Short answer
To find AI opportunities in your company, start from the information work teams repeat, not from a list of tools. Assess each candidate by volume, current effort, cost of errors, data access and human oversight, and choose the first pilot from a narrow task whose output is easy to review.
Start with processes
Ask teams to list repetitive information tasks rather than AI tools. Reading requests, classifying documents, researching and drafting are separate opportunities.
Information work often looks like this: reading an incoming request and routing it, extracting fields from a document, searching similar past records, drafting a reply or report and moving data between two systems. These tasks repeat and partly rely on unstructured text, which is where AI most often creates value.
A simple assessment matrix
Record volume, current effort, error cost, data access and review needs for each process. A high-volume task without accessible data may not be ready for a pilot.
When filling in the matrix, take a short sample instead of guessing figures: have the team note for a few days how often the task is done and how long it takes. Estimates rarely reflect the real picture.
Example: a consulting proposal
Past proposals may help with structure and service information, but permissions and current prices still matter. The first pilot might only draft content while staff retain sending authority.
The same logic applies elsewhere: classifying support requests, extracting certain clauses from contracts or drafting a task list from meeting notes. In each case, decide first who will review the output and how.
Make the selection explicit
Document the chosen use case, required data, measurement plan and stop criteria. Defining failure conditions before starting prevents an unsuitable solution from expanding.
Add a "what we will not do" section to the selection output. Writing down which processes are outside this pilot, and why, stops expectations from growing during the pilot.
Opportunity assessment matrix
Fill in the five criteria below for every candidate process. The table is not a scoring method but a framework that structures the discussion.
A practical evaluation exercise
In a sample working week, record tasks involving information search, drafting and data transfer. Assess volume, variability, error impact, data access and human review for each candidate. Choose a narrow task with reviewable output for the first pilot, and name a process owner and success criterion before building.
At the end of the weekly sample you should have at least three candidates; fill in the same five criteria for each and review the results with the process owners. If the process owner does not agree, that candidate is not suitable for a first pilot.
Frequently asked questions
The questions we hear most often when identifying AI opportunities.
Do we need to collect data before anything else?
Not always. First decide which process suits AI, then check whether the data that process needs exists. If data is scattered, the first step of the pilot may be data preparation.
Which tasks are not suited to AI?
Calculations that can be written as exact rules, high-impact decisions whose errors cannot be reversed, and work without data access are usually unsuitable for a first pilot. AI or rule-based automation? explains the distinction in detail.
How long should a pilot last?
It depends on how narrow the scope is. What matters is writing the success and stop criteria before starting; the duration should be the shortest period needed to reach those criteria.
How is success measured?
By comparing the starting time, errors and review effort with the values at the end of the pilot. See How to measure benefit and total cost in an AI pilot for details.
Sources and further reading
