Connectivity was the easy part: now comes governance.

The race to connect AI to business systems happened faster (and proved easier) than anyone expected. The harder question is one most enterprises can’t yet answer.

 

When an AI agent updates a customer record, triggers a financial workflow or modifies an inventory position, three things suddenly matter: who is accountable; under what rules; and how anyone would know if something went wrong.

 

The new accountability problem

 

Consider an enterprise using AI to handle refund requests. The agent reads an incoming complaint, retrieves the order history, checks the return policy and issues the refund automatically.

 

Now suppose the refund was authorised in error: the customer had already received a replacement, the policy didn’t cover the claim, or the agent misread the invoice.

 

Who is responsible? The model, the policy logic it relied on, or the team that approved the workflow without a human checkpoint?

 

This isn’t hypothetical. AI is being moved into production faster than the governance frameworks around it.

 

Production isn’t the same as operational maturity

 

The enterprise conversation has moved on. What was once a debate about which models to deploy has become a much harder question about how to operationalise them. Research from MIT Technology Review Insights, conducted in partnership with Celigo, finds that 76 per cent of mid- to large-sized US firms have at least one AI workflow running in at least one department. But Gartner predicts more than 40 per cent of agentic AI projects will be cancelled by 2027 – citing cost, accuracy and governance failures, not model capability.

 

The gap is structural, not technical. Enterprises know how to deploy AI inside a single application. They don’t yet know how to govern AI operating across multiple systems in real time.

 

Operational maturity looks different. It means AI that runs reliably across multiple systems, with consistent oversight, decisions that can be reconstructed and costs that can be measured by the workflow they support.

 

As April Rassa, VP of Product Marketing at Celigo, frames it: “Agents can only be as reliable as the infrastructure they run on.”

 

A patchwork of vendor guardrails

 

Many vendors have responded by building governance into their own products. But none of those efforts combine into a coherent governance model. Every CRM, ERP and support platform now ships with some version of guardrails, approval flows and audit logs. The problem is that each one defines acceptable behaviour differently, logs to a different place and enforces policy through its own mechanism.

 

A finance leader trying to understand who approved a credit decision made by an AI agent can’t simply ask one system. The CRM may have logged the trigger under one identity; the ERP committed the transaction under another; the support system recorded the customer interaction under a third. Each operates with its own logging, retention and access controls. Reconstructing what happened means reconciling all of them.

 

The parallel to the API era is instructive. Enterprise systems have exposed APIs for years, but companies still needed automation platforms to govern how those APIs were called. Agentic AI is creating the same problem at higher stakes, except the callers are now autonomous and the consequences can be customer-facing within seconds.

 

Celigo addresses this by treating cross-system context as platform infrastructure: more than 1,000 pre-built connectors that already understand how enterprise systems interact, paired with support for the MCP open standard so AI agents can act within that context, not outside it.

 

‘Rules versus AI’ is the wrong frame

 

Many leadership teams are mapping where rules-based automation fits and where AI takes over. It feels like a sensible starting point. But it’s the wrong frame.

 

Real business processes don’t split that cleanly. A single expense reimbursement workflow can touch a fully scripted policy check, a manager’s approval for exceptions and an AI agent flagging anomalies, sometimes within the same transaction. Payroll sits at one end of this spectrum and customer support at the other, but most enterprise workflows fall somewhere in between, blending modes within a single process.

 

Deterministic automation isn’t something to be replaced as AI matures – it remains essential to processes that demand precision. As Jan Arendtsz, Celigo’s founder and CEO, puts it: “Predictable workflows aren’t legacy. They’re essential.”

 

A related discipline applies to the AI side of the spectrum. Even where AI is the right tool, more of it isn’t better. Celigo’s CPO Matt Graney calls the underlying principle “least agency”: solve the problem with the minimum amount of AI required, because every additional AI step adds cost, latency and risk of hallucination.

 

The discipline is to apply AI precisely where it produces the best outcome and to use deterministic logic and human oversight everywhere else.

 

Governance has to sit above the stack

 

The governance layer cannot live inside any single vendor’s product. It has to sit above all of them, applying consistent policy and oversight regardless of which system executed the action.

 

In practice, that means embedding guardrails as explicit steps in the workflow rather than leaving them to individual systems: a PII filter before an agent receives data, a policy check after the agent acts, an audit trail capturing not just what was done but under which rules. It also means cost governance: knowing how AI spend breaks down across functions, so ROI can be measured where it’s earned.

 

Some enterprises are already operating this way. Egnyte runs AI agents across its enterprise systems through Celigo’s integration platform – an approach its CIO Frank Scilia describes as the ability to “scale Egnyte’s workflows while maintaining our governance standards.”

 

That architecture rests on two concrete capabilities: support for the MCP open standard, which gives enterprises a consistent way to connect AI agents to any system; and human-in-the-loop controls that keep consequential decisions accountable before they execute.

 

The data backs the structural case. MIT Technology Review Insights found that 90 per cent of organisations with AI workflows fully in production already use an integration platform. Integration was once treated as plumbing. In an AI-driven enterprise, it’s becoming the control plane.

 

The shift ahead

 

For most enterprises, the integration layer was always how work moved between systems. Agentic AI is changing what that layer has to do. The platforms that once connected systems are now being asked to govern what runs across them, enforcing policy, tracking decisions and holding autonomous actions to a consistent standard regardless of which tool executed them.

 

The answer is a deliberate marriage of deterministic and agentic, each constrained to its lane, with the integration layer holding the boundary between them. That architecture pays off practically. Fewer AI steps means fewer tokens, lower latency and less surface area for error.

 

The companies that win this shift won’t be running the most AI. They’ll be running it on a platform built to connect and control how their business operates, with the governance to prove it, across every mission-critical process, at enterprise scale. Celigo is built for that.

---

Celigo just launched Operationalize AI playbooks for IT, finance and accounting, and sales and revenue operations teams to show the framework for operating these critical processes. Access the playbook.