Build an AI Literacy Program

info

This content was developed with AI assistance and is regularly reviewed for accuracy.

Snapshot: October 2026. Model names, prices, and product features change monthly; What's New in AI is the maintained chronology.

Most organizations still treat AI training as an event: schedule a workshop, send a video series, mark it complete. Six months later the usage data shows a fraction of people genuinely working differently, and everyone else remembers the training the way they remember every other mandatory module, vaguely and not practically.

Effective literacy programs are systems, not events. They start from a measured gap, they are designed differently for different roles, they put learning inside the work, and they measure behavior rather than completion. The pressure to get this right has grown in 2026. Deloitte's State of AI in the Enterprise (January 2026, reported) found skills shortages were the biggest barrier to AI value, and that broad AI literacy (53%) and upskilling (48%) were the top responses, well ahead of hiring specialists (36%). The EU AI Act's Article 4 has required proportionate literacy measures since February 2025. And the skills employers now prize have shifted: in Microsoft's 2026 Work Trend Index, AI users ranked quality control of AI output (50%) and critical thinking (46%) as the most important human skills, and 86% said they treat AI output as a starting point rather than a finished product. Coursera's 2026 Job Skills Report shows demand for critical-thinking content among generative-AI learners up 185% year over year (reported). Prompt mechanics are not the scarce skill. Judgment is.

Learning focus:

• Define AI literacy for 2026 using the 4D fluency framework and map it to concrete skills
• Run a skills-gap assessment that tells you where training will pay off
• Design a program across four modalities with role-specific curricula
• Measure behavior change, and keep the record a regulator would ask for

What AI literacy means in 2026

Three public frameworks dominate how organizations describe AI literacy, and they agree on the shape of it. Anthropic's AI Fluency framework, taught in a free, Creative Commons course created by Margaret Vo with professors Rick Dakan and Joseph Feller (reported), defines four competencies: Delegation (deciding what to hand to AI and what to keep), Description (communicating the task, context, and output you want), Discernment (evaluating what comes back), and Diligence (using AI responsibly, transparently, and safely). UNESCO's AI Competency Framework for Students organizes 12 competencies across four dimensions and three levels, Understand, Apply, Create, which makes a good proficiency ladder for adults too. And the big vendor academies all treat prompting as one module among four to six, after "how it works" and before responsible use.

This site's Fluency track teaches the same competencies as a workflow. The mapping below is what you will build your matrix on.

4D competency	What it means at work	Where this site teaches it
Delegation	Decide which tasks to hand off, which tool and mode to use, and when a plain workflow beats an agent	Decide
Description	Write a brief: objective, context, resources, constraints, output, verification; supply files, projects, memory, connectors, skills	Brief, Supply Context
Discernment	Check output, verify citations and actions, iterate, and know when to restart	Check
Diligence	Protect data, scope permissions, resist prompt injection, avoid irreversible actions, follow policy	Protect

Article 4 fits this picture rather than fighting it. Compliance commentators agree it prescribes no course, exam, or certificate; measures should reflect people's technical knowledge, experience, education and training, the context the systems are used in, and the people affected by them (reported; confirm with counsel). A role-based program with records does that. A one-size video does not.

Start with a skills-gap assessment

Designing training before you know where the gaps are is guesswork. An assessment takes an hour or two to design and run, and it makes everything after it cheaper. It should tell you two things: where people actually are, not whether they attended something, and which gaps cost the most.

Four approaches, each with a trade-off. Self-reported surveys are quick but overestimate capability; use them for relative confidence and to find pockets of high and low engagement. Observed task completion (ask someone to brief an assistant on a realistic task, supply the right file, and judge the output) is more accurate and more expensive. Manager input often gives the clearest picture of where a team struggles and how work actually flows. Usage data, where you have it, shows breadth: people using one feature of a tool that offers ten usually have a knowledge gap, not a motivation problem.

Then prioritize by impact. A gap in writing a good brief on the sales team might cost 30 minutes per person per day; the same gap on the legal team might cost 20 minutes a week but carry sanction risk. Train where the payoff, in time or in avoided harm, is highest.

"I'm designing an AI skills-gap assessment for [role, e.g. customer service
representatives] who mainly use [tool] for [use cases]. Create a 10-question survey
covering current comfort level, the tasks where they feel least confident, whether
they know how to supply context (files, connectors) and verify output, and what
stops them using AI more. Mix Likert-scale and open-ended questions."

The skills matrix

AI literacy is a cluster of capabilities people need in different combinations. A matrix lets you build modular training instead of one program for everyone. The 2026 version below drops the 2023 assumptions (that each chat starts fresh, that the model cannot browse, that a persona improves accuracy) and adds the skills that delegated work requires.

Foundational (everyone, regardless of role)

• How assistants work in 2026 and where they fail: long context, always-on reasoning with an effort setting, web search, memory, and task-dependent hallucination
• Delegation: what to hand off, what to keep, and which mode (fast, thinking, research, agent) fits
• Description: a six-field brief and attaching the right file instead of pasting it
• Discernment: checking facts and citations, and the rule that after two failed corrections you restart with a better brief
• Diligence: the organization's policy, data classification, memory settings, and what never goes into a prompt

Most of a workforce needs this layer plus the role-specific intermediate layer, not advanced training.

Intermediate (people who use AI daily in their work)

• Supplying context deliberately: projects, custom instructions, approved connectors, and reusable skills
• The five to ten highest-value tasks for this role, each with a tested brief
• Judging "good enough" versus "needs careful review," and recognizing fabricated detail
• Reviewing what an agent did, not just what it said, before accepting the result
• Recognizing a prompt-injection attempt in an email, web page, or document the assistant reads

Advanced (power users, champions, and managers who oversee AI-assisted work)

• Designing a team workflow or a scheduled task with approval gates, scoped permissions, and spend limits
• Evaluating tools and vendors with the organization's own test set (Evaluate and Measure)
• Building and maintaining the team's brief and skill library
• Coaching colleagues and running error reviews
• Keeping up with capability changes and judging which ones matter

Map your population to these levels and you have a curriculum outline. Most of the time you will find the foundational layer is thinner than people assume and the intermediate layer is where the money is.

Four training modalities

Effective programs mix delivery formats matched to the kind of learning each produces.

Self-paced modules

Best for foundational concepts, reference material, and a distributed workforce. The failure mode is passive content: watch this, take a quiz. Make modules active by ending each one with a real task in the learner's own tool: "write a brief for your most common weekly task, attach the file it needs, and review the output against the checklist." A module that ends with a produced artifact changes behavior; a module that ends with multiple choice changes a completion rate.

Cohort-based learning

Best for building shared language and norms, role-specific application, and peer connection. A cohort of eight to fifteen people from the same function works through two to six weeks together. Keep cohorts role-homogeneous so the use cases apply immediately, build in structured sharing ("what did you try since last session, and what happened?"), and end with a team artifact: a shared brief library, a documented workflow, or a short showcase.

Embedded-in-workflow learning

Best for driving actual behavior change. The biggest barrier is not knowledge but integration: people finish training and return to work that has no natural entry point for the new skill. Embedded learning builds AI into the work itself. Add an "AI-assisted draft" step to the weekly report process. Make the shared brief library the default starting point for recurring writing. Set a team norm that any task over 30 minutes gets an AI attempt first, with the result shared. Give the team a shared project with custom instructions so nobody has to reconstruct context. None of this happens without the manager who assigns and evaluates the work setting the expectation, which is why Drive AI Adoption spends so long on managers.

Peer learning

Best for ongoing development and for capturing knowledge no central curriculum would include. Peer learning happens anyway; the question is whether you design for it. Useful formats include champion-led 30-minute demos, a monthly 20-minute brief review where the team shares what worked and what did not, low-stakes "I tried this" posts in a team channel, and paired review of each other's AI-assisted work, which builds quality judgment faster than solo practice.

Role-specific curriculum examples

Generic training covers the tools. Role-specific training covers how this job uses them. Three non-technical examples, written as behaviors with a practice task for each.

Customer service

Focus	Learning objective	Practice task
Response drafting	Draft a reply that matches company tone from a brief that includes the policy document	Take a real complaint, attach the relevant policy, produce and edit a draft
Ticket triage	Summarize a thread and recommend resolve or escalate	Summarize a 10-message thread; identify the core issue and next step
Knowledge base	Turn last month's top questions into FAQ entries	Extract five recurring questions from a ticket export and draft answers

Finance and accounting

Focus	Learning objective	Practice task
Report summarization	Produce an executive summary and check every figure against the source	Summarize a P&L; reconcile each number back to the document
Variance narrative	Draft plain-language explanations of budget variances	Explain three variances for a board deck, flagging anything unverifiable
Data cleanup	Describe a cleanup task and evaluate the proposed approach before running it	Brief a cleanup of a messy export; review the plan; run only after approval

Operations and administration

Focus	Learning objective	Practice task
SOP drafting	Turn a bulleted process into a structured SOP	Draft from bullets; verify each step with the process owner
Meeting workflow	Generate agendas and pre-reads from the calendar and prior notes	Connect the calendar; produce an agenda for next week's standing meeting
Recurring reports	Set up a scheduled task that drafts the weekly status report	Build the task with an approval step before anything is sent

The Brief and Prompt Library and the career guides hold tested briefs for these and other roles.

Integrate into existing L&D, and use the free academies

The fastest way to kill a training program is to position it as something extra. Wherever possible, fold AI development into what already exists. Include AI skills in growth conversations during review cycles, as a development area rather than a score. Add an AI orientation to onboarding and AI tasks to each functional onboarding track, because new hires who learn the norms on day one never have to unlearn anything. Host modules in the LMS people already log into. And add AI coaching to manager development, since managers multiply or neutralize everything else you do.

The vendor academies are now free and credible enough to form the foundational layer, so you do not have to build it. Reported as of October 2026: Anthropic's academy (launched March 2026 and reported relaunched as Claude Academy in August 2026) offers the AI Fluency course plus Claude-specific tracks; OpenAI Academy added self-paced courses with completion certificates in June 2026 and role-based paths in September 2026, including one titled "Lead AI Adoption"; Google AI Essentials remains a five-course, four-to-ten-hour sequence. Details and dates are secondary and change often; Learning & Education and the Certifications Directory are the maintained lists. Use an academy for "how it works" and the four competencies, then spend your own effort on the role-specific and embedded layers no vendor can write for you.

Measure what changed

Completion rates are the least useful metric for a literacy program. A framework adapted from Kirkpatrick's four levels gives you something to report.

Level 1, reaction. Did people find it relevant? Survey: "rate how applicable today's session is to your actual work, 1 to 5." Target 4.0 or above; below 3.5 signals a relevance problem in the curriculum.

Level 2, learning. Did they gain the skill? A task completed at the end of the module (a brief written, a planted error caught) and a before-and-after self-assessment on specific skills.

Level 3, behavior. Are they working differently? Usage breadth at 30, 60, and 90 days, manager observations, and a pulse question ("I now regularly use AI for X").

Level 4, results. Did the business change? Self-reported time saved, quality improvements judged by manager or peer review, and use-case breadth over time. Most programs can measure levels 1 to 3 well; level 4 needs a baseline and patience, and it is the subject of Evaluate and Measure.

Keep the record

Whether or not Article 4 applies to you, keep a simple register: who was trained, on what, when, at which level of the matrix, mapped to their role and to the risk tier of the systems they use (the inventory in Govern supplies the tiers). Add the policy acknowledgment and the date of the last refresher. A spreadsheet is fine. What matters is that when someone asks "what measures did you take?", you can answer with evidence rather than an attendance list.

Hands-on exercise

Design one 90-minute module for one role you know well.

1. Three learning objectives, written as behaviors. "Will draft customer responses from a brief that includes the policy file," not "will understand AI capabilities."
2. The gap it closes. One sentence from your assessment, with the number that justifies it.
3. Two use cases. For each: the task, the brief, the context it needs (file, project, connector), and what good output looks like for this role.
4. The practice activity. What every participant produces during the session.
5. The 30-day behavior metric. How you will know the module changed anything, and what level of the matrix it moved people to.
6. The register entry. What you will record for each participant.

Key takeaways

• The scarce skill in 2026 is judgment, not prompt mechanics: quality control and critical thinking lead every 2026 skills survey, and the public frameworks (4D fluency, UNESCO) agree
• Start with a measured gap and prioritize by cost, in time and in avoided harm
• Build a three-level matrix on Delegation, Description, Discernment, and Diligence; most people need the foundational layer plus role-specific intermediate skills
• Mix modalities: self-paced for reach, cohort for norms, embedded for behavior change, peer for durability; the manager makes embedded learning happen or not
• Use the free vendor academies for the foundation and spend your effort on role-specific and embedded layers
• Measure behavior and results, and keep a register that maps training to roles and risk

Next steps

Continue to: Govern: Frameworks, Policy, and Compliance

• Want ready-made exercises for the foundational layer? Use Hands-On AI Exercises
• Building the team's shared library? Follow Build Your Brief and Skill Library
• Training people who will supervise agents? Pair this page with Agent Safety and Guardrails