Find where AI is worth implementing.
Identify the processes where AI can create value, compare the options, and choose what to test first.
How to use the method
Bring a business process and a problem. Leave with candidate interventions, the evidence behind them, and a practical next decision. Start with quoting, fulfillment, service resolution, tenders, onboarding or planning, then use your own process names. You do not need an AI idea to begin.
- A process is slow, costly or unreliable, and you want to understand where AI could help.
- You need to compare process repair, integration, automation and AI before committing budget.
- You have several opportunities and need to separate pilots, enabling work and missing evidence.
01. Process and outcome
Define the process from its trigger to the result delivered to a customer. Name an owner, the current performance gap and the outcome worth improving.
A faster task matters when it improves the process outcome. Boundaries help reveal delays, shifted workload and double-counted benefits.
What begins and ends this process, who needs its result, and what is currently failing?
Starting with “we need a copilot” before establishing the problem, baseline or owner.
Request to approved quote: measure total turnaround, clarification loops, rework and critical omissions. Drafting time is one part of that process.
02. Constraint and evidence
Locate waiting, rework, exceptions and handoffs. Distinguish knowledge or interpretation problems from missing capacity, unclear authority, policy or incentives. Record competing explanations.
Friction does not establish AI suitability. The cause determines whether better information, a system connection or an organizational change could help.
What evidence suggests this is the constraint, and what else could explain it?
Treating a polished account or a long answer as verified operational evidence.
If engineering approval takes two days, inspect incomplete requests, reviewer workload and decision rights before proposing AI extraction. Mark each explanation as unknown, assumed, reported, observed or measured.
03. Intervention and alternatives
Compare process repair, integration, deterministic rules, predictive models and generative AI. Describe the mechanism linking each intervention to the outcome. Combinations are allowed.
Architecture and sourcing follow requirements. Retrieval, agents, fine-tuning, building and buying are later decisions, and do not describe mutually exclusive interventions.
Which feasible alternative can improve this outcome with the least unnecessary complexity?
Choosing an agent because the process has many steps, or AI because the process has much friction.
A quote process may need structured intake, integration, rules for price validation and AI-supported requirement extraction with source references. Engineering still approves exceptions.
04. Delivery conditions
Check input access, acceptable error, evaluation, human review, action authority, integration and adoption. A consequential missing condition becomes enabling work before deployment.
Potential value cannot compensate for an unresolved condition needed to use a system responsibly. A pilot must include representative cases and exceptions.
What must be true for this intervention to be tested and used in this process?
Averaging a serious deployment blocker into a readiness score, or testing only easy examples.
For extracted tender requirements, define critical-omission checks, source traceability, a human approver and a fallback. Evaluate difficult formats as well as routine requests.
05. Economics and next decision
Estimate benefits and total costs as ranges, with adoption and timing. Separate released capacity from cash savings. Choose the next measurement, enabling task or pilot and state what could change the decision.
A shortlist should guide investment and learning. An incomplete business case may be a useful opportunity that needs measurement, rather than a weak idea.
What is the smallest experiment that could change our investment decision?
Calling saved minutes cash savings, importing a vendor ROI claim, or counting the same benefit in multiple opportunities.
If a pilot reduces drafting effort but adds equal review work, no net capacity was released. Compare total process effort, quality and turnaround with the improved non-AI baseline.
The next decision
Use the process to discover opportunities, the intervention to compare options, and the process outcome to judge success. Every recommendation should name its assumptions and the evidence that could change it.
Sources
- APQC: Identifying and prioritizing AI use cases
Supports starting with a business problem, mapping processes and comparing candidate interventions.
- Dell’Acqua et al.: Navigating the Jagged Technological Frontier
Supports evaluating representative work and exceptions. Results from a consulting experiment with GPT-4 cannot predict performance in your process.
- NIST: AI Risk Management Framework Core
Informs checks for context, capability limits, oversight, evaluation and viable alternatives. NIST does not validate this tool’s recommendations.
Use and cite the method
This is a practical synthesis, not an empirically validated ranking model. There are no peer benchmarks or universal ROI assumptions.
The explanation and worksheet are openly available. When referencing the method in a briefing, article or workshop, credit Juan Beltrán and link to this page so readers can inspect the assumptions and sources.
Suggested citation: Juan Beltrán. AI Opportunity Map. 2026.
Updated: 2026-10-02.
Key takeaways
- Define the process from its trigger to the outcome someone needs.
- Source the baseline and suspected constraint; keep unknowns visible.
- Compare process repair, integration, rules, predictive models and AI.
- Establish access, accountability, evaluation and adoption before a pilot.
- Separate capacity from cash savings and define stop and scale conditions.