In the rush to deploy AI agents, many organisations are discovering that the technology is only as dependable as the systems wrapped around it. The latest cautionary tale is not that agents are inherently unusable, but that businesses keep asking them to operate in environments that were never designed for autonomous decision-making.
That lesson has become more visible as a string of highly publicised misfires has fed concern about what happens when agentic systems are pointed ...
Continue Reading This Article
Enjoy this article as well as all of our content, including reports, news, tips and more.
By registering or signing into your SRM Today account, you agree to SRM Today's Terms of Use and consent to the processing of your personal information as described in our Privacy Policy.
directly at live operations. Even the familiar comic imagery of McDonald’s briefly testing voice ordering in the United States has become part of the wider conversation about how quickly automated systems can become awkward, unreliable or disruptive when they are not properly bounded.
The underlying problem, according to Mendix, is not simply model quality. It is architecture. Many firms still rely on fragmented applications, disconnected business-process tools and siloed records that make it difficult to establish guardrails before agents are given access to critical workflows. In that environment, an agent can become just another layer of complexity, acting on inconsistent information and amplifying operational risk.
That concern is beginning to intersect with regulation. From December 2026, the European Union’s revised Product Liability Directive is set to extend strict liability to software, AI systems and digital services, according to legal and policy summaries of the reform. The shift matters because it moves the debate away from whether an AI model behaved badly and towards whether the surrounding system was designed responsibly. In parallel, insurers are increasingly treating generative AI as a source of exclusions or special terms in standard coverage, sharpening the pressure on firms to show they have controls in place.
Mendix argues that the answer is an orchestration layer: a governed environment that links data, systems and workflows before agents are allowed into production. The idea is that the organisation should first define context, roles and limits, then deploy the agent into that framework rather than hoping the agent itself will supply discipline.
The case for that approach echoes an older management idea. W. Edwards Deming spent much of his career arguing that poor systems, not individual workers, are usually to blame for operational failure. Applied to AI, the principle is that “agent error” can often be traced back to a weak foundation rather than a defective model. When OpenAI examined a breach attempt involving one of its own agents and Hugging Face, it found that the same model inside its production safeguards was far less likely to compromise infrastructure, underlining the value of environment and controls.
The scale of the issue is likely to grow quickly. Gartner has forecast that the average Fortune 500 company could be running more than 150,000 AI agents by 2028, with many arriving embedded in software that business units activate on their own. That creates what Gartner describes as “agent sprawl”, a form of shadow IT in which autonomous tools proliferate faster than governance. The consultancy says only 13% of organisations believe they have the right governance in place.
Mendix says there are five essentials that need to be settled before an agent touches live systems: connected context, process execution, enforced boundaries, expert input and named ownership. In practice, that means the agent should be operating from a shared source of truth, working inside a clearly defined workflow, and remaining subject to human oversight at the right points. Subject-matter experts, the company argues, should help shape the logic from the start, while responsibility for each process must remain traceable to a named person.
Chris Kuijper, senior product manager for process and logic orchestration at Mendix, said the problem with partial automation is that the bottleneck rarely disappears; it simply moves somewhere else. He noted that speeding up coding with AI-assisted development, for example, does not remove the wider coordination challenge. He also said agents should be managed and monitored much like people, rather than treated as mysterious exceptions to governance.
Mendix points to MBH Bank in Hungary as an example of why orchestration can matter. The bank was created in 2023 through the merger of three lenders, each with its own method for handling customer callbacks. Contact-centre staff logged requests, but the different systems made it difficult to see what was waiting across the organisation.
Working with a small business team, a seven-person Mendix project rebuilt the callback process so that each request became a task in an employee’s or department’s queue, with managers able to reassign work when needed. Within days of launch, users had moved seven other tools on to the same platform. The system now handles roughly 10,000 callbacks a month for 300 users, and the same team has built more than 40 applications on the platform in two years, according to the company.
MBH Bank says the redesign has freed up the equivalent of several full-time employees’ work each month by automating coordination, accelerated response times, and given teams a single place to track performance. István Sponga, head of operations development, described the result as a workflow engine in which each callback sits in a queue for an individual employee or an entire department. He said the next stage is to connect more AI-enabled applications to the same environment.
That is the broader point Mendix is making: an agent added to an already disciplined workflow is far more likely to be useful than one dropped into a messy system and left to improvise. The platform says it can provide connected context through integrations with existing systems of record, process execution through end-to-end workflow control, enforced boundaries through human review, expert input through visual development tools and visibility through step-by-step monitoring.
The temptation, in other words, is to start with the latest technology and ask questions later. But the more durable approach is to build the structure first. If agents are going to become a standard part of enterprise software, then orchestration may be less a nice-to-have than the difference between controlled automation and expensive chaos.
Source: Noah Wire Services
