AI WORKSHOPS FOR ORGANIZATIONS

AI trainings usually die on the 'what is a token' slide — I decided every module of mine starts from a real company document: tender documentation with attachments, a JIRA report, daily-meeting notes, a photo of a retro board. For each case I prepared a ready input, a system prompt as a checklist, and the expected output, so attendees don't watch a demo — they walk a process they'll repeat on their own data the next day. Building 17 such cases took more work than many a coding project, because each had to survive contact with a real document, not a staged example.

A decision maker won't write a line of code and shouldn't have to — yet most AI materials assume the opposite. I translate LLMs into the language of decisions: capacity planning instead of 'context window', a decision-meeting pre-read instead of 'completion', the cost of a bad decision instead of the cost of a token. Zero coding prerequisites was a hard design constraint — everything I show must work from a browser with a ready-to-paste prompt.

Some data may not leave the infrastructure — regulation, contract or security policy — and that's not a footnote, it's an adoption blocker. I run a dedicated local-models track: LM Studio on the attendee's machine, matching the model to the task, an honest conversation about what a local model can and cannot do. Paradoxically this module builds the most trust — it shows AI doesn't have to mean sending company documents out into the world.

An 80+ slide deck doesn't fit the model's context window — halfway through, Claude lost state and filling fell apart. So I split the build into three commands run in separate chats (init → fill × N → verify) and put the process state into the deck itself: a hidden slide holds JSON in its notes field, fill reads it at start, fills a batch of 25–30 slides, updates the phase and stops itself at ~60% of the window, asking for a new chat — checkpoint/resume instead of a single pass. The second trap was the Master itself: it has no fields for card descriptions, and shape ID order doesn't match the visual A/B/C order — a mechanical 'i-th shape = i-th card' would write content under the wrong letter. So I built a dictionary of 21 representative slides with shape roles and warnings embedded inline — fill must read the role, not the position, and descriptions with no fields go into speaker notes instead of breaking the OOXML geometry.

Training materials must comply with the digital accessibility act (WCAG 2.1 AA): no italics, expanded abbreviations, a jargon glossary, repeated table headers, correct language metadata — and with 323 DOCX files, manual proofreading doesn't scale. Instead of a human checklist I wrote a Markdown→DOCX converter (8 scripts, 1,351 lines of Python) that executes the accessibility rules as automatic OOXML transformations: italics become bold, abbreviations get expansions, the glossary builds itself, tables get w:tblHeader, and the document gets the corporate document theme and language=pl-PL. Regulatory compliance stopped being a stage — it's a property of the generator.