How ENEOS Materials made training analysis easier with AI

The problem.

ENEOS Materials collected feedback after employee training, but the work of combining and analyzing it limited what HR could learn from each session.

What changed.

The team introduced a custom GPT for training analysis. An HR colleague also used ChatGPT to build an internal data-aggregation tool without prior coding experience.

As described by OpenAI and ENEOS Materials.

What was reported.

OpenAI’s September 2025 account reports a 90% reduction in time spent on HR data aggregation and analysis. This is a result for that workflow, not the manufacturer’s total operating time. Source: OpenAI

What we take from it.

Recurring administrative work can be a useful starting point because the team already knows what a good result looks like. For training feedback, that might mean a complete set of responses, consistent categories, and a short account of issues worth addressing. Before introducing AI, we would write down those requirements and identify which parts are repetitive preparation and which need a person's judgment.

Those two kinds of work need different treatment. Moving approved columns into a standard table may be better handled by a simple rule. Interpreting open-ended comments may benefit from a language model, but the interpretation should remain connected to the original responses. A useful summary should let the reviewer inspect its evidence, recognize disagreement, and notice a serious concern mentioned by only one person.

Feedback also carries context that can disappear in a polished paragraph. A summary can flatten criticism or make a small group sound unanimous. We would compare generated themes with a human-reviewed sample, record omissions, and check whether the language overstates what respondents actually said. Anonymous responses should remain anonymous throughout the workflow, including in any exported reports.

The practical test is whether the team reaches a trustworthy decision with less total effort. That includes preparing the input, reviewing the output, correcting mistakes, and maintaining the workflow between sessions. A pilot should finish with something the owner can operate: a documented input format, clear access rules, an exception path, and an agreed review step. The strongest saving is work that stays manageable after the first demonstration.

How to evaluate a similar idea.

Start with your situation and a question you can test. These are evaluation steps we would discuss before choosing an implementation.

1. 01

   Choose one recurring task

   Follow a real feedback cycle from collection to review. Identify repeated preparation that delays a useful decision.

2. 02

   Record the whole baseline

   Measure preparation, analysis, checking and corrections together. Keep a small set of examples with agreed outcomes.

3. 03

   Keep the input controlled

   Use approved fields and access permissions. Remove unnecessary personal information before any model processes the responses.

4. 04

   Make summaries inspectable

   Let reviewers trace themes to source responses and retain minority concerns. Test omissions as well as incorrect statements.

5. 05

   Measure sustained usefulness

   Compare several cycles, including review time and maintenance. Keep the simpler process if it delivers the same reliable outcome.