Model-Agnostic AI Training: Skills That Outlast Tools

Model-agnostic AI training teaches the skills that work in any AI assistant (assigning work, setting standing context, connecting sources, routing work by its shape, verifying output, following data rules and keeping work portable) and uses the tools your organization already licenses as the practice ground. It matters because vendors change their products faster than click-path courses can be rewritten: between 2025 and October 2026, OpenAI, Microsoft, Google and Anthropic retired models, renamed apps and plans, and scheduled replacements for the features teams saved their instructions in. Administrators and builders still need vendor-specific training; the model-agnostic approach keeps that layer thin and gives it only to them.

Key figures (updated 23 October 2026)

• OpenAI launched GPT-5 on August 7, 2025 and retired it from ChatGPT on February 13, 2026, along with GPT-4o, GPT-4.1, GPT-4.1 mini and o4-mini: about six months.
• Microsoft added Anthropic's Claude models to Microsoft 365 Copilot on September 24, 2025; they are now on by default for most commercial-cloud customers outside the EU, EFTA and UK.
• OpenAI plans to retire custom GPTs in favor of plugins (announced September 11, 2026), and Google says skills will soon replace Gems (September 30, 2026).
• In BCG's 2025 survey of more than 10,600 leaders, managers and frontline employees, only 36% said they felt properly trained.
• Australia's 2024 whole-of-government Microsoft 365 Copilot trial found staff needed both training tailored to agency use cases and general generative AI training.
• The AIDOLS AI Fluency course is free to read, vendor-neutral by design, and takes 228 minutes for one full path.

Cite as: AIDOLS Research Team, Model-Agnostic AI Training: Skills That Outlast Tools, aidolsgroup.com, 23 October 2026.

This guide covers the definition, dated evidence of how fast tool training goes stale, the skills that transfer, the vendor layer, a neutrality checklist, a single-tool plan and a one-page template.

Model-agnostic AI training, also called vendor-neutral or tool-agnostic AI training, teaches durable working skills first and translates them into each licensed tool second. The test is whether it would survive a change of assistant: briefing a task, checking an output and deciding what data may go in would; menu paths and model names would not.

Public frameworks draw the same line. The European Commission's minimum content for AI literacy under Article 4 of the EU AI Act names no product: a general understanding of AI, the organization's role as provider or deployer, the risk of the systems involved, and literacy actions built on that analysis (European Commission).

Principle, worked example, translation

Each lesson in a model-agnostic course has three layers: a principle true in any assistant, a worked example in one named tool, and a translation into the others. The AIDOLS AI Fluency course works this way: lessons label the worked example ("Shown here in ChatGPT") and include an "In your tool" table for Claude, ChatGPT, Gemini and Microsoft 365 Copilot, dated when the surfaces were last checked (August 4, 2026 for the current lessons). Those rows go stale: the ones that point to ChatGPT's canvas or to Gemini's Gems are already overtaken by the changes below. The principles above them are not.

What it is not

• Training without tools. Every exercise runs in an assistant the learner can open, on a task they already do.
• A vendor ranking. Claims that one tool is better at something age within a release cycle; in the course's words, a tool map that ranks vendors "stops being a map and becomes an argument."

Because what it teaches (model names, app and plan names, and the features people save work in) changed repeatedly between 2025 and October 2026 at every major vendor. Each change below broke something a click-path course would have taught; none changed how to brief a task, check an output or decide what data may go in.

Models, and the features built on them, retire within months

GPT-4.5, a research preview released on February 27, 2025 (OpenAI), was retired from ChatGPT on June 26, 2026, and OpenAI o3, introduced on April 16, 2025, was scheduled to leave on August 26, 2026; OpenAI says ChatGPT models generally stay available for 90 days after a successor is released (OpenAI).

Saved work moves with them. When OpenAI retired models in 2026, custom GPTs built on them moved automatically to the closest newer model (OpenAI): same name and instructions, different model underneath. Now the containers are going too: custom GPTs are slated to give way to plugins, which can combine reusable instructions with connected apps (OpenAI), and Gems to skills (Google). Training that taught "build a GPT" taught a container; training that taught people to write the instruction down, prove it on fresh inputs and keep a copy taught the asset.

The vendors are now inside each other's apps

"Copilot or ChatGPT training" is a false choice, because the products no longer separate cleanly. In enabled tenants, Microsoft 365 Copilot users can select Claude in Edit with Copilot and in Researcher (Microsoft Learn), and by default a real-time router picks the underlying model for each prompt (Microsoft Learn). Since September 17, 2026, ChatGPT also works inside Word, through the same Microsoft add-in as Excel and PowerPoint (OpenAI). When vendors share the document, routing by vendor fails; routing by the shape of the work still holds.

Governance knowledge dates too. At launch, Microsoft said Anthropic models were hosted outside Microsoft-managed environments under Anthropic's terms (Microsoft). Its documentation now says Anthropic has onboarded as a Microsoft subprocessor, so Microsoft's Product Terms and Data Protection Addendum apply, except for models labeled "Anthropic models with Data Retention", which are off by default and fall under Anthropic's own terms (Microsoft Learn). Australia's evaluators asked for adaptive planning to reflect "the rolling release cycle nature of generative AI tools" (Digital Transformation Agency).

Nine skills transfer intact because they describe the work, not the product. The table maps each to the AI Fluency lesson that teaches it; only the last column, which is structural, needs editing when a vendor changes.

Verification and data rules carry the most weight

Verification is the skill no license includes. Australia's evaluation found that time spent verifying and editing Copilot output negated some of the efficiency savings, and that 61% of managers in its pulse survey could not confidently identify Copilot outputs (Digital Transformation Agency). The European Commission says even staff using ChatGPT for advertising text or translation should be informed of specific risks such as hallucination (European Commission).

The course's answer is a fixed four-pass check (trace the numbers, check what points outward, find the load-bearing sentence, look for what is missing) sized to the cost of being wrong, plus four facts that decide paste, redact or refuse: the class of the data, the account it goes to, who can retrieve it later and how long it is kept. Neither depends on the assistant.

Configuration does. Transferable skills will not switch on a tenant's AI-provider setting, set sensitivity labels, grant model access or configure retention; the people who own those settings need training on the exact product they administer, and everyone else needs only a thin product layer.

Tenant settings decide what each user sees

Two people in one company can see different tools. Microsoft turns Anthropic models on by default for most commercial-cloud customers but off in the EU Data Boundary and the UK; changing that requires the AI Administrator or Global Administrator role, and access can be limited to named users or Microsoft Entra security groups (Microsoft Learn). Microsoft's license comparison also separates three kinds of user: Copilot Chat (Basic), with no Copilot in Word, Excel, PowerPoint or OneNote; Microsoft 365 Copilot (Basic), with no add-on but standard access in those apps; and Microsoft 365 Copilot (Premium), with the add-on, priority access and grounding in work data through Microsoft Graph (Microsoft Learn).

Whether Copilot can display text from an encrypted file depends on the EXTRACT usage right in the Microsoft Purview sensitivity label, and the audit log does not capture the actual prompt or response; for that, Microsoft points administrators to eDiscovery (Microsoft Learn). Click-path training breaks in mixed tenants because the clicks differ by person. The skills do not.

Vendors already publish the click-level material

Microsoft publishes separate user training for each Copilot license tier, a Copilot Success Kit covering admin controls, licensing, training materials and onboarding email templates, and a Copilot Trainer Kit, a downloadable PowerPoint presentation for Copilot Chat (Microsoft Learn). If a provider's paid hours mostly reproduce that material, you are paying for what the vendor already publishes. Spend them on judgment and your own work.

The available surveys point the same way: several models and apps per organization, and employees going beyond the licensed tools. Each measures a different population.

• Models in production. In a16z's 2025 survey of 100 enterprise CIOs across 15 industries, 37% were using five or more AI models, up from 29% a year earlier (a16z). That counts models in production use cases, not employee assistants.
• AI apps across the enterprise. In a Zapier survey of more than 500 enterprise leaders published in December 2025, 28% said their enterprise uses more than 10 different AI applications, 34% said tool sprawl makes training employees on AI a major challenge, and only 35% said their AI tools go through proper approval channels (Zapier). Zapier sells automation software, so treat this as a vendor survey.
• Employees filling gaps. More than half of employees in BCG's survey said that when they lack the AI tools they need, they will find alternatives and use them anyway (BCG).
• Even in a single-vendor trial. Australia's Copilot evaluation noted that other generative AI tools may be more effective for reviewing or writing code, generating images or searching research databases (Digital Transformation Agency).

Why a single-tool policy does not create a single-tool workforce

A license decision settles what the organization pays for, not what people use. Training that only works inside the licensed product leaves the rest untrained, and that is where data rules matter most: the person pasting a client file into a personal account needs the paste, redact or refuse habit more than anyone. For licensing and data terms, see how Microsoft 365 Copilot, ChatGPT Enterprise and Claude compare.

For a workforce, AI vendor lock-in is mostly instructions, configuration and habits that exist only inside one account. Avoid it with plain-text copies the organization owns, re-tests after vendor changes and open standards where they exist.

What each vendor lets you take with you

• ChatGPT. Self-service data export is not available in ChatGPT Business workspaces, and Business-to-Business workspace merges are not supported, so GPTs cannot move that way (OpenAI).
• Claude. On Team and Enterprise plans, members can export their own chats only after an owner turns on "Allow members to export their own data", and the export leaves out the contents of uploaded files; Primary Owners export the whole organization's data (Anthropic).
• Gemini. Google says Gems will migrate into skills automatically when Gems go away (Google); copy the instructions out first anyway.
• Microsoft 365 Copilot. Retrieving what employees typed is an administrator's eDiscovery task in Microsoft Purview (Microsoft Learn).

Keep the instruction, the setup and the outputs as plain text

The course's export lesson keeps each build as plain text in a folder the organization owns: the instruction, the setup written as prose, the sources, the checks and a dated file for each run. Its rule for setup notes, "name what I set rather than where I clicked", lets the record survive the next interface change. The a16z report noted that "prompts can be more easily ported from one model to another", while "the rise of agentic workflows has started making it more difficult to switch between models" (a16z). Re-test on a trigger: Anthropic commits to at least 60 days' notice before retiring a publicly released model (Anthropic), while OpenAI gave GPT-4.5 a 30-day sunset in ChatGPT (OpenAI).

Prefer open standards where they exist

Connections and packaged instructions are where lock-in grows as teams move from chat to agents. In December 2025 Anthropic donated the Model Context Protocol (MCP), the open standard for connecting AI assistants to data and tools, to the Agentic AI Foundation, which Anthropic, Block and OpenAI co-founded under the Linux Foundation; Anthropic reports that ChatGPT, Gemini, Microsoft Copilot and Visual Studio Code have adopted it (Anthropic). When a provider builds something for you, ask which parts use an open format, and keep a plain-text copy of the rest.

Model-agnostic training is the better default for the workforce; tool-specific training is the better choice for administrators, builders and narrow feature rollouts. The judgments in this table are AIDOLS assessments.

AIDOLS recommends a blend: a model-agnostic core for everyone, a thin vendor layer by role, and a refresh on a trigger rather than the calendar alone. Vary the formats: in Australia's trial, 75% of participants who received three or more forms of training were confident using Copilot, 28 percentage points more than those who received one (Digital Transformation Agency).

Ask who pays them, which tools their examples use, and whether you keep the work. Partner programs are legitimate and common; the issue is disclosure. Microsoft's partner site says partner incentives provide direct funding to eligible partners to deliver standardized workshops to qualified customers, and its engagement catalog, funded and unfunded, includes Copilot and Agents at Work, covering Microsoft 365 Copilot, Copilot Chat and Copilot Studio (Microsoft). Anthropic launched the Claude Partner Network in March 2026 with an initial $100 million for training courses, dedicated technical support and joint market development, and says a significant share goes directly to partners (Anthropic).

Questions to send before you sign

1. Do you resell licenses for any AI product we use, or earn referral fees, rebates or partner incentives tied to our purchases or usage?
2. Is any part of this engagement funded by a vendor program, and what does that program require you to cover?
3. Which tools do your worked examples use, and do you rotate them?
4. Will every exercise run in tools we already license, on tasks our people actually do?
5. Where does your training stop and the vendor's admin training start?
6. Will we keep plain-text copies of every instruction, template and tool map you build with us?
7. Does any part of your curriculum name a "best" vendor?
8. Do you measure attendance and quiz scores, or a countable change in the work?
9. What would lead you to recommend a tool you have no partnership with?
10. How do you keep the material current when vendors rename or retire things, and who pays for the update?

Red flags in a proposal

• A course named after one product, sold as "AI training".
• A tool recommendation before anyone has looked at your work.
• Success measured in activated seats or logins.
• No written handover of the instructions built during the program.

AIDOLS does not resell platform vendor licenses (ServiceNow, Microsoft, Salesforce, SAP, Oracle). Send us the same ten questions.

Teach the same skills inside Copilot, label every example "shown here in Microsoft 365 Copilot", and keep the portable record outside it. Copilot Chat (Basic) is included with eligible Microsoft 365 licenses, so organizations on those plans already have an assistant to practice on (Microsoft Learn). The steps suit any single-tool estate; for the rollout itself, see our 90-day Microsoft 365 Copilot adoption plan.

1. Inventory what is licensed and switched on. Record each person's tier, Copilot Chat (Basic), Microsoft 365 Copilot (Basic) or Microsoft 365 Copilot (Premium), because in-app access and grounding in work data differ by tier. Check whether Anthropic models are enabled for your region and groups, and which agents and connectors your administrators allow.
2. Name one recurring task per person and record a baseline. Pick a task each person repeats on a schedule, can describe end to end and could afford to get wrong once, and write down how long it takes today.
3. Teach each skill in Copilot, principle first and clicks second. A reusable instruction can live in an agent, and administrators control how users create and use agents (Microsoft Learn). Where the tenant allows it, practice model choice too: users can select Claude in Edit with Copilot and Researcher.
4. Write the team tool map by work shape. Name surfaces, not vendors: Copilot in Word, Excel or PowerPoint for work inside one file, where the tier includes it; Copilot Chat for work across files and mail, which draws on Microsoft Graph automatically only with the Premium license, while Basic users attach the files themselves (Microsoft Learn). Rank nothing.
5. Set the one-page data ground rule and the verification tiers. Decide which classes of data may go into which account, who owns the rule, and which outputs get which depth of check.
6. Export every working instruction as plain text. Keep the folder somewhere the organization owns, and re-test each instruction on fresh inputs whenever a vendor announces a retirement, rename or new default model.
7. Measure a countable change in the work. Count the share of built assets reused at two weeks, the named people outside the first cohort using them at 30 days, and one change in cycle time, volume or a removed step at 90 days, as the course's adoption lesson sets out. These checkpoints are course guidance, not research findings.

When the job is configuring or building on one product, or a rule requires training on one specific system.

• Tenant administrators. People who set Purview labels, run eDiscovery, manage Copilot settings or hold the Gemini Settings administrator privilege need training on that product.
• Builders on one platform. Someone building Copilot Studio agents or ChatGPT plugins needs that platform's documentation.
• A single feature rollout where the click path is the job. The vendor's own kits may already cover it.
• High-risk AI systems in the EU. The European Commission says deployers of high-risk AI systems remain obliged to ensure their staff are trained for human oversight (European Commission); general literacy training does not replace training on that system.
• Teams that have never opened the tool. A short vendor orientation first makes the skills sessions faster.
• A deliberate long-term single-vendor commitment. Weight the vendor layer more heavily, but still keep instructions as plain text.

Most organizations need both layers; the harder decision is who gets which: see whether to train AI champions or the whole team. Regulated professions add a layer of their own, as our guide to AI training for law firms shows.

One page, one owner and a review trigger. Copy the template; the quarterly cadences in it are AIDOLS planning estimates, not vendor rules.

VENDOR-NEUTRAL AI TRAINING PLAN
Organization: [name]   Owner: [role]   Next review: [date]

1. TOOLS: licensed [tool, plan, seats]; switched on [models, agents,
   connectors]; unlicensed tools in use [from a staff survey]
2. WORK SHAPES: [team]: [shape] -> [surface], one row per shape
   the team meets
3. CORE SKILLS: course path, 228 minutes; hands-on sessions on each
   person's own task [dates]
4. VENDOR LAYER: administrators [admin training]; builders [platform
   docs]; all staff [where saved instructions live]
5. DATA GROUND RULE: [link]; owner [role]
6. PORTABILITY: instruction folder [location the organization owns];
   copy back after every edit; export check [quarterly]
7. MEASURES: 2 weeks, share of built assets reused; 30 days, named
   people outside the first cohort using them; 90 days, cycle time,
   volume or a removed step
8. TRAINING RECORD: sessions, attendees, assets built (internal)
9. REFRESH TRIGGERS: vendor retirement or rename notice; new admin
   setting; license change; quarterly vendor-layer review

In the EU and EEA, the record matters: the European Commission says AI literacy under Article 4 needs no certificate and that organizations can keep an internal record of training. Since the Digital Omnibus amendments entered into force on July 27, 2026 (Regulation (EU) 2026/1744), Article 4 keeps AI literacy as an obligation without mandating a specific level, and national authorities supervise and enforce it from August 2, 2026 (European Commission). References to the EU AI Act in this guide are general information, not legal advice; check current guidance or ask counsel.

The AI Fluency program is built on a course that is vendor-neutral by design.

• One skill, several tools. Each lesson's worked example is translated into Claude, ChatGPT, Gemini and Microsoft 365 Copilot, and the team tool map never ranks vendors.
• Your tools, your work. Team programs run hands-on sessions by role, built around each team's own tools and tasks.
• Free to read. Every written lesson, including all three role tracks, is free and publicly readable with no sign-up.
• Enrollment adds the doing. Guided builds, starter templates, instructor video, cohort office hours, a graded capstone and a certificate issued only for a build that was submitted and marked, never for time spent, plus roster and completion reporting for organizations. Individual enrollment is not open yet (there is a waitlist); organizations can ask about a cohort or team program.

Reading is a floor. BCG found regular AI use sharply higher among employees with at least five hours of training and access to in-person training and coaching (BCG); the course path is 3.8 hours, so plan hands-on sessions on live work alongside it. AIDOLS is the wrong fit if you need administrator training for one product, which the vendor's documentation serves better, or want a provider to name the best vendor.

• This week, for free: have two or three people read Units 1 and 3 of the course, then write the default line and first three rows of a tool map for your team. If a new joiner can route a real task from it without asking, the approach fits.
• For a program built around the tools you already license: book a 15-minute call and bring your license list and three recurring tasks.
• If the bigger question is where AI should go first: the free workflow map takes one workflow: AI agents interview the two or three people who run it, and within three business days you get a one-page map and a written verdict on what each bottleneck needs (software, a process change or AI), with no obligation.