Choose a first AI initiative from a recurring task with known inputs and an output your team can review. A useful starting point has a bounded scope, errors you can detect and correct, and a result you can compare with the current process. Define the problem before selecting software.
Your shortlist might include report summaries, enquiry classification, or preparing website updates for an editor. Compare the workflows and their conditions rather than tool names. Use the business AI readiness checklist if ownership and permitted inputs are still unclear.
Shortlist three tasks
Ask staff what repeats, where information is copied between places, and which mistakes regularly need correction. For each opportunity, record a one-sentence description, frequency, input source, and the person using the output. Without a recognizable recurring burden, automation may not be the priority.
Set aside decisions with direct consequences for people’s rights or financial commitments until the relevant specialists have assessed requirements. Preparing a draft for review has different consequences from sending it or acting on it automatically.
Use a comparison you can explain
This is a hypothetical planning example, not a client assessment or a universal ranking.
| Candidate task | Starting opportunity | Prerequisite |
|---|---|---|
| Summarize an approved internal report | A short output can be checked against its source | Required information and a qualified reviewer |
| Categorize contact requests | Repeatable work with defined categories | Permitted samples and an unresolved category |
| Send quotations automatically | Fewer steps, but an external business consequence | Pricing rules, approval, and exceptions; begin with reviewed drafts |
Resolve the choice with three questions: are suitable inputs available, can a reviewer recognize mistakes, and can the pilot stop without disrupting the process? If two options are similar, prefer the one with fewer unprepared integrations. This is a planning judgment, not a formula that predicts return.
Define the boundary
- Input: one document type or request channel initially.
- Output: a draft or classification with a reason, without automatic external sending.
- Reviewer: someone who understands the work and can reject the result.
- Exceptions: incomplete or unusual cases return to manual handling.
- Acceptance: required information is retained, the format is usable, and reviewer changes are recorded.
- Stopping: a defined response to repeated material errors or unauthorized data use.
Why we began with our own website
In our internal work, website content and theme changes had inspectable outputs: a published page, metadata, a translation, or a verified file hash. That made review more specific than an open-ended objective such as “run the company with AI.”
For example, we added the new service to the iKetch homepage in Arabic and English and checked links, the main heading, and FAQs. People retained responsibility for the message, design review, and execution decision. The internal iKetch Platform case study documents that process. A completed release is not a measurement of time saved.
Decide what follows the pilot
Review accepted cases, rejected cases, and correction effort. If results are usable, expand one variable, such as a document type or user group. If every output needs extensive rewriting, reassess the task, inputs, and method before purchasing more subscriptions. See the AI results measurement guide for the comparison record.
Questions about the first initiative
Should we start with the biggest opportunity?
Choose useful value you can test within available boundaries. A large project dependent on unprepared data may belong later, after a smaller pilot builds operating experience.
When is custom development necessary?
When the workflow requires structured reading or writing across systems, after checking existing capabilities. Define data scope, permissions, and acceptance tests before commissioning an integration.
What should we bring to a scoping discussion?
Three candidate tasks, their current problems, non-sensitive examples, and the person who reviews the output. We can compare the options and define a starting scope through iKetch Managed AI.
Reference for considering context and risk before implementation: the NIST AI RMF core functions. The comparison and pilot boundaries above are iKetch working guidance, not a NIST-approved assessment.
