How to Use AI to Create Training Materials in 2026: Step-by-Step Guide with Prompts

You can use Perplexity to gather industry standards with citations, tl;dv to turn expert interviews into structured transcripts, ChatGPT to design the curriculum, Claude to write the modules, Gamma and Canva AI to produce the slides and job aids, Synthesia to film the training video, and Quizlet to generate the assessments. That is how to use AI to create training materials in 2026, and this guide walks through the full workflow with copy-paste prompts for every step.

• A six week production cycle compresses to about one week. Industry analysis of L&D teams using AI reports course creation time reductions of 40 to 60 percent, with a 30 minute course that once took three to four weeks now shipping in days, and every step below is sized for that reality.
• The stack costs almost nothing to start. Every tool in this guide has a genuine free tier, and the first paid upgrade most teams make is Synthesia Starter at $29 per month, against agency quotes of $10,000 or more for a single custom module.
• Grounding beats guessing. Each step feeds the AI real source material: policy documents, expert interview transcripts, verified process notes, and the prompts in this guide instruct every tool to use only supplied facts, which is what keeps the training accurate enough to certify people on.
• Companies are already moving. The corporate e-learning market reached $108.2 billion in 2025 on the way to a projected $121 billion in 2026, and industry surveys show 37 percent of companies now use AI in their training programs, up from 25 percent in 2024.
• Top rated picks for this task: ChatGPT leads the stack at 4.7/5 in our database, Claude and Gamma follow at 4.6/5 each, and tl;dv matches them at 4.6/5 for the interview step that most teams skip and most trainings suffer for.

Training production has always been a bottleneck disguised as a priority. Managers agree employees need onboarding, compliance refreshers, and skills training, then the materials sit in a queue for months because building them means pulling subject matter experts off revenue work for interviews, then waiting on writers, designers, and video crews. The economics have been brutal at every company size, and the scale of the market reflects it: the corporate e-learning market was valued at $108.2 billion in 2025 and is projected to reach $121 billion in 2026, much of it spent on production labor that AI now performs in minutes.

The time savings are no longer anecdotal. Analysis of L&D teams using AI in content production reports course creation time reductions of 40 to 60 percent, with a 30 minute course that previously required three to four weeks of work now completed in a fraction of that, and separate ROI studies report AI training tools cutting onboarding time by up to 50 percent while raising trainer capacity two to three times. Adoption has followed the results: industry surveys show 37 percent of companies now use AI in training programs, up from 25 percent in 2024, and the pressure keeps rising because the World Economic Forum projects that skills will change by 68 percent by 2030 once generative AI acceleration is counted, which means the half life of training content keeps shrinking and manual production cannot keep up.

There is a second benefit that has nothing to do with speed: consistency and reach. An AI-produced module delivers the same message to the new hire in Austin and the technician in Manila, updates everywhere at once when the process changes, and generates versions in dozens of languages from the same script through Synthesia or HeyGen at no extra filming cost. In our database of 300 tools, the education, presentation, and video categories alone hold dozens of free-tier options, and the eight tools in this workflow, rated between 4.2 and 4.7, cover every job from research to assessment. The rest of this guide shows the exact sequence, with the prompts to copy.

This step produces the raw material every later step depends on, because AI-generated training is only as good as the knowledge you feed it. Start by assembling what already exists: process documents, policy PDFs, slide decks from past sessions, and the notes of whoever currently explains the job to new hires. Then fill the gaps. First, open Perplexity, switch to Pro Search, and enter this prompt, replacing the example role with yours:

I am building internal training for new customer support agents at a mid-size SaaS company. Research the current industry standards for this role: the core skills training programs cover, the typical 30-60-90 day expectations, common failure modes for new agents, and how leading companies structure the first two weeks. For every fact give me the source name, the year, and a link, and flag anything where sources disagree.

Next, record a twenty to thirty minute interview with the person who knows the job best, because internal reality beats industry averages. Open tl;dv, connect it to your Zoom or Google Meet call, and let it record and transcribe automatically. Structure the interview around three questions: what do new people always get wrong in the first month, what does good look like on this specific team, and which resources do veterans keep returning to. After the call, tl;dv hands you a timestamped transcript in minutes, and its AI summary organizes the pain points and the vocabulary your best people actually use, which is the language your training should speak.

Save the transcript, the Perplexity research, and every existing document into one source folder in Notion AI. Perplexity is free to start with Pro at $20 per month, and tl;dv offers unlimited recordings and transcripts on its free plan, so this step costs nothing but the interview time. Do not skip the interview: the expert answers become the grounding document for Steps 2 and 3, and they are the difference between training that sounds right and training that is right.

This step converts raw knowledge into a learning structure, which is the skeleton every asset hangs on. Open ChatGPT, paste in your source folder contents or the interview summary from Step 1, and use this prompt:

Here is my source material for a training program on [topic] for [audience, for example: new warehouse associates]. Design the curriculum. Give me: 3 to 5 learning objectives written as observable behaviors using action verbs, a module breakdown of 4 to 6 modules with a one sentence purpose for each, the key points per module drawn only from my source material, and one real-world scenario per module from my transcripts. Mark anything my sources do not cover with a GAP flag so I can interview my expert again.

[paste your source material and interview summary here]

Push back on the draft before accepting it. Ask which module a skeptical line manager would cut first, request the sequence that gets a new employee productive soonest rather than the sequence that reads most logical, and check every learning objective against what the business actually measures. This iteration is where ChatGPT earns its 4.7/5 rating in our database, because curriculum design is a conversation with constraints, not a one-shot generation. Lock the module structure, then paste it at the top of every prompt you run from here on so all later outputs inherit the same skeleton.

For scheduling the production work itself, Taskade turns the module list into a project with owners and deadlines in one import, free to start with Pro at $10 per month. ChatGPT costs nothing to begin, with Go at $8 per month and Plus at $20 per month removing the usage caps free plans apply at peak hours. The deliverable of this step is a one page curriculum map: objectives, modules, key points, scenarios, and the gaps you still need the expert to close.

This step produces the reading core of the training: the module texts, the participant workbook, and the facilitator guide. Start with module one in Claude, because it sustains tone and structure across long-form documents better than any competing model, which is why it holds 4.6/5 in our database. Create a Project for the training, paste in the curriculum map from Step 2 and the source documents, and use this prompt:

Write the full text for Module 1 of my training program using the curriculum map and sources in this Project. Structure it as: a short why this matters intro written from the pain points in my interview transcript, the core content in sections that follow my module key points, one worked example using the real scenario from my sources, a common mistakes box with 3 errors new people actually make, and a summary of 5 takeaways. Rules: use only facts from the supplied sources, reuse the exact terminology my expert uses, keep the reading level conversational, and keep the total under 1,200 words. Flag any place you were tempted to invent a detail.

Repeat the prompt for each module, one call per module, because single-module generations stay more accurate than one attempt at the whole course. After the module texts, run a second prompt for the participant workbook: the exercises, the checklists, and the fill-in sections employees complete during the session. Run a third for the facilitator guide: timing per section, the discussion questions, and the answers to the workbook exercises, because the person delivering the training is often not the person who wrote it.

Claude Pro costs $20 per month, or $17 per month billed yearly, and the free tier covers a shorter program if you work within daily limits. Route every finished module through the subject matter expert from Step 1 for review before moving on, and file the approved version back in the Notion AI source hub with a date, because review order matters: Steps 4 and 5 turn these texts into slides and video, and regenerating those downstream assets after a factual fix is the kind of rework a review pass prevents.

This step converts the written modules into the visual layer: the session deck, the printed job aids, and the diagrams that make procedures stick. Start in Gamma, choose generate from text, paste in one module of your approved Step 3 content, and use this prompt:

Create a 10 slide training deck from my Module 1 content. Slide one states the learning objective as a question the learner cares about. Then cover the core content with one idea per slide, using the real scenario from my source as the worked example, a common mistakes slide, and a closing recap slide with the 5 takeaways. Keep each slide under 35 words, suggest one simple visual per slide, and match the terminology of my source exactly.

Gamma assembles the full deck in one pass, and the one click restyle feature lets you test visual themes before committing, which is why it earns 4.6/5 in our database as the fastest prompt-to-deck tool. It is free to try with 400 credits, with Plus at $9 per month and Pro at $18 per month. Export to PowerPoint for anyone who presents from their own machine, and keep the deck linked to the module text so both update together.

For the job aids and reference cards, open Canva AI and use Magic Design to produce one page summaries sized for printing and for screen: the process map, the escalation flowchart, and the checklist people pin at their desks. Save the colors and fonts in the Brand Kit so every aid matches, and let the bulk translate feature produce the Spanish, Mandarin, or Portuguese versions your locations need in one pass. Canva AI holds 4.4/5 in our database with a free tier and Pro at $18 per month, or from $12 per month billed yearly. Screenshots of the real software beat invented illustrations for procedure training, so capture genuine screens and let Canva frame and annotate them rather than asking AI to draw what the interface should look like.

This step produces the video layer, and for procedural training it is the layer employees actually finish. Start in Synthesia, the platform rated 4.3/5 in our database for avatar-led corporate video in more than 140 languages. Choose an avatar, paste the module script, and structure the script with this prompt first in ChatGPT:

Convert my Module 1 text into a video script for a training presenter. Target 4 to 5 minutes of spoken delivery, about 600 to 700 words. Structure it as: a 20 second hook naming the mistake this module prevents, the core steps in spoken language with no slide jargon, one walk-through of the real scenario, and a closing that bridges to the next module. Write for the ear: short sentences, no parentheses, and spell out anything an avatar must say naturally. Use only facts from my module text.

Paste the finished script into Synthesia, and it renders the presenter video in minutes: no studio, no camera, no retakes. When the process changes next quarter, edit the script line and re-render the same video instead of booking a reshoot, which is the specific economics that makes AI video the default for corporate training. Synthesia starts at $29 per month on Starter with Enterprise custom, and the free tier lets you test the workflow before committing.

For the human touch moments, HeyGen at 4.4/5 creates avatar video from a single photo of a real team member with consent, with a free trial and Creator at $29 per month, which is how companies put an authentic welcome from the team lead on top of avatar-delivered instruction. Close the step in Veed.io, rated 4.4/5 with a free tier and Lite at $12 per month billed yearly: upload the screen recordings of your real software, cut them into the avatar video, and let the AI subtitles generate captions in one pass, because captions are the single accessibility feature training videos cannot ship without. The pattern that works is hybrid: one authentic human introduction, avatar-delivered instructional body, real screen recordings for the procedure itself.

This step closes the loop: the checks that prove learning happened and the metrics that prove the training mattered. First, generate the knowledge check in ChatGPT with this prompt:

Create a 10 question quiz for Module 1 from the text in this Project. Use 7 multiple choice questions with 4 options each and 3 scenario questions where the learner picks the correct action. Draw the wrong options from the common mistakes my sources document, not from invented errors, so each distractor tests a real misconception. Provide the answer key with one sentence explaining why each answer is correct, quoting the module where the fact appears.

Then load the questions into Quizlet, rated 4.4/5 in our database, which turns them into flashcard decks and spaced repetition sessions employees review on their phones between shifts, free to start with Quizlet Plus at $7.99 per month or $35.99 per year. For skill checklists and scenario scoring, build the observation form in Notion AI so a supervisor signs off on demonstrated behavior, because knowledge quizzes measure recall while the business usually needs proof of practice.

Finally, measure the program the way you would measure any process change. Export completion rates, quiz scores, and the onboarding time of the cohorts before and after, then upload the spreadsheet to Julius AI, rated 4.2/5 with a free tier and Plus at $20 per month, and ask it to compare cohort performance, chart quiz scores by module to find the module where understanding drops, and flag any question more than 60 percent of learners miss, because a failed question usually marks a teaching gap rather than a learning gap. Industry ROI analysis reports AI-supported training cutting onboarding time by up to 50 percent, and the spreadsheet is where you prove that number for your own program instead of quoting it.

The six steps stay the same for any program, but the emphasis shifts with the training type, and knowing where your learners will struggle lets you spend production time in the right places. Here is how to tune the workflow for the four most common programs.

• Onboarding: split the program into a role-general track, company values, tools, and policies, and a role-specific track built from the Step 1 interview with the team lead. Record one authentic welcome video from a real leader in HeyGen, and let the avatar body carry the procedural content. The 50 percent onboarding time reduction in the ROI studies comes mostly from putting the repetitive explainer content on demand instead of repeating it live for every hire.
• Compliance training: grounding becomes legal protection, so every module must trace to the policy document and every draft routes through compliance review before Step 4 begins. Use Synthesia heavily here, because policy explainers are exactly the content employees rewatch at their desk, and log the source versions in Notion AI so auditors can see which policy edition each module was trained on.
• Software and systems training: screen recordings replace most generated visuals, so Step 4 shrinks and the Veed.io part of Step 5 grows. Capture the real interface with a veteran clicking through, let the AI subtitles and scene detection do the editing, and regenerate the walkthroughs after every release instead of letting stale screenshots teach the wrong interface.
• Soft skills training: the roleplay scenario is the asset, so Step 3 writes scenario branches and the facilitator guide becomes the main document, with the deck in Gamma reduced to prompts and debrief questions. AI video works for modeling a bad example, watching the avatar mishandle a difficult conversation is genuinely instructive, but the practice itself stays human.

Whatever the type, keep the order of the steps intact: sources feed the curriculum, the curriculum feeds the writing, and the writing feeds every asset downstream. The sequence is what lets one person produce what used to take a five person team.

The difference between training people skim and training they use comes down to workflow habits. These seven tips come from the patterns that separate programs that change behavior from programs that merely get completed.

• Give each tool one job. Perplexity researches, tl;dv captures, ChatGPT structures, Claude writes long-form, Gamma designs, Synthesia films. Mixing jobs produces mediocre output at both, while a single-purpose chain lets each tool do what its rating in our database was earned for.
• Build a terminology glossary and paste it into every prompt. Your company calls it the escalation path, not the support flow, and the AI will happily use both unless told otherwise. Thirty lines of approved terms at the top of each prompt keeps all eight tools speaking the same language as your veterans.
• Write the GAP flag into your prompts. Instruct every tool to mark anything your sources do not cover with a GAP flag instead of filling the hole with plausible invention. The flagged list becomes your next expert interview agenda, which is how the workflow improves itself.
• Review modules before producing assets. Every downstream artifact, the slides, the video, the quiz, inherits the module text. A factual fix costs one prompt in Step 3 and a full afternoon if it surfaces in Step 5, so the expert review gate belongs exactly where this guide puts it.
• Keep a prompt library in the training hub. Store every prompt that produced good output in Notion AI with a note on what it produced. The next program starts from proven prompts instead of the blank page, and the module updates next quarter take minutes rather than days.
• Version everything with dates. Name files with the policy edition and the render date, because the question every auditor and every new manager asks is which version of the truth this material teaches. The source hub in Notion AI makes the answer a link instead of an archaeology project.
• Test on one real learner before launch. Hand Module 1 to one actual new hire and watch where they slow down, reread, or skip. Thirty minutes of observation beats any amount of internal proofreading, and the fix is usually one prompt away.

Most AI-produced training fails in predictable ways, and every failure mode has a simple fix. Four mistakes account for the large majority of weak programs.

• Generating without grounding. Feeding a bare prompt like write me a safety training course produces content that sounds authoritative and describes a company you do not work for. The fix is the entire Step 1: sources, interviews, and prompts that instruct the tool to use supplied facts only, with GAP flags for the holes.
• Skipping the expert review gate. AI drafts that ship unreviewed eventually contradict the actual process, and one employee following invented instructions costs more than every hour the workflow saved. The SME review in Step 3 is not bureaucracy; it is the step that makes the speed safe.
• Wall-of-text modules. AI happily produces 2,000 word blocks that no shift worker will read. The fix is structural: instruct the writer to break content into sections with the mistakes box and the worked example, and let Gamma and Canva AI carry the visual load, because the reading core should be under 1,200 words per module with everything else chunked.
• Measuring completion instead of capability. A 100 percent completion rate with quiz scores of 60 percent means employees clicked through, and the business learned nothing. Track the score-by-module view in Julius AI, watch for the module where understanding drops, and fix the teaching, not the completion reminder.

None of these mistakes come from the tools; all of them come from skipping the checkpoints this workflow builds in. Teams that keep the sequence ship training that passes audit and changes behavior, in a fraction of the old timeline.

Training materials are a recurring asset, not a one-time deliverable, and the maintenance burden is where traditional programs quietly die. Processes change, policies get revised, and software ships new interfaces, so a manual course starts aging the day it launches. The AI workflow changes the maintenance equation completely, because the real source of truth lives in your Notion AI hub and every asset regenerates from it: update the source document, rerun the affected prompt, and the module text, the deck in Gamma, and the video script in Synthesia all come back current without a reshoot or a rewrite from scratch.

Run the maintenance on three cadences. Monthly, export the quiz data into Julius AI and read the score by module view, because a module where scores sag is a module that needs a rewrite before anyone complains. Quarterly, rerun the Perplexity research prompt from Step 1 and compare the answers against your current material, since industry standards drift faster than internal processes do. Immediately, when a process or policy changes, edit the source document, regenerate only the modules the change touches, and re-render the affected video segments the same week, which is the exact edit-the-script-and-rerender loop that makes AI video economical for training in the first place.

Discipline is what holds the system together as the library grows. Name every file with its source version and render date, retire superseded versions the day replacements ship, and keep the retirement log in the same hub, because the audit question is never whether training exists but which version of the truth it taught and when. Under this cadence, maintenance drops from a rewrite project measured in weeks to a regeneration routine measured in minutes, and the materials stay as current as the source documents they were built from.

The table below summarizes the full training production stack so you can see the whole workflow at a glance. Every tool has a free tier or trial, and the complete paid stack would cost roughly $150 per month at list prices, though most teams pay for only two or three subscriptions at a time as the program grows.

Free tiers carry the entire workflow in this guide, and every upgrade has a specific moment when it becomes worth buying: the moment its constraint actually interrupts your program. Buying in that order, instead of subscribing to everything up front, keeps your costs matched to your progress.

• Synthesia Starter at $29 per month is the first upgrade most teams make, when the free video minutes run out and the module videos need re-rendering after process changes. Video is the layer employees consume most, so this constraint bites first.
• ChatGPT Go at $8 per month removes the usage caps free plans apply at peak hours. Upgrade when curriculum iteration sessions start getting interrupted, typically once you are producing more than one program at a time.
• Claude Pro at $20 per month, or $17 billed yearly multiplies your usable message volume by roughly five and keeps long module-writing sessions from cutting off. Upgrade when the program has more than three modules or you maintain two tracks at once.
• Gamma Plus at $9 per month refills credits faster and removes the made-with-Gamma branding from shared decks, which matters the moment the deck goes in front of leadership or external auditors.
• tl;dv Pro at $29 per user per month unlocks unlimited AI notes and deeper search across the interview archive, worth it once you are running structured discovery for several role types.
• Quizlet Plus at $7.99 per month or $35.99 per year adds the deeper spaced repetition scheduling for programs that run over months rather than days, usually the last upgrade in the stack.

For context, subscribing to every tool in this guide costs roughly $150 per month at list prices, and that is rarely necessary. The practical pattern is one or two subscriptions at a time, bought against the bottleneck you are actually feeling, and revisited after the first cohort completes, because the shape of your second program will not match the first.

Reading about a workflow does not build a course, so here is the whole guide compressed into a checklist you can start today. Each block maps to a step, with the realistic time it takes on free tiers across a one week part-time schedule.

• Session 1, about 2 hours: collect every existing document into one Notion AI source hub, run the Perplexity prompt from Step 1, and book the expert interview.
• Session 2, about 1 hour: record the interview with tl;dv, export the transcript, and file it with the research.
• Session 3, about 2 hours: run the ChatGPT curriculum prompt from Step 2, push back on the generic parts, and lock the one page curriculum map with GAP flags assigned.
• Session 4, about 4 hours: write the modules with the Claude prompt from Step 3, route each through the expert, and file the approved versions.
• Session 5, about 2 hours: generate the deck in Gamma, the job aids in Canva AI, and the quiz in ChatGPT, then load the questions into Quizlet from Steps 4 and 6.
• Session 6, about 3 hours: script and render the video in Synthesia, cut in the screen recordings with Veed.io from Step 5, and test the full module on one real learner before launch.

Six sessions produce a program that grounds every claim in your own sources, ships in every format employees use, and leaves you with the metrics to prove it worked. Start with the source folder this morning, and by the end of the week the training backlog stops being a queue and becomes a process.