Implementing AI in a small business, one workflow at a time

An operations leader or an L&D manager who has approval to try Elvina and now has to make it real without a dedicated project team. This is the week after a yes. The usual mistake is starting with an integration project instead of one workflow that someone will actually use next month.

Discovery leads to a source/access review, bounded draft, human approval, authorized delivery, observed result and a decision about expansion. NIST's AI Risk Management Framework asks organizations to understand and document an AI system's intended purposes and deployment settings, and to define and document processes for human oversight.

• Discovery: pick one recurring task with a business owner, an approved source and a success criterion a person can check.
• Source and access review: add the approved source to Company Knowledge with its owner and version, and confirm who may see the output.
• Bounded draft: produce one artifact in Creation Studio, such as a lesson and quiz, and have the owner review and edit it.
• Authorized delivery: assign it to a small group in the Learning Platform, observe the result for two weeks, and decide whether to expand.

A fictional plumbing services company with 70 technicians chooses the water heater install checklist as its first task. The field operations manager owns it, the checklist v4 is the source, and the success criterion is that eight new technicians complete a supervised install without a callback in the first month.

• Week one: checklist v4 sits in Company Knowledge with the manager as owner; a lesson, a job aid and a six-question quiz exist as reviewed drafts.
• Week two: the eight technicians are assigned in the Learning Platform and their completion is tracked by content version.
• Week six: the manager compares supervised install outcomes and decides to add the next checklist rather than five at once.

Confirm the actual product operations and support responsibilities before rollout. The website does not provision your environment or connect systems.

• One task, one owner and one source are written down before any content is produced.
• The first artifact was reviewed and edited by the owner before assignment.
• The expansion decision cites an observed result, not completion counts alone.

This page describes the intended sequence and does not provision your environment, connect systems or set support responsibilities. Confirm the product operations available to your workspace and who handles support before rollout. Every artifact stays a draft until a person approves it, and delivery to learners is a deliberate action by your team.

How many workflows should the pilot include? One. A second workflow starts only after the first produced an observable result and a decision to continue.

• Who needs to be involved from IT? Someone who can confirm source access and permissions for Company Knowledge. A first workflow built from documents needs no system integration.
• How long should the first AI workflow run before we expand? Assign the first lesson and quiz to a small group, observe the result for two weeks, and then decide whether to expand. The expansion decision cites an observed result, not completion counts alone.

Public sources, checked on the date shown.

• NIST's AI Risk Management Framework, MAP 1.1: “Intended purposes, potentially beneficial uses, context-specific laws, norms and expectations, and prospective settings in which the AI system will be deployed are understood and documented.” MAP 3.5: “Processes for human oversight are defined, assessed, and documented in accordance with organizational policies from the GOVERN function.” Source: https://nvlpubs.nist.gov/nistpubs/ai/NIST.AI.100-1.pdf (checked 2026-10-26).