AI or a workflow fix? How to tell before you spend

Most problems put forward for AI turn out to be workflow, integration, data or software problems, and those are usually cheaper and more reliable to fix without AI. That is not an argument against AI. It is an argument for sorting the problem before choosing the tool. An operations or IT leader who can tell the difference will spend on AI where it pays, and fix the rest in ways that are easier to support.

The pressure to use AI is real. Boards ask about it, vendors have added it to nearly every product, and staff are already using general-purpose assistants whether or not there is a policy. The risk is not that organizations ignore AI; it is that they point it at the wrong problems, then conclude it does not work.

A typical example: a team is overwhelmed by incoming requests and someone proposes an AI assistant to handle them. When the requests are examined, most of them are people asking where something is. The real problem is that the process has no status anyone can see. A workflow that shows requesters the state of their request removes most of the volume, and no model is needed. The same pattern appears with reports nobody trusts (a data problem), information re-keyed between systems (an integration problem) and a step no system supports (a software problem).

Seven questions separate a genuine AI opportunity from the rest. Work through them for any proposed use before anything is bought or built.

Is the process itself defined? If the steps, rules and owners are unclear, AI will inherit the confusion and make it harder to see. Define the process first. If the work follows clear rules, ordinary workflow automation will usually do it more predictably than a model.

Is the input unstructured? AI is strongest where the information arrives as free text, documents, emails, images or speech and has to be read and interpreted: classifying incoming requests written in the requester's own words, extracting fields from varied documents, summarizing long case histories, or finding the relevant passage in a large body of internal guidance. If the input is already structured, such as form fields, database records or fixed codes, a rule or an integration is usually the better answer.

Is the data available, and is it allowed to be used? A model can only work with information it can reach. Establish where the data lives, whether it is accurate enough to rely on, and whether it can be processed by the tool being considered, given privacy, contractual and sector obligations. Decide which data the model can see, where it is processed, whether it is retained and who sees the output, before anything goes live.

What happens when it is wrong? Every AI system will sometimes produce a wrong answer. The question is what that costs and who notices. Where a person reviews the output before it takes effect, such as a draft reply, a suggested classification or a summary for a caseworker, occasional errors are manageable. Where the output acts directly on a customer, an account or a payment, the bar for accuracy and oversight is much higher, and many uses should not be automated at all.

Is there enough volume to matter? Building, testing and supervising an AI capability takes effort. If the task happens a few times a week, a person doing it well may be the better answer. The case for AI grows with volume, variety and the time each item takes to read.

Can it fit inside the workflow people already use? AI that lives in a separate tool, which people have to remember to open and copy results out of, rarely lasts. The value comes from placing the model inside a structured process: the request arrives, the model classifies or extracts, a person confirms, the workflow continues. That usually means the workflow and integration work matter as much as the model.

How will you know it is working? Agree the measures before the pilot starts: accuracy against a sample checked by people, time saved per item, share of items a person had to correct, and the effect on the process as a whole. Without a baseline, a pilot produces impressions rather than evidence.

AI earns its place where the input is unstructured, the volume is real, and a person can check the result. Where those three conditions hold, the gains can be substantial and the risks manageable. Where they do not, the money is usually better spent on the process, the integration or the data.

Governance does not need to be elaborate to be useful. The National Institute of Standards and Technology published its AI Risk Management Framework 1.0 in January 2023 as a voluntary framework, organized around four functions: govern, map, measure and manage. It is a practical reference for deciding who is accountable for an AI capability, what risks it carries in its context, how it will be measured and how issues will be handled. Organizations in regulated sectors will have additional obligations of their own, and those should be settled with the people responsible for compliance before deployment.

Label every piece of AI work as a prototype, a pilot or production, so everyone knows what they are relying on. A prototype shows something is possible. A pilot tests it on real work with close supervision and agreed measures. Production means it is supported, monitored and owned, with a way to handle errors and a way to switch it off. Confusing these stages is how organizations end up depending on something that was only ever meant as a demonstration.

In CnergyPro's delivery model, applied AI starts in Assess like any other change: understand the work, the data and the constraints. Architect decides whether the answer is workflow, integration, software, AI, or no new technology at all. Implement builds the chosen solution into the process with the oversight it needs. Optimize measures it in use and improves it, or retires it if it does not earn its place.

The organizations that get value from AI are rarely the ones that adopted the most tools. They are the ones that knew which problems they were solving.

CnergyPro helps organizations improve critical business processes, then configures, integrates, builds and optimizes the technology that makes the better process operational, including AI where it is the right answer and not where it is not. If you are being asked where AI fits in your operation, tell us what's not working.

Source: National Institute of Standards and Technology, AI Risk Management Framework, https://www.nist.gov/itl/ai-risk-management-framework (checked 25 September 2026).