AI produces a measurable return only when it solves a defined business problem. Solving that problem depends less on software code than on understanding the organization’s business rules and the goal they serve. These rules define a correct, measurable outcome and provide the basis for reliable AI.
Many executives have directed their organizations to adopt AI. The intention is sound, but adoption without a defined business problem is not. A company that deploys AI without defining the problem may gain little beyond email that appears better written. Better email may save time, but that gain alone does not establish a return on the broader investment.
Start With the Most Valuable Problems
Every organization recognizes problems. Few rank them by the value of solving them.
Our first task is to identify the problems whose solutions would yield the greatest return. The second task is to divide each problem into smaller problems. The third is to sequence the work so that each solution increases the value of the last. AI may solve part of a smaller problem. It rarely solves all of it.
Each selected problem requires metrics. Measurements before work begins establish a baseline. Measurements afterward show the return. Continuous monitoring confirms that the solution still performs as intended and remains within budget. Without metrics, we cannot demonstrate a return. Without monitoring, we cannot know whether that return lasts.
An Organization That Began With a Solution
An organization recently adopted a new operating strategy. Its core system, built for its industry, no longer served its needs. Its first step was to evaluate industry replacements. We proposed discovery before product evaluation.
Discovery redefined the problem. The existing system could not accommodate each project’s requirements. It also confined information within departments, causing teams to work in silos without recognizing knowledge the organization already held. Replacing it with another industry-standard system would not have solved either problem, because none of the systems under consideration addressed the organization’s core needs.
The organization needed flexibility more than features. We recommended a workflow collaboration system, then used AI to configure it to the organization’s requirements in a fraction of the time manual configuration would require. We also used AI to build deterministic connections between the new system, the finance system and the CRM. The system is scheduled to enter production on January 1, 2027.
AI accelerated the implementation. Discovery established what the organization needed.
The Rules Matter More Than the Code
Code specifies what a system does. Business rules define what the organization requires and why. Correct code cannot compensate for rules that serve the wrong goal.
Most rules reflect decisions made with the best information available at the time. Some remain correct. Others have outlived their reasons. The code cannot distinguish the two.
Rules also depend on one another and changing a process without understanding its purpose can disrupt operations, remove efficiencies and harm processes that were previously operating correctly. Understanding rule dependencies and examining their purpose can also reveal a more direct route to the goal.
A Record of Decisions
Properly conducted, this work produces two records. The first explains why existing systems and processes were built as they were. The second explains why their replacements are built as they are. Software practice calls this an architecture decision record. We call it a decision record because it documents business decisions as well as technical ones.
Problems cross departmental and system boundaries and affect customers. The records must follow those connections. AI makes documentation at that scale practical, as described in Complex Technology Problems Begin with Questions, Not Solutions. People responsible for the work must verify its accuracy.
Each solved problem extends the record. Future projects can then examine earlier decisions instead of reconstructing their reasons from code or memory.
Rules Make AI Governable
Data becomes information when organized. Information becomes knowledge when understood within the rules of the business. Knowledge becomes wisdom when experienced people apply it to decisions. AI processes data and information. Decision records preserve business knowledge. Experts supply the wisdom to apply it.
Recorded knowledge gives people and AI a common account of how operations should work. That account provides a basis for governance, but documentation alone does not enforce the rules.
Tests check whether the system follows those rules. Security controls restrict what AI can access and do. Monitoring checks whether processes run as designed and continue to meet the original metrics.
The Problem of Consistency
A CFO and a COO who ask the same question must receive consistent figures when they use the same definitions and data. Access controls determine which information each may see. A language model alone cannot guarantee consistent calculations.
Where a question has one correct answer, a tested algorithm should produce it. The language model interprets the question and explains the result. The algorithm supplies the figures. The application must validate the model’s interpretation before executing the calculation.
Each response should identify its data source, data snapshot and algorithm version. If the CFO asks in the morning and the COO asks in the afternoon, those records help explain any difference. Updated data may justify different figures. Different interpretations of the same business term do not.
A semantic layer supplies shared definitions and metric calculations. Reports and AI assistants then use the same definitions and calculations. In our view, this is becoming a standard approach to consistent AI reporting.
Guardrails, such as those in Amazon Bedrock, provide additional checks on questions and responses. They supplement tested calculations and access controls.
When There Is No Single Answer
Many questions have no single correct answer. Identifying a contract’s principal risks requires interpretation and professional judgment.
Claude and ChatGPT can identify issues and draft an analysis. A model trained on contract law might perform better on particular tasks. Neither replaces a lawyer. For questions of judgment, AI prepares a draft and a qualified expert evaluates and approves the result.
Exact calculations require tested algorithms. Questions of judgment require accountable reviewers. Both require evidence sufficient to examine the answer.
The Problem Comes First
AI produces a measurable return only when it solves a defined business problem. Solving that problem depends less on software code than on understanding the organization’s business rules and the goal they serve.
Select the problems of greatest value. Record the rules that govern them. Build AI around those rules.
Has your organization adopted AI without a measurable return, or does it face a problem that AI might help solve? Oxer can help.