AI Opportunities
How to identify AI opportunities in your business
A repeatable method for finding AI opportunities: start from workflows and cost, not from technology. Includes signals to look for, a discovery sequence and disqualifiers.
Miguel Torres, Founder, Merjora · Published 18 August 2026 · Updated 3 September 2026
In short
Identify AI opportunities by starting from workflows rather than tools: list the processes that consume the most repetitive human hours, isolate the decision or document step inside each one, estimate today's volume, handling time and error cost, then score each candidate on value, data availability, risk and organisational readiness. The opportunities worth running are high-volume, rule-adjacent, text- or data-heavy steps where the current cost is measurable and a human can still review the output.
Key takeaways
- Opportunity discovery is a costing exercise before it is a technology exercise.
- The unit of analysis is a workflow step, not a department or a tool.
- Volume × handling time × loaded hourly cost gives you the baseline that every later ROI claim depends on.
- Half of a good discovery output is a list of things you decided not to do, and why.
Start from where time and money actually go
The most reliable input to AI discovery is an honest inventory of repetitive work. Ask each team leader which tasks their people would drop tomorrow if they could, how often those tasks run, and how long each instance takes. You are looking for volume and repetition, not novelty.
Anything that cannot be described as 'X times per month, Y minutes each, done by someone paid Z' is not yet an opportunity. It is an idea. Ideas are fine, but they do not belong in a prioritized portfolio until they are quantified.
- Inbound requests that are read, classified and routed by hand
- Documents that are re-keyed from one system into another
- Standard replies assembled from templates plus a lookup
- Recurring reports compiled manually from two or more sources
- Follow-ups that depend on someone remembering to send them
Isolate the step, not the process
'Automate claims' is not an opportunity. 'Extract the eight fields on a first-notification-of-loss document and pre-populate the claim record, with an adjuster confirming before submission' is an opportunity: it has a boundary, an input, an output and a reviewer.
Narrowing to a step also makes failure survivable. If extraction accuracy is 88% rather than 97%, the human reviewer catches it. If you attempted the whole process end-to-end, the same accuracy would be an incident.
Signals that a step is a genuine candidate
| Signal | Why it matters |
|---|---|
| High frequency, low variance | Value accrues from repetition; variance drives exception handling cost |
| Text, documents or structured records as input | Language and extraction models are strongest here |
| An existing quality bar | You need something to measure against, ideally already tracked |
| A natural human checkpoint | Review keeps residual error economically containable |
| Data already reachable in a system | If the data needs a new integration, cost and timeline roughly double |
| A named owner | Unowned processes do not get adopted after go-live |
Disqualifiers worth respecting
- The current process is undocumented and disputed between teams — fix the process first.
- The output drives a legally consequential decision with no practical review step.
- Volumes are low enough that even perfect automation saves less than the run cost.
- The required data sits in a system nobody is willing to integrate with this year.
- The only stated benefit is 'innovation' or 'being seen to use AI'.
A discovery sequence you can run in two weeks
- Week 1, days 1-2: workflow inventory interviews with four to six team leads.
- Week 1, days 3-5: quantify volume, handling time and error/rework cost for each candidate step.
- Week 2, days 1-2: score candidates on value, data availability, risk and readiness.
- Week 2, day 3: rank, and explicitly park everything below the line with a stated reason.
- Week 2, days 4-5: write the business case for the top two, including the measurement plan.
Write down the negative recommendations
A discovery output that recommends everything is worthless. The parked list is what protects the programme six months later, when someone asks why a plausible-sounding idea was not pursued.
Record the reason in the same language as the positive cases: insufficient volume, unavailable data, unacceptable residual risk, or readiness gap. Reasons expire, so revisit the list once a year.
Frequently asked questions
- How do you identify AI opportunities without technical expertise?
- Discovery is mostly operational, not technical. Anyone who can quantify volume, handling time and error cost for a workflow step can produce a credible candidate list. Technical input is needed later, to validate data availability and integration effort.
- Should opportunity discovery start top-down or bottom-up?
- Both, but bottom-up produces better candidates. Leadership sets the themes and the constraints; the people doing the work know where the repetitive hours are. Reconcile the two lists before prioritizing.
- How many opportunities should a mid-sized company expect to find?
- A typical 50-500 person company surfaces between eight and twenty candidate steps in a first pass, of which two or three usually justify immediate investment. The rest form a backlog with stated conditions for revisiting.
Where would this apply in your business?
Merjora maps your workflows, scores each opportunity and quantifies the likely value range before you spend budget.
Discover your AI opportunitiesRelated reading
Editorial standard. Merjora publishes analysis, frameworks and publicly documented examples. We do not publish invented statistics, unattributed benchmarks or unverified customer stories.