Does this theory hold up under the optimal conditions of domain-team involvement, workflow oversight, documented policies, reliable data security and monitoring?
As they move AI systems into operational workflows, enterprises need to formalize decision-making rules and accountability, argues a senior executive of a US-based enterprise analytics and automation software firm.
The premise is that enterprise AI programs may be investing heavily in data, models and infrastructure while overlooking a less visible part of deployment: the rules that determine how a business makes decisions.
That is the argument from Philip Madgwick, Regional Vice-President (Asia), Alteryx. He says organisations need to document and govern the business rules, approval processes and risk tolerances that sit between an AI system’s output and an operational decision. “The next frontier in enterprise AI isn’t data. It’s business logic,” Madgwick notes. He defines business logic as “the rules, calculations, policies, workflows, and decision frameworks that determine how an organisation operates”.
The concept can include relatively simple rules, such as pricing thresholds, as well as more complex arrangements involving compliance requirements, market-specific policies and approval hierarchies. Madgwick argues that these rules can vary significantly between organisations, even where the companies use similar data and models. “Two organisations can have access to the same data and the same model, and arrive at completely different conclusions, because their business rules, priorities, and risk profiles differ,” he says.
Business logic as an obstacle
That observation is uncontroversial in itself. A credit assessment, pricing recommendation or fraud alert cannot be evaluated without knowing the organisation’s objectives, risk appetite and obligations. The more contentious part of Madgwick’s argument is his suggestion that business logic has become the main obstacle to scaling AI.
“The answer cannot be found in the data or the models,” he argues. “Rather, it lives in business logic.” That framing presents a false choice if read literally. Data quality, model performance, infrastructure, security and monitoring remain necessary to the operation of reliable AI systems. A clearly documented business rule cannot compensate for biased data, an unsuitable model or an insecure deployment.
Madgwick’s arguments draws partly on his firm’s research involving a survey of 1,400 business and IT leaders globally, including 350 in APAC. Madgwick sees the survey results ass evidence of a governance and workflow problem. “That is not a model problem. It is a business logic problem, and it is costing organizations time they cannot afford,” he says.
There is a practical case for involving domain experts. They may understand exceptions, regulatory requirements and operational constraints that are absent from a model specification. However, business-level ownership also carries risks, including inconsistent controls, duplicated systems and the growth of shadow AI. A workable governance model is likely to require cooperation between domain teams and central functions such as IT, security, legal, privacy and risk.
Nevertheless, Madgwick proposes four characteristics for AI workflows: Visible, Understandable, Repeatable and Auditable.
- Visible means users can see where an output came from
- Understandable means non-technical users can validate it
- Repeatable means the same inputs produce consistent outputs
- Auditable means there is a record of accountability when a decision is challenged
“These are the foundations for trust,” Madgwick asserts. The principles are familiar from wider discussions of explainability, reproducibility, traceability and governance. Such principles expose a difficulty in applying business logic to generative and agentic systems. A workflow may be repeatable at the process level while still relying on a probabilistic model whose output varies. Users may also understand the sequence of steps without being able to determine whether the data, assumptions or rules were appropriate.
For Madgwick, the answer is to ensure that AI is connected to approved organisational rules rather than left to infer operational context from data alone. “AI should adapt to the business it serves, not the reverse,” he believes.
That argument may be most relevant when AI is used to support decisions with clear rules and defined approval processes. It is less straightforward where the organization’s objectives are contested, where policies change frequently, or where the relevant expertise is tacit and difficult to formalize.
Are analysts’ roles changing?
Madgwick also predicts a change in the role of analysts. “They are becoming the custodians of the business logic that AI depends on,” he asserts, describing analysts as responsible for defining rules, validating outputs and keeping automated processes aligned with business practice. Whether that becomes a durable role will depend on how organisations divide responsibility for AI systems.
Analysts may help translate operational requirements into workflows, but accountability cannot rest on them alone. The broader lesson is less dramatic than the claim that business logic will replace data as the next enterprise AI frontier.
Organisations moving AI into production need both: trustworthy data and models, plus explicit rules, decision rights and oversight. Without those elements, a technically capable system may still produce an answer that no one is authorised — or willing — to act on.
