Workflow evidence
Shadow users and sample records to map triggers, inputs, judgement, systems, handoffs, waits, rework, exceptions, outputs, and downstream consequences.
Enterprise AI discovery
How should an enterprise identify the right AI opportunities?
Start from costly or consequential operating decisions, not a catalogue of AI capabilities. Observe the real workflow, establish a baseline, identify the constraint, test whether AI can change it, and rank opportunities by achievable value, evidence cost, adoption path, and downside.
By Onyx Devs Lab · Published · Updated
ObserveReal work, decisions, exceptions, and baseline
DiagnoseConstraint, intervention, data, and adoption
PrioritiseValue, risk, evidence cost, and reversibility
Find a decision worth changing
A frequent task is not automatically valuable, and an impressive output is not automatically operational improvement.
Baseline and constraint
Measure volume, cycle time, cost, quality, loss, delay, review burden, and variation; identify the binding constraint rather than automating every visible step.
Intervention fit
Decide whether the constraint needs deterministic software, process change, better data, search, prediction, generation, optimisation, an agent, or no new technology.
Build a comparable opportunity portfolio
Each candidate should be expressed as a falsifiable intervention with a bounded next evidence step.
Achievable value
Estimate the reachable population, adoption, time saved or loss avoided, quality effect, implementation and operating cost, delay, and attribution boundary.
Feasibility and risk
Assess representative data, ground truth, integration, identity, permissions, legal and sector controls, failure consequence, reversibility, support, and owner capacity.
Learning value
Prefer tests that resolve a material uncertainty quickly and generate reusable knowledge, even when the result is to narrow or stop.
How to prioritise an AI opportunity portfolio
Collect workflow evidence
Interview owners and operators, observe work, sample normal and exceptional records, and document actual decisions and consequences.
Write an intervention hypothesis
State user, trigger, baseline, proposed change, expected measurable effect, necessary data, failure consequence, and disconfirming evidence.
Score with ranges
Compare value, feasibility, risk, adoption, strategic leverage, evidence cost, and reversibility without hiding uncertainty in one precise number.
Fund the next evidence step
Approve a bounded assessment, prototype, or pilot with owner, budget, decision date, acceptance thresholds, and stop conditions.
Frequently asked questions
Should ideas be collected through an employee survey?
Surveys can reveal friction, but observation and record sampling are needed to distinguish perceived inconvenience from a valuable, changeable operating constraint.
Should the highest ROI estimate go first?
Not automatically. Include confidence, risk, dependencies, time to evidence, adoption, reversibility, and whether the benefit can actually be attributed.
How many opportunities should be piloted?
Only as many as the organisation can own, evaluate, integrate, and make decisions about. A small evidence-bearing portfolio is better than many unattended demos.
Start with one concrete operating problem.
We will first clarify the workflow, evidence, data boundary, and outcome standard before recommending the next step.
Contact Onyx