Council Post: Enterprise AI Should Start With The Workflow, Not The Model
Eranda Maldeniya is Director of Product, Solutions and AI at Enterprise Analytics, with 20+ years of experience in enterprise systems.gettyEnterprise AI adoption often starts with the wrong question: “Which mo...
Eranda Maldeniya is Director of Product, Solutions and AI at Enterprise Analytics, with 20+ years of experience in enterprise systems.
getty
Enterprise AI adoption often starts with the wrong question: “Which model should we adopt?” A better starting point is, “Which business process are we trying to improve?” This is especially important for organizations that rely on complex enterprise platforms such as enterprise resource planning (ERP) systems.
I have seen organizations select an AI model first and then search for use cases to fit it. The same pattern is now happening with agentic AI, where teams look for places to deploy agents before defining the business problem. The result can be unclear outcomes, wasted time and difficulty proving the value of the investment.
The workflow is the real unit of value.
Enterprise AI is often evaluated at the task level. A model may extract information from an invoice, summarize maintenance instructions or predict the likelihood of a machine breakdown on the shop floor. These capabilities can be useful, but such actions alone do not guarantee a better business outcome.
For example, extracting invoice data is only one part of the accounts payable (AP) process. In practice, the harder challenge is the full AP workflow. Invoices must be collected from the AP mailbox, separated from supporting documents, matched to purchase orders, checked for taxes and charges, assigned the correct payment terms and routed for approval. If those steps remain manual, better extraction alone will not create much value for the finance team.
I have seen several AP automation projects where AI-enabled solutions were deployed on top of ERP systems but focused mainly on data extraction. Although the extraction worked, the wider process remained manual, making it difficult to demonstrate meaningful value to the finance team.
This is why AI should be assessed by the operational impact it creates. Did it reduce data entry time? Did it reduce the error rate? Did it resolve exceptions faster? Did it improve the quality or speed of a decision?
ERP systems can be the foundation for AI.
AI can generate an insight, prediction or recommendation, but enterprise systems are where that output becomes operational. ERP platforms often hold much of the business context, controls and transaction history needed to act on that output.
If AI is not connected to the existing workflow, someone still has to interpret the result and complete the action manually in another system. The insight may be useful, but the process remains fragmented.
Organizations generally have three options: use AI already embedded in the ERP platform, connect an external AI tool or build a custom solution for a specific workflow. There is no universal best option. The right choice depends on what fits the business, what the solution can realistically deliver, what it costs and how it will be used, not on which technology sounds most advanced.
Start with the problem, not the technology.
Every business function has operational challenges, such as manual data entry, inaccurate records, delayed closing, poor reconciliation, limited inventory visibility or recurring equipment failures. Once the problem is clearly defined, the organization can determine the right solution. In some cases, AI may be appropriate. In others, conventional automation, process redesign or better system integration may deliver more value. This avoids forcing AI into problems that may have a simpler solution.
Consider a treasury team that manually enters currency rates into the ERP system each day. This is primarily an integration problem, not an AI problem. On the other hand, repeated machine breakdowns on the shop floor may require a different approach. If maintenance teams struggle to identify early warning signs, AI-based anomaly detection could help support preventive maintenance and reduce unplanned production delays.
When AI is the right solution, the goal should be to use the smallest effective capability. A specialized model may be better suited than a large general-purpose model for tasks such as document extraction, classification or anomaly detection. It can be faster, more predictable and less expensive to operate. Don't deploy the most advanced AI if you can solve the problem reliably and measurably.
Measure the workflow before and after.
A measurable AI rollout starts before the technology is selected. The organization first needs to define how the workflow performs today using measures such as cycle time, error rates, manual effort or cost per transaction. Without this baseline, it is difficult to prove that AI has improved the process.
McKinsey’s five-layer framework supports this approach by linking technical performance to user adoption, operational KPIs, strategic outcomes and financial impact. A model may work well technically, but it creates little value if people do not use it or if the workflow does not improve.
Each initiative should therefore have one clear business outcome. This could be a higher straight-through processing rate, faster invoice registration, fewer planning exceptions or shorter contract review times. Technical measures such as accuracy, response quality and token usage still matter, but they should support the business outcome rather than replace it. In my experience, a proof of value is more useful than a proof of concept because it shows not only that the technology works, but that the business performs better.
The Takeaway
Enterprise AI should begin with a business problem, a clearly understood workflow and an outcome that can be measured. The right solution may be embedded AI, connected AI, a specialized model or no AI at all. What matters is whether it fits the process and improves business performance.
From what I have seen, the best results do not always come from the most powerful model. They come from applying the right capability to the right workflow and proving that the business performs better.
Forbes Technology Council is an invitation-only community for world-class CIOs, CTOs and technology executives. Do I qualify?