A PowerPoint localisation pipeline that turned weeks into hours
Localising a blended learning programme used to mean rebuilding every tutor-led PowerPoint from scratch, one language at a time, including redrawing infographics with text baked straight into the image. I built a workflow in Claude that does the heavy lifting and keeps a designer in charge of the parts that matter.
The problem
CIPS serves a global membership, and its blended learning programmes need to reach professionals in more than one language. Getting there was painfully slow. For every language, a designer effectively rebuilt the tutor-led materials from scratch: extracting the text baked into each PowerPoint infographic, translating it, and redrawing the graphic by hand before the deck could go back to a tutor. It was expensive, it did not scale, and it quietly turned multilingual delivery into something the team avoided rather than offered.
The learner and the constraints
The people at the end of this are expert, time-poor professionals who expect content in their own language and to a professional standard. Nothing about being fast could change that. Every localised deck still had to be accurate, on-brand, accessible and safe to publish under a chartered body's name. Machine translation on its own was nowhere near trustworthy enough for that. Speed could not come at the cost of quality, so I designed for both.
The question I set myself: how do I make localisation an order of magnitude faster without taking the human judgement out of it?
The approach, and why
I did not treat AI as a translate button. I built a governed pipeline in Claude that takes on the mechanical work and leaves the decisions to a designer:
- Ingest. The pipeline reads an existing PowerPoint deck and extracts the text baked into each infographic, alongside the rest of the slide content, into a reviewable form.
- Translate in context. A custom AI skill translates the content with instructional intent preserved, using CIPS terminology and register rather than generic translation.
- Human-in-the-loop review. A designer checks accuracy, tone, terminology and cultural fit before anything is rebuilt. This gate is never skipped.
- Reassemble. The approved text is redrawn back into the infographics and repackaged into a production-ready PowerPoint deck that drops straight into the tutor-led session.
- QA and accessibility check. Final review confirms the deck reads correctly and still meets WCAG 2.2 AA.
Artefacts
Here's the pipeline itself: five stages, from raw CIPS slides to a finished, accessible deck.
Analyse
Extract every slide's content, text, notes, layout, into a single reviewable document.
Deck → structured contentTranslate in context
A custom AI skill rewrites each line for instructional intent, checked against the CIPS glossary.
Custom AI skillReassemble & rebuild
Every infographic and slide rebuilt natively, fully editable, batch-built in groups.
Batch buildHuman review
Each batch is checked for accuracy against the source before the next one begins.
Per-batch sign-offQA & accessibility
Final pass confirms correct reading order and WCAG 2.2 compliance across the module.
WCAG 2.2 checkFinal sign-off
I review the finished module myself before it goes to the SME.
Ready for SMEThe outcome
The pipeline brought localisation from weeks down to hours, roughly 80% faster than traditional methods: six decks localised in four days. Multilingual delivery went from something the team avoided to something it could offer as standard, and every published deck still had a designer's name on the sign-off.
What I took from it
The value was never the translation. It was the governance around it. Building the human review gate in from the start is the only reason this was ever safe to use on regulated content. And the time it saved did not disappear into doing more, faster. It went back into analysis, evaluation and design, which is where a designer actually earns their keep.
Further reading: the governance principles behind this pipeline, and why AI should support, not replace, the designer.