Automation vs AI in training: rules or judgment

Both words cover a lot, so this post narrows them. Automation here means rule-based automation: if a person joins a role, assign this lesson. The same input should give the same output, so you can test each rule against your own cases and keep checking it as roles and lessons change. AI here means generative drafting: a model that writes a lesson, a question or a translation from a source.

The NIST AI Risk Management Framework describes AI systems more broadly, as machine-based systems that, for a set of objectives, generate outputs such as predictions, recommendations or decisions. It adds that AI systems operate with varying levels of autonomy. Many products mix the two, using rules to route work and a model to draft it.

A step is a rule job when you can write the decision down in advance and it does not change from person to person. These steps are worth putting on rules first, because they are easy to check and cheap to fix.

• Assigning a lesson when someone joins a role or a location.
• Sending a nudge when an assigned lesson has not been opened by a set date.
• Recording who finished which version of a lesson.
• Moving a person into a new group after a promotion.
• Retiring an old version of a lesson on a set date.

Other steps depend on meaning, context or the person in front of you, and no simple rule covers them. Some suit generative drafting with a person checking the result, and some should stay with a person entirely. OSHA's 2010 policy statement says training required by its standards must be presented in a manner employees can understand, using a language and vocabulary they understand. The same idea applies to any lesson: understanding is a judgment, not a checkbox. Elvina does not replace required safety training, but the principle shapes how every lesson should be built.

• Turning a messy procedure or a recording into a clear lesson.
• Writing quiz questions that test a decision, not memory of a sentence.
• Adapting a lesson for another language, or for staff who read little.
• Noticing that the work changed and the lesson did not.
• Deciding whether a person is ready to work alone.

List the steps in one training flow, from hire to working alone, and sort each one with the table below. The third column matters most. If you cannot say why a step belongs where you put it, it probably needs a person.

Putting a judgment step on rules makes it rigid. A rule that marks someone ready after a passed quiz will pass people who memorized the answers. Using a generative model for a rule step adds risk without adding value. The NIST framework notes that failure modes of large pre-trained models are harder to predict, and that AI is harder to test than traditional code. Assigning a lesson does not need that uncertainty. So keep rules where rules work, and put AI where a draft spares real effort and a person can check it.

Three points in a training flow should always belong to a person. The first is approval: nothing drafted by AI reaches staff until its owner signs off. The second is readiness: a manager decides when someone can work alone, using quiz scores as evidence, not as the verdict. The third is exceptions: when a case does not fit the procedure, a person answers and decides whether the source needs an update. Write these three names down for each role before you automate anything. If one person holds all three, check that they have the time, or the approval step will quietly turn into a rubber stamp.

Start with the rule steps that eat the most admin time and carry the least risk. Assignment at hire usually comes first, because it runs for every new person and a mistake is easy to spot. Recording completions comes next, since records kept by hand drift out of date. Leave AI drafting until your sources are current and each one has an owner, or the drafts will repeat old mistakes. Keep readiness decisions with managers from the start, and never let a rule mark a person ready.

Elvina keeps the same split, and each part has its own status today. Assignment, groups and progress by person run on the Learning Platform and are available now. Reminders and progress digests are planned, not available today, and so are reports to managers on a schedule. On the drafting side, Creation Studio turns sources approved in Company Knowledge into lessons, job aids and quizzes, and translates them, all available now. Approval of every piece by a named person before release is available now. When a source changes, a person starts the redraft, and right now our team does the drafting with you on the platform.

A fictional orchard with a farm market and a cider barn hires seasonal staff every fall. A few rules change every year, such as which varieties are ready and how picking bags are handed out.

• A rule handles the routine: each new hire gets the market or orchard lessons for the role on the first day.
• Generative AI drafts this season's picking lesson and the market checkout lesson from the manager's updated notes and a short recording.
• The orchard manager approves the drafts and corrects the rule about which rows are open for picking.
• A field lead decides when each picker can work a row alone, after watching a short shift.
• When a customer asks about a variety that no lesson covers, the field lead answers and adds it to the notes for next year.

Measure the rule side and the judgment side separately, because they fail in different ways.

• Rule steps: assignments that went to the wrong person, and nudges sent after a lesson was done.
• AI steps: corrections the approver makes per draft, and how many drafts are sent back.
• Person steps: time from hire to the manager's ready decision, and how often that decision is reversed.
• Questions on shift that no lesson covers.
• Hours the manager spends each week on training admin, before and after.

What is the difference between automation and AI? In training, the useful comparison is rule-based automation versus generative drafting. Rules run steps defined in advance, such as assigning a lesson at hire. Generative AI drafts new text from a source and can be wrong. Many tools combine both, so put assignment and records on rules, use AI for drafting and keep approval with a person.

• Should training nudges run on a rule or on AI? Use a rule. A nudge depends on a date and a status, so a simple rule is easier to test and cheaper to run. Save generative AI for steps that need wording or adaptation, such as drafting a lesson from a procedure, and have a person check the result.
• Can AI decide when an employee is ready? It can collect evidence, such as quiz scores and finished lessons, but a person should make the call after watching the work. In Elvina, readiness checks in the first week are available now, detailed quiz answers and scores are not stored, and the manager's sign-off remains the decision.

Public sources, checked on the date shown.

• NIST AI 100-1 defines an AI system as a machine-based system that generates outputs such as predictions, recommendations or decisions, with varying autonomy. Source: https://nvlpubs.nist.gov/nistpubs/ai/NIST.AI.100-1.pdf (checked 2026-10-02).
• NIST AI 100-1 notes that failure modes of large pre-trained models are harder to predict, and that AI is harder to test than traditional code. Source: https://nvlpubs.nist.gov/nistpubs/ai/NIST.AI.100-1.pdf (checked 2026-10-02).
• OSHA's 2010 policy says required training must be presented in a manner, language and vocabulary that employees can understand. Source: https://www.osha.gov/laws-regs/standardinterpretations/2010-04-28 (checked 2026-10-02).