The least useful AI workshop begins with a blank whiteboard and the question, “What could we do with AI?” The room quickly fills with ideas: chatbots, agents, forecasting, document generation, customer assistants. Most are technically possible. That does not make them good investments.
A useful assessment works in the opposite direction. Start with expensive or frustrating work. Quantify it. Understand the data and decisions involved. Then determine whether AI is actually the right tool.
Build an opportunity inventory from real work
Interview people who perform the process, observe the workflow, and sample actual transactions. Capture frequency, average handling time, wait time, error/rework rate, systems touched, information types, decisions, exceptions, and required approvals.
High-frequency tasks involving unstructured information are especially interesting: reading requests, extracting fields from documents, comparing specifications, summarizing project updates, searching knowledge, classifying tickets, or drafting repetitive communications.
Score value separately from feasibility
A high-value use case may be difficult because the data is inaccessible, permissions are unclear, or the decision carries significant risk. A technically easy use case may save almost no time. Score them separately.
A simple model can rate annual labor opportunity, cycle-time impact, quality impact, customer impact, data readiness, integration complexity, model risk, regulatory risk, and change effort. Plot value against feasibility and prioritize the upper-right quadrant.
Choose AI only where uncertainty is useful
Use deterministic automation for deterministic rules. If “when field A equals X, create task B” solves the problem, do not insert a language model. Use AI where the input is ambiguous or unstructured: language, documents, images, semantic search, summarization, extraction with variation, or reasoning across context.
Many strong solutions combine both. AI interprets the request; rules validate it; APIs execute the transaction; a person reviews exceptions.
Design the evaluation before the pilot
Define a test set representing normal cases, difficult cases, and known exceptions. Decide how correctness will be measured. For extraction, measure field accuracy. For retrieval, measure whether the correct source was found. For classification, measure precision and recall. For generated summaries, evaluate factuality and omission of material issues.
Do not declare success because ten curated demonstrations worked.
Calculate the economics with review time included
If a task takes 20 minutes today and AI reduces preparation to 3 minutes but requires 5 minutes of human review, the saving is 12 minutes, not 17. Multiply the net saving by annual volume and adoption. Add infrastructure, licensing, integration, support, and governance costs.
Then consider non-labor value: faster response, improved consistency, increased throughput, better auditability, or the ability to perform analysis that was previously too expensive to do at all.
The practical objective is not more technology. It is a better-performing operation with clearer ownership, less friction, and technology that can be supported over its full lifecycle.