturity becomes the template for the newer layers.
Where should the unified program start?
At the newest risk: instrument GenAI and agent traffic through the control plane first, then register the ML portfolio into the same inventory and reporting. The 90 day path runs instrument, enforce, unify.
What is a policy profile?
The generation appropriate control set an AI actor’s identity carries: deployment gates and drift monitors for ML, allowlists and budgets for GenAI, scoping and approval gates for agents, all enforced by the same plane at the point of use.
How does the blueprint handle the next AI generation?
By construction: new actor types get identities and a new profile on the existing plane. The five parts are generation agnostic, which is the point of building them once.
One blueprint, whole portfolio
Cloudgov.ai governs your ML, GenAI and agent generations on one plane: one inventory, one enforcement point, one evidence stream, one score. Talk to our team about unifying your three regimes, or start with our resources.