How PMOs Can Govern a Growing AI Agent Portfolio With Creatio AI Studio
Key Takeaways
- AI agent sprawl is a portfolio problem: agents come from employees, business and operations teams, SaaS tools and developers, usually with no shared record of who owns them, what they can reach or what they cost.
- The disciplines a PMO already runs (intake, a register, health reporting and budget tracking) transfer directly to AI agents.
- Once agents take thousands of actions a day, governance has to be enforced by the platform while agents run.
- In Creatio AI Studio, policies, human approvals, observability and AI spend controls apply to every agent built on Creatio, and to outside agents at the point they touch Creatio data or processes.
- The layer that does this across a whole agent estate is increasingly called an AI control plane.
Ask a PMO lead how many projects are in flight and the answer comes back in seconds, with an owner, a budget and a status for each one. Ask the same question about AI agents and the answer usually takes much longer, if it comes at all.
Employees build personal assistants, service teams deploy case-handling agents, SaaS tools arrive with built-in automation switched on, and developers build custom ones on OpenAI or Anthropic models. Each agent reads data, takes actions and consumes paid model capacity, yet few organizations track them as closely as a two-week website refresh.
PMOs have solved this kind of problem before. The five steps below apply familiar portfolio disciplines to AI agents, with templates and an example of how each step works in Creatio AI Studio.
What AI Agent Sprawl Looks Like in a Delivery Organization
AI agent sprawl is the unmanaged growth of AI agents across an organization: agents built by different teams, on different tools, with no shared record of what each one does, what it can reach or what it costs.
It rarely starts as a problem. Each team adds an agent that makes sense locally, and a year later nobody can produce a complete list, two agents give customers conflicting answers, and AI costs have grown with no clear explanation. Shadow IT followed the same pattern, and so did project portfolios before intake processes existed. The fix then was to give every initiative an owner, a scope and a budget, and the same questions map closely onto AI agents:
| What a PMO Tracks for a Project | The Equivalent for an AI Agent |
|---|---|
| Owner and sponsor | Who is accountable for the agent's behavior and results |
| Scope | Which tasks the agent handles, and which it must hand off |
| Access and dependencies | Which systems, data and tools the agent can reach |
| Approvals and stage gates | Which actions need a person to sign off before they run |
| Budget | How much AI capacity the agent may consume per hour, day or month |
| Status and health | Whether the agent still performs at the quality it was approved at |
| Change control | Who can change its instructions, model or permissions, and how changes are tested |
Most enterprises I speak with are well past a handful of agents. A handful is easy to keep track of. Beyond that, almost none of them can tell me what all of their agents are doing right now, what they cost or who approved them. That gap is the whole problem.
A spreadsheet works while the list is short. As it grows, the answers need to live in the platform the agents run on.
What Creatio AI Studio Is and Where It Fits in the Agent Portfolio
Creatio AI Studio is the AI-native environment organizations use to build, run and govern enterprise AI agents across the full agent lifecycle, from design and testing through monitoring in production.
is the AI-native environment organizations use to build, run and govern enterprise AI agents across the full agent lifecycle, from design and testing through monitoring in production. Launched in Creatio's 10x release in July 2026, it works alongside Creatio's AI CRM on a shared platform and can also be used separately, connecting agents to the systems an organization already runs. It brings together three groups of capabilities:
- Building agents. Business teams design agents without code in the Prompt Agent Designer and the Workflow Agent Designer. Developers use coding agents such as Claude Code, Codex and Cursor, and the Creatio Agent SDK supports agents built with OpenAI and Anthropic technologies.
- Connecting agents. Agents reach business data and workflows through Creatio's native MCP server, and each agent can use a different model, including OpenAI, Anthropic and Google models.
- Controlling agents. Organization-wide and agent-level policies, human approvals, an audit log, observability, usage limits and AI Credits for AI spend management.
Personal agents follow the same rules. Employees build them in Creatio AI Twin, a conversational, no-code experience where they describe a goal in natural language, test the result and deploy it. An AI Twin agent inherits the exact permissions of the person it acts for, so it never gains access its owner does not have.
That gives a PMO a clear boundary. Agents built and running on Creatio, whether pre-built, designed in AI Studio, described in AI Twin or written against the Agent SDK, share one set of policies, spend controls and monitoring. Agents running elsewhere are governed at the point they reach into Creatio data, processes or approvals, through permissions, human approval and the audit log.
Step 1: Build an AI Agent Register Before the Next Agent Goes Live
An AI agent register is the single record of every agent in use: what it does, who owns it, what it can reach and how it is performing. It doubles as the intake gate. One rule, no register entry and no production access, stops new agents from slipping in unrecorded.
The example column below uses the Creatio Data Enrichment Agent, a pre-built sales agent that finds missing account details in external sources and adds them to the CRM record. The values are illustrative.
Template: AI Agent Register
| Field | What to Record | Example |
|---|---|---|
| Agent name | Full name, as it appears in the platform | Creatio Data Enrichment Agent |
| Purpose and scope | The task it handles and where it stops | Completes missing company details such as industry, size and location; does not change ownership or pipeline data |
| Business owner | The person accountable for its results | Sales operations lead |
| Where it runs | Platform and build method, inside or outside Creatio | Pre-built agent, runs on Creatio |
| Systems, data and model | What it can reach, and the model it calls | Account records and public web sources; model approved at organization level |
| Autonomy | Suggests, acts with approval, or acts within set limits | Assistive: proposes updates for a sales rep to review |
| Human approval points | Actions that need sign-off before they run | Overwriting a value a person entered |
| Budget | Allowed consumption per day or month | Share of the sales team's AI Credits |
| Quality check | How quality is measured, and the last review date | Monthly spot check of updated records |
| Status and last change | Lifecycle stage, and the latest approved change | Production; instructions updated and approved by the owner |
Review the register on the same cadence as the project portfolio. Agents with no owner, or no recent quality check, are the first candidates to pause or retire.
How Creatio AI Studio Keeps the Register Close to Live
For agents that run on Creatio, much of the register can be read straight from the platform. Creatio AI Studio holds each agent's configuration, model, applicable policies and version history. Its monitoring views show run status, errors, escalations and consumption by agent, and versioning with rollback covers change control. The PMO still owns the fields a platform cannot know, such as the business owner, the intended scope and the review cadence.
A register shows what exists. The next step decides what each agent is allowed to do.
Step 2: Turn AI Agent Governance Policies Into Rules the Platform Enforces
Most organizations start AI governance with documents: an acceptable-use policy, a list of approved models, a rule that customer data stays inside approved systems. Those documents hold while a few people can check agents by hand. Once agents run across dozens of teams, manual checks fall behind, and executable governance takes over: the platform applies policies every time an agent runs, covering rules such as:
- Which models an agent may call
- Which personal data is masked before it reaches a model
- Which tools need a person to confirm before the agent uses them
- Which actions need human approval first, such as a refund above a set amount
- How much an agent may consume per hour, day or month
At enterprise scale, the only governance that means anything is the kind the platform applies on its own, every time, whether or not anyone is watching.
For a PMO, the practical exercise is translation: for each written policy, decide what rule would enforce it, where that rule sits, and who is notified when it blocks an action.
Organization-Wide and Agent-Level Policies in Creatio AI Studio
Creatio AI Studio applies policies at two levels. Organization-wide policies cover every agent: approved models, masking of personal data, tool confirmation, human approval and usage limits. Agent-level policies add rules for a single agent, alongside its own instructions and knowledge. A new agent inherits the organization-wide rules automatically, so nobody has to remember to add them.
Approval requests can be handled inside AI Studio, in chat or in the CRM, wherever the approver already works. A governance dashboard shows pending approvals, policy decisions and the agents where decisions cluster most often. Those clusters are useful in a portfolio review: an agent that keeps stopping at the same approval may need a narrower scope, or the rule may need adjusting. The audit log records every action.
What Agents Built in Creatio AI Twin Inherit
Personal agents are where sprawl usually starts, because employees can build them without a project or a ticket. In Creatio AI Twin, an agent inherits the role-based access of the person it acts for, follows Creatio's existing roles and business rules, and logs every decision back to that person. It moves from sandbox to production under the controls IT already applies.
IT sets the guardrails once, and employees keep building inside them. A sensible PMO rule is to add personal agents to the register once they are shared with a team or start working with customer data.
Rules control what a single agent may do. Most business processes, though, involve more than one agent and more than one person.
Step 3: Keep People Inside the Process When Agents Take on the Work
Enterprise processes rarely sit inside one application. Customer onboarding at a bank can touch the CRM, document processing, identity verification, core banking and compliance, with several people signing off along the way. Adding one agent per system spreads the complexity around without reducing it.
Those processes also include participants that are not agents at all: a business rule, an approval queue, a compliance checkpoint, a manager who signs off before the next step runs. Human-in-the-loop design means coordinating all of them in one flow, with the right context passed at every handoff.
A handoff map makes this visible before anything is built. For each step, record who does the work, what triggers the handoff, what context moves with it, and what happens if nobody responds.
Template: People-and-Agents Handoff Map (Illustrative Onboarding Example)
| Step | Done By | Hands off When | Context Passed On | If Nobody Responds |
|---|---|---|---|---|
| Collect application and documents | Onboarding agent | All required documents are received | Application data and document set | Reminder to the customer after an agreed interval |
| Validate data and flag gaps | Onboarding agent | Validation is complete or an inconsistency is found | Validation results and list of issues | Routes to an operations specialist |
| Compliance review | Compliance officer | Decision is recorded | Structured summary with flagged issues | Escalates to the team lead |
| Open the account | Workflow step | Account is created in the core system | Approved application | Error routes to operations |
A regional bank that runs on Creatio works this way. Its onboarding agent extracts information from applications and identity documents, validates the required data, identifies missing or inconsistent information and prepares a structured summary for human review, while employees keep control of compliance decisions and final approval.
The result is less manual document handling, faster and more consistent onboarding, and a clearer audit trail. As the deployment expands, the same workflow can connect e-signature, account funding and the bank's core system without adding disconnected agents.
How Creatio AI Studio Workflow Agents Pause for Approval and Resume
In Creatio AI Studio, the Workflow Agent Designer builds this kind of process visually. Deterministic steps, business rules and approval points sit in the same flow as the agent's reasoning, so a workflow agent can stop at an approval, wait for a person's decision and continue once it arrives. Multi-agent orchestration coordinates specialized agents, workflows, applications, tools and human approvals across one end-to-end process.
Context travels with the work. Agents draw on CRM records, workflow state and conversation history, so an agent handling a service case works from the same customer record the sales team uses, and existing access controls stay in force. Combining deterministic steps with agent reasoning also helps the budget, because workflow steps consume no AI Credits and only the steps that need judgment call a model.
Once agents are live, the question shifts from what they may do to how well they are doing it.
Step 4: Report Agent Health the Way You Report Project Health
An agent that passed every test before launch can still drift as models, data and the business change around it. Without regular measurement, the first sign of trouble is often a customer complaint.
PMOs already have the format to catch this early: the status report. Applied to agents, it rates a few dimensions red, amber or green against thresholds agreed when each agent was approved.
Template: AI Agent Health Report
| Dimension | What to Check | Green | Amber | Red |
|---|---|---|---|---|
| Quality | Scores against the agent's test set and production quality scoring | At or above the approved threshold | Trending toward the threshold | Below the threshold |
| Escalations | How often the agent hands work to a person, and why | Stable, with expected reasons | Rising, or new reasons appearing | Sharp rise or a recurring failure reason |
| Errors | Failed steps and what happened to the work | Rare, and the work recovered | Repeated in one step | Work lost or stuck |
| Business outcome | The result the agent was built to deliver | On target | Below target for one period | Below target for two periods or more |
| Spend | Consumption against budget | Within budget | Approaching a limit | Over budget or hitting limits |
| Policy activity | Blocked actions and pending approvals | In line with expectations | Approvals queuing up | Repeated blocks on the same rule |
A red rating on quality or policy activity is a reason to pause the agent or reduce its autonomy until the cause is found. Red on spend calls for a conversation with the owner about scope or model choice.
AI Observability in Creatio AI Studio
Creatio AI Studio supplies most of this report from the platform. Before release, agents are tested against golden datasets (real scenarios with known correct answers) with quality thresholds. In production, quality scoring, drift detection, outcome measurement and user feedback keep running, and monitoring views show run status, latency, errors, escalations, quality trends and cost by agent and by model.
Findings from production feed back into the test set, so the next version of an agent is tested against what went wrong in the previous one.
The spend line in that report needs a step of its own, because AI costs behave differently from most software costs.
Step 5: Put AI Spend on the Budget Before It Lands on the Invoice
Traditional software is budgeted per license or per seat. AI consumption accumulates per model call, per agent run and per workflow, which makes it variable, spread across teams and hard to attribute. Many organizations reconcile it at quarter end, when the money is already spent.
A PMO can apply the discipline it uses for project budgets. Allocate AI capacity by department or team and record each allocation in the register. Set thresholds that warn before a limit and act at the limit.
Match the model to the task, since classification rarely needs the most capable model and predictable steps may not need a model at all. Then forecast consumption at two or three times today's volume, so scaling becomes a planned decision.
Usage Limits and AI Credits in Creatio AI Studio
Creatio AI Studio meters AI consumption in AI Credits, so spend no longer has to be tracked separately with each model provider. Organizations allocate credits by department and team, and managers can see their team's consumption. The AI Command Center shows credits consumed by user, team, department and AI skill, along with trends over time and high-consumption outliers.
AI Credits budgets can be set for the whole organization and for individual agents, by hour, day and month. Each budget has a warning threshold, and an optional hard stop blocks requests once the budget is exceeded. Cost breakdowns by model and burn rate show where consumption comes from and how fast it is growing, and each agent can be assigned the model that fits its task.
These controls cover consumption that runs through Creatio. An agent that bills its own model provider directly sits outside them, which is worth noting in the register: the way an agent connects decides how much of it can be managed.
From Agent Portfolio to AI Control Plane
Taken together, the five steps describe a layer many organizations are trying to build, and analysts now track agentic control plane solutions as a category of their own. An AI control plane is the enterprise layer that governs, orchestrates, observes and economically manages AI agents across an organization, whoever built them and wherever they run.
The term comes from networking and cloud infrastructure, where the control plane decides and the data plane executes. For a delivery organization, the control plane also has to understand the work itself: routing a case to the right agent, pausing it for a manager's approval and resuming it inside a business process.
Creatio describes this as the direction for Creatio AI Studio. Agents built on Creatio, from no-code agents to those written against the Agent SDK, are governed through one set of models, policies, spend controls and orchestration, and agents running elsewhere are governed whenever they reach into Creatio data, processes or approvals.
Enterprises will not standardize on a single AI vendor. Teams should be able to build agents anywhere, as long as nothing acts on business data without permission.
6 Questions for Your Next Agent Portfolio Review
Put these questions to the team before the next portfolio review. Each maps to one of the steps above, so a "no" points straight to the work that needs doing.
| Question | Step |
|---|---|
| Can we produce a complete list of the agents running today, each with a named owner? | 1 |
| Can we apply a new policy to every agent at once, without rebuilding any of them? | 2 |
| Can we enforce our rules on an agent we did not build? | 2 |
| When an agent hands work to a person, does it pick the work back up once that person acts? | 3 |
| Would we notice an agent's quality slipping before a customer did? | 4 |
| Do we know what each agent cost last month, and whose budget paid for it? | 5 |
FAQ
What Is an AI Control Plane?
An AI control plane is the enterprise layer that governs, orchestrates, observes and manages the cost of AI agents across an organization. It applies policies automatically while agents run, coordinates agents with workflows and human approvals, monitors quality in production and controls AI spend, whoever built the agent and wherever it runs.
What Is AI Agent Sprawl, and How Can a PMO Prevent It?
AI agent sprawl is the unmanaged growth of AI agents built by different teams on different tools, without a shared record of ownership, access or cost. A PMO can prevent it with an agent register that doubles as an intake gate, policies the platform enforces at runtime, regular health reporting and a budget for AI consumption.
Can Creatio AI Studio Govern AI Agents Built Outside Creatio?
Yes, at the point where those agents touch Creatio. Agents built and running on Creatio, including agents written with the Creatio Agent SDK using OpenAI or Anthropic technologies, share one set of policies, spend controls and monitoring. Agents running on other platforms are governed when they reach into Creatio data, processes or approvals, through permissions, human approval and the audit log.
Where to Start: One Register and One Review
Governance is often treated as a brake on AI adoption. At portfolio scale it works the other way: when the platform enforces the rules, builders move faster, because the guardrails travel with them.
For most PMOs the first step is small. List the agents already running, who owns each one and what each one costs. That exercise usually shows whether the gap is a policy problem or a platform problem, and which of the five steps to tackle first.
