AI-ready architecture starts with engineering discipline

Data-driven decision-making rests on a reasonable premise: better data leads to better decisions.

For Wayne Yan, chief technology officer at Dariel, AI hasn’t changed that premise or the problem underneath it.

“The basic pattern of input, processing and output is essentially the same as it was 90 years ago,” he says. “If you think of an AI model as a multivariable function, the old rule still applies. Garbage in, garbage out.”

The definition of input has grown

What has shifted is what counts as input. Training data and the data supplied to a model at runtime are the obvious parts. Yan adds everything else that shapes how a model behaves: the instructions it receives, the agentic skills it can call on, and the constraints, rules and assumptions built around it.

“Traditional software came with guardrails we took for granted,” he says. “A compiler would reject an invalid expression. A syntax tree told you when something was malformed. With AI, there are 50 ways to say the same thing, and the machine will happily interpret most of them.”

A poorly specified instruction or an ambiguous rule doesn’t fail loudly. It produces an answer, and that answer usually looks plausible. This puts more weight on the quality of every input, including the ones nobody used to think of as data.

Why switching off old systems isn’t enough

A common response is to treat AI readiness as a clean-out exercise: find the systems that can’t expose clean data and retire them. Yan understands the appeal but believes it skips the hard part.

“Clean data is one thing. Getting to clean data is another,” he says. “Retiring a system that produces poor data doesn’t create good data. The underlying issues of ownership, governance, security, architecture, integration and data quality are still there after the system has gone.”

In many organisations those issues sit in the processes and responsibilities surrounding a system as much as in the system itself. If a platform is replaced without settling who owns the data, how it is validated and who may access it, the same problems resurface in the new environment.

Principles first, then a roadmap

Yan argues that AI-ready architecture should be approached the way any good architecture is: from principles. He points to information security, data governance, data stewardship, componentisation, design clarity, controlled and standardised accessibility, and a clear roadmap for system maturity.

“If existing systems violate those principles, the answer isn’t necessarily to switch them off tomorrow,” he says. “Your enterprise architecture roadmap should identify those gaps and establish a programme of modernisation that progressively moves those systems towards the architectural state you need.”

For agentic AI, that target state makes specific demands. Systems need well-designed APIs and, increasingly, support for the Model Context Protocol (MCP), which gives AI agents a standard way to connect to tools and data. They also need data integrity and enough overall system quality for an agent to act on what it retrieves without a person verifying every step.

Getting there is rarely a single project. It means sequencing the work so that the systems most important to planned AI use cases mature first, with others following as the business case allows. “Being deliberate makes the difference between failure and success,” says Yan. “That means a practical programme of change, run with good old-fashioned software engineering.”

Familiar risk, greater reach

The risks are not new. Organisations have always been able to produce the wrong result from bad inputs.

“The risk today is not fundamentally different from the risk we faced yesterday,” says Yan. “We can still produce the wrong result. AI can produce it much faster and at a far greater scale.”

An agent working across connected systems can repeat a flawed assumption in thousands of transactions before anyone spots the pattern. In Yan’s view, most of these failures trace back to weak engineering, where organisations lose sight of the input, processing and output discipline that has applied since the early days of computing.

Engineering as critical thinking

Engineering is usually defined as the application of science and mathematics to design, build and improve systems that solve real-world problems. Yan prefers a shorter version.

“Engineering is the application of critical thinking to a problem,” he says. “Deciding which data to trust, which rules a model should follow and where it needs a person to check its work are judgement calls. AI can help with a great deal, but for now that judgement has to come from people.”

For organisations planning agentic AI, the practical starting point is an honest architecture review that establishes which systems meet these principles, which fall short, and what a realistic path to maturity looks like, before any model is chosen.

The post AI-ready architecture starts with engineering discipline appeared first on IT-Online.