DevLearn 2026 Recap — Joan (Calandria) Nelson

A packed ballroom faces the wide opening keynote screen reading Welcome to DevLearn.

If you read one thing

Three key takeaways

AI is not replacing L&D. It is changing it.

AI is our lever, to help us do our job, not replace our jobs.
Brandy M. Ford, PMP, CPTM, CRPKeynote · Learning Design in the Age of AI
Learning teams are not content engines.

“We have become a content engine.” The panel’s answer: it’s not content, it’s designing instruction.
Tessa Forshaw, PhDKeynote · Learning Design in the Age of AI
AI can accelerate the work, great or sloppy.

If you don’t care about your job, AI is going to make you not care about it faster.
Dakota Duffy, M.A., PMPSession 405 · Clone Your Best Evaluator

It’s the human touch that makes learning stick.

The idea to take home

AI gets us to version one. We decide what version is right.

From Penske's session, Less Grind, Better Design. Their team uses AI for first drafts and spends the time it saves on design.

Don't automate instructional design. Automate the friction around it.

Same session. The friction is the repeated manual work around the design decisions, like combing through policies, decks and transcripts by hand.

Overview

AI is changing how L&D works. What hasn’t changed is the judgment that makes learning stick.

I spent the week in two full-day AI workshops, twelve sessions and two keynotes. This page pulls together what I heard: where the field is, where AI fits in a design workflow, and where our skills matter most.

It is a report back from the conference, not a plan. People there were split on AI too, and most of the useful advice was about using AI carefully and on purpose.

The week, by themeWorkshopKeynoteSession5 days · 16 sessions

AI in the design workflowAI Intensive for Workflow LearningMon 2 · Workshop AI-Supported Learning DesignTue 3 · Workshop Less Grind, Better DesignWed 4 · Session Cohesive by DesignWed 4 · Session
Quality and measurementClone Your Best EvaluatorThu 5 · Session AI in L&D: The Instruction LayerWed 4 · Session Built to LearnWed 4 · Session
Leadership and changeThe Courage to Think DifferentlyWed 4 · Keynote Learning Design in the Age of AIThu 5 · Keynote Purpose-Paced LeadershipFri 6 · Session The HP BlueprintFri 6 · Session
Platforms and customer educationBeyond the LMSWed 4 · Session Clean Up Your LMMessWed 4 · Session Customer Education Is DistinctiveWed 4 · Session Customer Education Buzz DiscussionWed 4 · Session You’re Not Starting From ScratchWed 4 · Session

Sessions and speakers16 sessions · 28 speakers

Who said it

Every session this page draws on, with its speakers. Some ran at the same time, so a few are drawn from the slides alone. Session titles link to the DevLearn program.

Notes + slidesNotesSlides

Mon 2 Nov · Pre-conference workshop

8:30–4:30
AI Intensive for Workflow Learning: Enabling the 5 Moments of Need
Josh CavalierDr. Conrad GottfredsonBob MosherHandout by APPLY Synergies
Notes + slides

Tue 3 Nov · Pre-conference workshop

8:30–4:30
P06: AI-Supported Learning Design
Garima Gupta Founder and CEO, Artha Learning Inc.
Notes + slides

The L&D landscape

Where L&D is with AI

Most L&D teams now use AI. The research shared at DevLearn says the hard part has moved from trying it to using it well.

68%of L&D respondents say their organization is using AI in L&D, up from 42% in 2023Around the Core, Oct 2026 · from the report, Figure 2

41%of employees rated AI-produced learning only adequate or poor, while 62% used AI to learnElucidat, State of Digital Learning 2026 · via Garima Gupta's slides

33%of workers got AI training from their employer in six months, while 55% use AI regularlyThe Conference Board, Skilling for AI · via Garima Gupta's slides

29%of workforce and learning leaders report mature AI capability, though 83% rank AI their top priorityedX / 2U and NewtonX, Oct 2026 · via Garima Gupta's slides

AI becomes business as usual for L&DShare of respondents at each stage, 2023 to 2026

9%49%42%

2023

19%41%40%

2024

10%36%54%

2025

6%26%68%

2026

UsingExperimentingPre-use

Redrawn from Figure 2 of AI in L&D 2026: Around the Core, by Donald H Taylor and Eglė Vinauskaitė (October 2026).

What the research says

Adoption is growing faster than learning quality. Garima Gupta made this point early in her workshop, citing several 2026 reports.
Content and design are still the main uses. Content production and learning design remain the most common ways L&D uses AI.
Most of the change has stayed inside L&D. AI use has risen every year, but uses that connect to the wider business are barely more common than four years ago.
A small group is doing different things. 14% of teams say the business now comes to them for new kinds of help. They are twice as likely to have AI working with employees directly, and almost three times as likely to automate their own processes and build their own tools.

"Change has to be chosen."Around the Core, on four years of data

Use of AI in L&D by category, 2023 to 2026Ranked by share of respondents. Circle size shows the share; hover a circle for its value.

Rose in rankFell in rankHeld its rank

View as a table

Use	2023	2024	2025	2026	Rank change
Content development	69.5% #1	75.9% #1	77.9% #1	77.6% #1	No change
Learning design	46.8% #2	43.2% #2	49.6% #2	49.0% #2	No change
Administration and communications	21.3% #3	19.7% #3	30.5% #3	24.9% #4	Down 1
User-facing delivery	19.9% #4	17.5% #4	22.3% #6	16.6% #7	Down 3
Discovery	13.5% #5	13.3% #6	22.3% #5	23.5% #5	No change
Skills intelligence	13.5% #6	11.1% #7	9.5% #9	12.6% #9	Down 3
Knowledge and point-of-need support	9.9% #7	8.9% #8	12.6% #8	12.8% #8	Down 1
Evaluation and analytics	7.8% #8	15.2% #5	27.6% #4	34.1% #3	Up 5
Operations	7.8% #9	5.1% #9	17.3% #7	18.3% #6	Up 3
Principled non-use	1.4% #10	1.6% #10	1.5% #10	1.7% #10	No change

Redrawn from Figure 4 of AI in L&D 2026: Around the Core, by Donald H Taylor and Eglė Vinauskaitė (October 2026).

What gets in the way: barriers to using AI, 2024 to 2026Share of respondents naming each barrier. They could choose any number from a list of 13.

Readiness36%49% → 49% → 36%Down 13 points since 2024

Skills, knowing where to start, the business case

Friction66%60% → 68% → 66%Up 6 points since 2024

Integration, cost, time, resistance

Risk74%69% → 77% → 74%Up 5 points since 2024

Data privacy, trust in outputs, regulation

Share naming at least one barrier in the group, 2024 → 2025 → 2026. Adding ethical and environmental concerns takes Risk to 77%.

Getting started is no longer the problem. Barriers L&D can fix by itself fell from 49% to 36%. “Not knowing where to start” has halved since 2024.
Putting AI to work is. Integration, cost and time are all higher than in 2024. Integration is now the third most-named barrier.
Risk still leads. Data privacy and security tops the list at 52%, down from 59% last year. Trust in AI outputs is second.
A new concern. Ethical and environmental concerns, added to the survey in 2026, were named by 29%.

202620252024

Data privacy and security concerns

52%

Lack of trust in AI outputs

36%

Integration issues

33%

Lack of skills in L&D team

29%

Ethical and environmental concerns

29%

Regulatory or compliance barriers

27%

Concerns about cost

23%

Time constraints

22%

Resistance or lack of engagement

21%

Not knowing where to start

10%

Other

12%

Difficult to make a business case

None of the above

View as a table

Barrier	2024	2025	2026
Data privacy and security concerns	54%	59%	52%
Lack of trust in AI outputs	37%	39%	36%
Integration issues	25%	33%	33%
Lack of skills in L&D team	35%	36%	29%
Ethical and environmental concerns	—	—	29%
Regulatory or compliance barriers	23%	33%	27%
Concerns about cost	17%	25%	23%
Time constraints	18%	25%	22%
Resistance or lack of engagement	22%	24%	21%
Not knowing where to start	20%	16%	10%
Other	9%	12%	12%
Difficult to make a business case	10%	12%	6%
None of the above	6%	4%	5%
Group: Readiness (skills, knowing where to start, the business case)	49%	49%	36%
Group: Friction (integration, cost, time, resistance)	60%	68%	66%
Group: Risk (data privacy, trust in outputs, regulation)	69%	77%	74%
Group: Risk plus ethics	69%	77%	77%

Redrawn from Figures 7 and 8 of AI in L&D 2026: Around the Core, by Donald H Taylor and Eglė Vinauskaitė (October 2026). The survey had 531 respondents in 55 countries, who were self-selecting. Bar values are read from the report’s chart to the nearest whole percent.

Hopes and concerns

Voices in the room

Speakers raised the hopes and worries most L&D teams have. Nobody treated either side as wrong, and most landed in the same place.

What gave people pause

Are we saving time at the expense of thinking?Question to the Keynote 2 panel
Skills fade if we stop using them. We will see atrophy if we don't keep exercising ours.Keynote 2 panel
If you don't care about your job, AI is going to make you not care about it faster.Dakota Duffy, session 405
Make sure the knowledge and verification are there, rather than opening things up to an unguarded LLM.Beyond the LMS panel
Model drift is real. AI can get worse over time.Dakota Duffy, session 405

What gave people energy

Because we can test a lot more, our ambition has risen.Garima Gupta, AI-Supported Learning Design
Interactions that used to take too long are now possible. Using JavaScript was huge for us.Penske, Less Grind, Better Design
Visual consistency is a lighter lift once it's set up.Penske, Less Grind, Better Design
Once you stop asking which block to use in the authoring tool, you can focus on designing to a higher quality.From conference notes

Where most speakers landed

AI is a companion to our thinking, to our responsibility to help people to learn, not a replacement for it.Dan Pontefract, opening keynote (his slide)
It's about augmentation, not automation.Dr. Conrad Gottfredson
AI is our lever, to help us do our job, not replace our jobs.Brandy Ford, FM, Keynote 2
In learning, AI should elevate the learning experience and learning itself.Tessa Forshaw, Harvard Graduate School of Education, Keynote 2
SMEs are being empowered by AI, not replaced by it.Dakota Duffy, session 405

Some quotes are paraphrased.

The designer's edge

The skills AI can't supply

Much of the Keynote 2 panel was about this: Tessa Forshaw, Karl Kapp and Brandy Ford, hosted by Luke Hobson. The message was consistent: AI output is only as good as the designer directing it.

Behavior change

We know what changes behavior

"We have become a content engine," Tessa Forshaw said. The panel's answer: it's not content, it's designing instruction. We know our learners and how to motivate them, and that is the competitive advantage.

Tessa Forshaw and the Keynote 2 panel

Outcomes

Clear outcomes are what make a good prompt

Writing behaviorally defined outcomes is how you instruct AI well. So is telling a good example from a bad one. IDs already have both skills.

Keynote 2 panel

The how

AI can say what to teach. We know how.

When to introduce a concept and how to deliver an exercise is design expertise. Successful salespeople were the ones coached by their managers. AI output improves the same way when a designer coaches it.

Karl Kapp, Commonwealth University, Keynote 2

Practice

Knowing isn't the outcome. Doing is.

Information is the easy part. Knowledge becomes an asset when people can act on it, so we design for practice and feedback: prework, live practice, scenarios, then spaced reinforcement.

Megan Wheeler, Built to Learn

Judgment

Useful friction

L&D is the friction, and that friction is what adds value. SMEs aren't 100% accurate either. Someone has to question the content before learners see it.

Keynote 2 panel

Focus

Choose fewer things and check they work

The discipline is choosing fewer things and checking whether they work. Purpose first, then pace, then what people can realistically absorb.

Rebecca Alimorong, Purpose-Paced Leadership

Launches ≠ adoptionReleasing something is not the same as people using it

Completion ≠ capabilityFinishing a course is not the same as being able to do the work

More content ≠ better performanceVolume doesn't change behavior

Speed ≠ sustained impactFaster delivery doesn't make it last

Built to Learn · Megan Wheeler, Articulate

Four shifts for an information-rich, learning-poor era

Her case for L&D as “capability architects”: people who turn information into things others can do.

1Spot the signal

Find where learners keep getting stuck, and make the problem visible while people can still learn from it.

2Find the experts

Turn what experts know into what many people do. Ask “Tell me about a time…” and “How do you know it’s working?”

3Design for practice

Information is the easy part. Design for some form of doing, with feedback.

4Measure and iterate

Completions tell you who showed up. What changed in the work tells you what worked.

Her sequence for behavior that sticks

Prework. Interactive and self-paced. It protects live time.
Live session. Discussion and practice, not lecture.
Scenario practice. Apply it before it’s real.
Spaced reinforcement. It comes back afterwards. Not one-and-done.

From my notes on Built to Learn: The Capability Architect’s Playbook for the AI Era (DevLearn 2026, expo stage), presented by Megan Wheeler, Director, People Development, Articulate.

What to hand to AI

The useful question is which task, and how much judgment it needs

The opening workshop started here. Work changes task by task, not job by job, so decisions about AI belong at the task level too.

Task intelligence · Dr. Conrad Gottfredson

Understand work at the job-task level.
Assess AI risk for each task.
Build the team's skill around that task-level view.

The 5 Moments of Need

Start from when people need help

The workshop’s base framework. Each job task can be supported at five moments, and only the first two are training.

1New

Learning it for the first time

2More

Going deeper: exceptions and the why

3Apply

Doing it on the job

4Solve

When something goes wrong

5Change

When the work itself changes

How the workshop used it. Each table built one AI skill per moment, all reading from a single task map made from one SOP. When the SOP changed, they reran the skills.

From AI Intensive for Workflow Learning: Enabling the 5 Moments of Need (DevLearn 2026, workshop P02), presented by Josh Cavalier, Dr. Conrad Gottfredson and Bob Mosher. Handout © 2026 APPLY Synergies.

Lens 1 · Impact and difficulty

Rate each task on two scales

How bad is it if the task goes wrong, and how much judgment does it take? Each is rated 1 to 7. Where a task lands decides how much AI leads and how much a person leads.

Level of difficulty

732211116432211155432211454332213654332227654332177654331234567

Critical impact of failure

Human–AI task scale

1Human
2Human-driven
3Human-guided
4Joint collaboration
5AI-driven
6AI-led
7AI agent

Warm colors: a person leads. Cool colors: AI leads.

Redrawn from the Gottfredson-Cavalier Human-AI Task Allocation Matrix, © 2025–2026 Josh Cavalier, presented in AI Intensive for Workflow Learning: Enabling the 5 Moments of Need (DevLearn 2026, workshop P02) by Josh Cavalier, Dr. Conrad Gottfredson and Bob Mosher. Interactive version at joshcavalier.com/matrix.

Lens 2 · Four questions

Ask before handing a task to AI

Accountability

If this is wrong, who is accountable?

Risk of error

How expensive is the mistake?

Irreversibility

How easily can it be undone or corrected?

Sensitivity

Does it need situational judgment?

Redrawn from AI-Supported Learning Design (DevLearn 2026, workshop P06), presented by Garima Gupta, Founder and CEO of Artha Learning Inc.

Lens 3 · Six choices

Not every step needs AI

For each step in a workflow, pick one of six moves. Automating is only one of them.

Fix the process

Eliminate it

No clear value. Remove the step.

Simplify it

Too many clicks, approvals or handoffs.

Standardize it

Different people do it different ways.

Bring in automation or AI

Automate it

Predictable rules and repeatable inputs.

Augment it with AI

Needs drafting, synthesis, search or variation.

Keep it with people

Keep it human-led

High judgment, sensitivity or accountability.

Redrawn from AI-Supported Learning Design (DevLearn 2026, workshop P06), presented by Garima Gupta, Founder and CEO of Artha Learning Inc.

Lens 4 · Seven checks

Is this task a good fit for AI?

Penske’s team runs a task through this filter before bringing AI in. The first six look for safe time savings. The last looks for work that wasn’t possible before.

Is the task repeated often?
Is the input structured enough?
Can the output be verified?
Is the risk manageable?
Does a person keep the consequential decision?
Will the time saved be worth the review effort?
Could AI make something valuable possible that wasn’t practical before?

From the takeaway guide to Less Grind, Better Design: A Case Study of High-Impact AI in Instructional Development (DevLearn 2026, session 307), presented by Melissa Bitten and Bronson Domasky, Managers of Learning Design, and Jessica Ortiz, Project Manager, at Penske Transportation Solutions.

Putting it together

Human–AI delegation

The more risk, judgment or context a task carries, the more human-led it must be.

Level 1Judgment and decision-making

Making decisions that commit the organization or individuals.

Level 2Interpretation

Determining what something means in context.

Level 3Synthesis

Turning raw, messy inputs into structured representations.

Level 4Content generation

Producing first-pass material within defined constraints.

Level 5Administrative work

Mechanical execution of known steps where correctness is easy to verify.

Redrawn from AI-Supported Learning Design (DevLearn 2026, workshop P06), presented by Garima Gupta, Founder and CEO of Artha Learning Inc. Her framework builds on the five levels of autonomy defined by Tonsen (2025) and the human-in-the-loop strategies proposed by Barnes (2025).

When someone asks for an agent

Agentic doesn’t mean autonomous

HP’s learning team hears the same requests as everyone: an AI tutor, a career coach, a coach for managers, an agent for internal mobility. Brian Ciccotelli’s advice is not to assume leadership wants an agent just because they asked for one.

Four levels of agency

Name the tool before you build it

Each level hands the AI more say over the next step, and each needs more scrutiny. The examples all come from performance reviews.

Level 1Workflow

“When X happens, do Y.”

Triggered by: An event or rule
Human role: Checkpoint or recipient
Value: Efficiency

A review is submitted, the manager gets an email, a reminder is set.

Level 2Chatbot

“Tell me.”

Triggered by: A prompt
Human role: Consumer
Value: Information

“What day are my performance reviews due?”

Level 3Assistant

“Help me.”

Triggered by: A user request
Human role: Driver
Value: Productivity

“Help me summarize my review notes and draft something for Workday.”

Level 4Agentic AI

“Achieve X within these guardrails.”

Triggered by: A goal and context
Human role: Supervisor, approver, exception handler
Value: Adaptability

“Work with me to get reviews done for my entire team.”

Start lower than you think. Workflows, the simplest level, give the largest return.

Redrawn from The HP Blueprint: Scaling High-Stakes Performance with Agentic Workflows (DevLearn 2026, session 810), presented by Brian Ciccotelli, Learning Technologist at HP Inc.

The Agentic AI Blueprint

Six questions before you build

An agent is like an intern: eager to help, with no idea what it’s doing until you teach it.

1Outcome

What are you trying to accomplish?

Understand the business process
Run a needs analysis
Get a baseline and set a goal
Develop the business case

2Tool

What technology fits the need?

More agency means more scrutiny
Workflows give the largest ROI

3Input

What does the user need to provide?

Meet people where they already go
Decide what data they provide
Think through personas and scenarios

4Context

What does the tool need to know?

Frameworks and core principles
Goals and strategy
Dates, deadlines and processes
Resources

5Skills and guardrails

What can it do, and where are the boundaries?

Teach it what good and bad look like
Questions it should not answer
Escalation paths and access control

6Humans

Where do people review, decide or step in?

People give final approval of outputs
Design checkpoints and validation points

Redrawn from The HP Blueprint: Scaling High-Stakes Performance with Agentic Workflows (DevLearn 2026, session 810), presented by Brian Ciccotelli, Learning Technologist at HP Inc.

Launching one? HP’s phased rollout and pre-launch reviews are under Adoption.

AI across the design process

Where AI fits in a design workflow, phase by phase

This is a general design flow pieced together from what Penske, Garima Gupta and the Less Grind guide shared. It isn’t any one team’s process. In every phase, AI can take on part of the work, and a designer makes the call.

1AnalysisSources, SMEs, the real problem
2ObjectivesWhat people must do on the job
3DesignOutline, story and practice
4DevelopmentMedia, interactions, builds
5ReviewAccuracy, quality, sign-off
6EvaluationWhat changed in the work

Where AI can helpWhere we lead

Phase 1

Analysis

"If you do the analysis right, you are halfway there." Garima Gupta

AI can help

Comb through policies, decks, transcripts and old courses for themes and terms
Separate performance-critical content from reference material
Flag gaps, conflicts and unclear language
Draft sharper questions for SME interviews, and play the SME to test them
Map the topic: common mistakes and the path from novice to expert

We lead

Decide what belongs in training, what belongs in a job aid, and what needs a stakeholder answer first
Run the SME conversation
Make sure gaps are flagged, not filled with plausible content

Heard at DevLearn

Source content doesn't have to be SME documents. It can be performance data or logs. Garima Gupta.

Phase 2

Objectives

Moving from content coverage to observable action

AI can help

Draft observable, job-centered objectives
Tie each objective to the business need and the source
Point out objectives that describe information instead of performance
Argue against the draft: what would a skeptical manager say?

We lead

Confirm the stated performance is the right one
Decide whether training can change it at all
Make sure the assessment reflects the real decision or action

Heard at DevLearn

Writing behaviorally defined outcomes is exactly what makes a good prompt. IDs already have that skill. Keynote 2 panel.

Phase 3

Design and storyboard

The first pass is fast. The design choices are ours.

AI can help

Draft a first outline or script from approved content
Suggest decision points, believable wrong choices and consequences
Offer other story structures, or several versions to compare
Turn technical language into conversational narration

We lead

Choose the story worth telling, and check it feels real to the audience
Decide when to introduce a concept and how to run an exercise
Design for practice and feedback, not just information
Validate timing and scope. AI's estimate of course length is not a commitment.

Heard at DevLearn

AI can tell you what to teach, but the how is the expertise we bring. Karl Kapp, Keynote 2.

Phase 4

Development

Things that were too hard or too slow before

AI can help

Build custom interactions, small JavaScript pieces and animations
Keep characters, icons and illustrations consistent
Draft video, narration and translation
Make a small test version before full development

We lead

Decide whether a new capability improves learning or just adds novelty
Check for a simpler option that needs no AI
Test accessibility and whether the team can maintain it

Heard at DevLearn

Penske's visual design team built a brand kit for the designers. Visual consistency became a lighter lift once it was set up. Less Grind, Better Design.

Phase 5

Review and quality

Where human review matters most

AI can help

Check drafts against our standards and checklists
Score work against a clear rubric, showing its evidence
Say when it doesn't know, or can't prove something

We lead

Make the final call on accuracy, tone and fit
Bring in SME, legal, compliance or accessibility review
Calibrate the rubric and decide who rules when AI is wrong

Heard at DevLearn

The designer isn't a gate at the end. They're the thought partner who checks, questions and improves what AI hands over. AI isn't good without that input. Dakota Duffy, session 405.

Phase 6

Evaluation

Completions tell you who showed up

AI can help

Analyze learner data and the questions people ask most
Group similar questions and find likely root causes
Point out content that is out of date

We lead

Decide whether it's a training need or a documentation need
Look at what changed in the work, not just completions
Choose what to improve next. Training is never one-and-done.

Heard at DevLearn

What changed in the work tells you what worked. Megan Wheeler, Built to Learn.

Penske's result: fewer drafting hours, the same timelines, better courses. Their 2026 estimate for a storyboard dropped from 45 hours to 12. Course timelines did not change. The team put the time back into design and quality.

Penske's design processFour phases, with design and development run as loops

PreparationBackgroundInformation gathering

Iterative designDesignPrototypeReview

Iterative developmentDevelopImplementEvaluate

RolloutLaunchScale and deliver value

Discovery
Content analysis
Project planning
Design
Storyboard
Alpha
Beta
Gold

Where AI works for them

01Analyzing existing content
02Developing actionable learning objectives
03Creating first drafts of scripts and storyboards

Redrawn from Less Grind, Better Design: A Case Study of High-Impact AI in Instructional Development (DevLearn 2026, session 307), presented by Melissa Bitten and Bronson Domasky, Managers of Learning Design, and Jessica Ortiz, Project Manager, at Penske Transportation Solutions. Their process is adapted from Allen Interactions' SAM (Successive Approximation Model), 2021.

Hours per task, before and after AIPenske's estimates for one course, in hours

Analysis

20 hrs2023 original estimate

14 hrs2025 target

6 hrs2026 new estimate

High-level design

6 hrs2023 original estimate

4 hrs2025 target

3 hrs2026 new estimate

Scripting

20 hrs2023 original estimate

15 hrs2025 target

3 hrs2026 new estimate

Storyboarding

45 hrs2023 original estimate

32 hrs2025 target

12 hrs2026 new estimate

2023 original estimate2025 target (25% less)2026 new estimate

Redrawn from Less Grind, Better Design: A Case Study of High-Impact AI in Instructional Development (DevLearn 2026, session 307), presented by Melissa Bitten and Bronson Domasky, Managers of Learning Design, and Jessica Ortiz, Project Manager, at Penske Transportation Solutions. Course timelines did not change; the time went back into design and quality.

Quality and review

The guardrails teams described

Every team that talked about using AI well described a way to check its work. Speakers were clear that this isn't a stop gate at the end. The designer is a thought partner all the way through.

Penske Transportation Solutions · Less Grind, Better Design

AI output review checklist

A reusable checklist for reviewing AI output. They added it when going "from good to great."

Factual accuracy
Alignment with the source material
Unsupported assumptions
Missing content
Alignment with learning objectives
Realism of scenarios
Accuracy of technical instructions
Accessibility
Bias and representation
Tone and audience fit
Privacy or confidential information
Need for SME, legal, compliance or security review

Redrawn from the Human in the Loop Checklist in Less Grind, Better Design: A Case Study of High-Impact AI in Instructional Development (DevLearn 2026, session 307), presented by Melissa Bitten and Bronson Domasky, Managers of Learning Design, and Jessica Ortiz, Project Manager, at Penske Transportation Solutions.

Session spotlight · Thu 5 Nov

Clone Your Best Evaluator: Build AI Rubrics That Actually Capture Quality

Dakota Duffy, AI Transformation and Governance Leader. My pick for the best session of the week. She showed how to turn an expert evaluator's judgment into a rubric that people and AI can both apply, and test it before anyone trusts the score.

The challenge AI does not remove ambiguity.

It can turn an undocumented assumption into an apparently definitive score.

Expert judgmentTacit experience

Rubric languageUnstated boundaries

Model inferencePlausible interpretation

Stakeholder trustHard-to-defend decisions

What almost went wrong Expert preference ≠ demonstrated competence

"Here is how I would do it."

"Here is what the escalation evidence shows."

Her team built assessment scenarios around common escalation pathways and how to prevent them, alongside the experts' own habits.

SME interview method Interview for evidence

CompareWhat makes one response stronger than another?
ProbeWhich warning sign changes your decision?
ChallengeWhich different approach is still acceptable?
BoundaryWhat looks plausible but creates real risk?

Before and after One sentence can hide five scoring decisions

Before · bundled

"Proposes an appropriate technical solution."

After · granular

RequirementsConnects the solution to stated needs.

ReasoningJustifies constraints and assumptions.

RiskIdentifies solution-specific pitfalls.

MitigationAddresses that risk with a feasible action.

TradeoffsExplains an important alternative or limitation.

Granular, not fragmented: each criterion is one meaningful decision with observable evidence and a defensible boundary. More criteria is not better measurement.

Three-level example Criterion: identifies and addresses implementation risk

1 · Developing

No material risk, or the mitigation does not address the risk.

2 · Proficient

A relevant material risk, plus a feasible, directly connected mitigation.

3 · Exemplary

Meets Level 2, and also explains a tradeoff, limitation or residual risk.

Her reminder: actual standards need SME agreement and testing.

From Dakota Duffy’s session deck6 slides. Select one to enlarge it.

Dakota Duffy's slide: Four types of scoring ambiguity. Interpretation, sufficiency, boundaries and construct. — **Four types of scoring ambiguity**Interpretation, sufficiency, boundaries and construct.

Dakota Duffy's slide: Four parts of an interpretable criterion. Observable behavior, quality threshold, boundary condition, level anchors. — **Four parts of an interpretable criterion**Observable behavior, quality threshold, boundary condition, level anchors.

Dakota Duffy's slide: Require evidence before assigning a score. Four questions to answer before any score stands. — **Require evidence before assigning a score**Four questions to answer before any score stands.

Dakota Duffy's slide: Four ways the rubric still breaks. Bundling, relative language, hidden assumptions, overlapping levels. — **Four ways the rubric still breaks**Bundling, relative language, hidden assumptions, overlapping levels.

Dakota Duffy's slide: Rubric testing comes first, then validation. Calibration, repeated scoring, disagreement analysis, boundary cases. — **Rubric testing comes first, then validation**Calibration, repeated scoring, disagreement analysis, boundary cases.

Dakota Duffy's slide: Who decides when AI is wrong?. Ownership, traceability, escalation and change control. — **Who decides when AI is wrong?**Ownership, traceability, escalation and change control.

Your best evaluators are a finite resource. Build the rubric to scale and augment them, not replace them.Dakota Duffy, closing slide

Cards redrawn and slides shown from Clone Your Best Evaluator: Build AI Rubrics That Actually Capture Quality (DevLearn 2026, session 405), presented by Dakota Duffy, AI Transformation and Governance Leader.

Ask for evidence before a score

Where is the evidence in the work?
Which requirement does it meet?
Which level boundary decides the score?
What is missing, or needs a person to review? AI should say when it doesn't know.

Dakota Duffy, session 405

Decide who rules when AI is wrong

Ownership. Who approves the standard?
Traceability. Can we reproduce the score and its evidence?
Escalation. When does a person step in?
Change control. What triggers a recalibration?

Dakota Duffy, session 405

Start from standards and sources

A reusable AI workflow needs rules, examples and a source of truth.
AI models are fluent, not factual. They predict; they don't verify.
If the knowledge AI draws on is out of date, its answers will be too.
Create the content, then have AI check it against your standards.

Josh Cavalier, Dakota Duffy, Customer Education Buzz Discussion

Accessibility

Four accessibility checks AI can run first

These speed up the first pass. A person still tests with real assistive technology.

Draft alt text for images, for a person to edit.
Rewrite jargon in plain language.
Read a draft as a persona, such as a reader with dyslexia, and say where it is hard to follow.
Scan for mouse-only wording like “click” and “hover.”

If you build interactions with AI. Custom HTML stays accessible when the full code goes into the course. Embedding it as an iFrame is where it breaks.

From AI-Supported Learning Design (DevLearn 2026, workshop P06), presented by Garima Gupta, Founder and CEO of Artha Learning Inc.

Being open about AI use builds trust. Garima Gupta suggested telling learners when content is AI-assisted and human-verified. It sets clear expectations and avoids the backlash other industries saw when AI use came out later.

Prompting

Less magic phrasing, more context and intent

The prompt frameworks and techniques shared across the week. Most came from Garima Gupta’s workshop, with formulas from Josh Cavalier and the Penske team.

The state of prompting, November 2026

What to stop, and what to do instead

Stop

Magic phrases like “you are a world-class expert”
“Think step by step” with reasoning models
Long, rigid templates

Context and intent: who it’s for and why it matters
Examples of what good looks like
The output shape, what to avoid or verify, and when to stop and ask

Context beats clever wording. Garima Gupta called it context engineering: give the AI the right background for the task. With retrieval (RAG), it looks things up in the documents you provide before it answers. Those sources can be performance data or logs, not just SME documents.

From AI-Supported Learning Design (DevLearn 2026, workshop P06), presented by Garima Gupta, Founder and CEO of Artha Learning Inc.

Prompt frameworks

Four formulas worth keeping

They overlap on purpose. Each one makes you say who it’s for, what you want back and what good looks like.

CAREGarima Gupta’s everyday formula

CContextBackground and constraints
AActionWhat needs to be done
RResultThe type of result and how it’s presented
EExampleA reference to work from

Garima Gupta, Artha Learning Inc. · AI-Supported Learning Design

TRACIJosh Cavalier’s structured prompt

TTaskWhat to produce
RRoleWho the AI is acting as
AAudienceWho it’s for
CCreateThe shape of the output
IIntentWhy it matters and what success looks like

AI Intensive, Josh Cavalier

RCTOGPenske’s five-part pattern

RRoleWho the AI is
CContextWhat it needs to know
TTaskWhat to do
OOutputThe format to return
GGuardrailsWhat to avoid or check

Less Grind, Better Design, Penske

4SGarima Gupta’s formula for images

SSubjectThe who and what
SStyleIcon, photo, cartoon
SShotCamera work and framing
SSpecsAspect ratio and other output details

Garima Gupta, Artha Learning Inc. · AI-Supported Learning Design

Make the best ones permanent. “A skill is a TRACI prompt made permanent, plus the rules a model needs every time.” Josh Cavalier

Techniques

Seven ways to get a better answer

Chain of thought

Give it the steps. Define the task, then show the steps to follow.

Use it for Skip “think step by step” with reasoning models; they already do.

Chain of density

Summarize, then rewrite two to five times at the same length, adding one to three new facts each pass.

Use it for Denser summaries of long source documents.

Prompt chaining

Ask for more, then select. Have it draft, then revise three times, one improvement per pass.

Use it for Storyboard descriptions you can pick from.

Meta prompting

Ask the AI to write the prompts: it breaks the job into sub-prompts, runs them and compiles the result.

Use it for Big tasks you haven’t broken down yet.

Flipped interaction

The AI interviews you first. “Ask me the questions you need answered to do the best possible job.”

Use it for Briefs where you don’t know what you haven’t said.

Self-consistency

Run the same prompt several times and choose the answer with a defined rule.

Use it for Assessment items and anything that must be consistent.

Chain of verification

Draft an answer, plan verification questions, answer them separately, then write the final answer.

Use it for Factual content that SMEs will check.

From AI-Supported Learning Design (DevLearn 2026, workshop P06), presented by Garima Gupta, Founder and CEO of Artha Learning Inc.

Structured responses

Ask for the shape you want, not just the content

1. Hook
2. Practice
3. Debrief
ListsSequences and steps
Try asking“Give me the steps as a numbered list.”
TablesComparisons
Try asking“Compare these three options in a table: cost, time, effort.”
```
{
  "question": "…",
  "answer": "B"
}
```
JSONOutput that feeds another tool
Try asking“Return the quiz items as JSON with question, options and answer.”
```
<section>
  <h2>…</h2>
</section>
```
HTML or codeWhen you’re building, not writing
Try asking“Write it as an HTML snippet I can drop into my course.”
```
Topic
├─ Goal
│  └─ Skill
└─ Risk
```
Text mind mapsTo see relationships
Try asking“Map this topic as an indented text mind map.”

From AI-Supported Learning Design (DevLearn 2026, workshop P06), presented by Garima Gupta, Founder and CEO of Artha Learning Inc. The “Try asking” prompts are my own examples.

Prompting for visuals

Consistent characters, believable photos

Image prompts need the same clarity as text prompts, plus a vocabulary for style, composition and specs.

Christy Tucker · Cohesive by DesignSeven parts of an image prompt

Image style
Subject
Action
Physical characteristics
Clothes
Setting
Additional details

Christy Tucker · Cohesive by DesignKeeping a look consistent

Ask an LLM to describe a style you like, then reuse its key phrases in every prompt.
Start a character from one reference image.
Change one or two elements at a time, and say what to keep.
If the character drifts, start a new chat or go back to the reference image.

Garima Gupta · AI-Supported Learning DesignMaking people look real

Ask for an “editorial photo,” not “photorealistic” or “hyper-realistic.”
Describe real skin: pores, freckles, fine lines, uneven tone.
Describe the foreground, midground and background so the scene feels lived in.

Garima Gupta · AI-Supported Learning DesignJSON prompting for edits

Have the AI convert an image into a structured JSON description.
Edit the JSON: the camera, the lighting, one character’s expression.
Apply it back. The AI follows JSON closely, so only what you edited changes.

Trying it in the workshop

My practice images from the day: one character, placed in a scene, then changed with a JSON edit.

AI-generated photo of a man with dark curly hair and a grey shirt in an office, used as a character reference. — 1. A character reference

AI-generated over-the-shoulder photo of the same man talking with a woman across a meeting table. — 2. The character placed in a scene

The same meeting scene with the woman’s expression and gesture changed, and everything else kept. — 3. The same scene after a JSON edit

Tips from the slides for Cohesive by Design: Generate Consistent AI Icons, Illustrations & Characters (DevLearn 2026, session 102), presented by Christy Tucker, LXD Consultant, Syniad Learning, and from AI-Supported Learning Design (workshop P06), presented by Garima Gupta, Founder and CEO of Artha Learning Inc. Images are AI-generated and mine.

Beyond the prompt

From prompt to skill to agent

The AI Intensive’s core idea: a team’s use of AI skills up the way people do, one step at a time.

1Prompt

A detailed prompt for one step of a task. You type it every time.

People: perform one skill

2Skill

The prompt made permanent, with the rules, examples and source of truth it needs. The AI loads it when the task matches.

People: learn the steps of a job task

3Automation or agent

An automation reruns fixed rules on a trigger. An agent decides what needs to change, and asks when it isn’t sure.

People: combine skills into a workflow

4Agent swarm

Several agents handling more than one process at once.

People: work as a team

Fix the skill, not the output. A hand-fixed output is right once. A fixed skill is right every time.

What’s inside a skill

A skill loads in three layers. If a skill never runs, fix the description first.

Description

The name and a short description. Always loaded, and it decides when the skill runs.

Instructions

The body of the skill file: purpose, steps, rules and output.

Reference files

The SOP, style guide, templates and examples. Loaded only when a step needs them.

A 30/60/90-day start

The workshop closed with this roadmap. Each step gets an owner, a success measure, its biggest risk and a human checkpoint.

30One skill in production

Run it on a real SOP and get SME sign-off.

60One automation

Trigger the skill when the SOP changes, and set the human review gate.

90One agent pilot

Score it on the task matrix. Name an owner, a success metric and a cost check.

If your team built on Custom GPTs. OpenAI plans to retire them on 11 December 2026 and is moving their instructions into skills inside plugins. The workshop handout also lists Gemini Gems as converting to skills. Dates can change, so check OpenAI’s retirement FAQ.

Content health

AI answers are only as current as the content behind them

Five sessions came at the same problem from different sides: libraries full of stale content, and AI tools that now answer from it. Governance is what keeps both trustworthy.

“Training sets the starting point. Knowledge sets every answer after.”Caleb Burrill, eGain · AI in L&D: The Instruction Layer

Advocate Health · Clean Up Your LMMess

Four steps to govern a learning library

Amanda Archer’s team cleaned up a large LMS by agreeing the rules first, then the process.

1Identify your goals

Decide what the library is for before deciding what stays in it.

2Create your governance

Set the rules for inclusion, ownership, lifecycle and assignment.

3Add processes and policies

Intake, sponsor approval, course renewal and deactivation.

4Communicate and get buy-in

From the start, not just at the end: webinars, a stakeholder committee and champions.

Four things the rules cover

The thresholds are Advocate Health’s own. The categories transfer to any library.

InclusionWhat earns a place

A true learning solution with clear objectives
Supports an organizational goal
For 100 or more learners, or part of orientation, a job requirement or a regulation

OwnershipWho keeps it current

One named owner per item
The owner answers questions about it
The owner names a backup

LifecycleWhen it is reviewed or retired

At most one update a quarter
A review at least every two years
Retired if it has no owner, no review in two years, or no use in 12 months

AssignmentWho gets enrolled

Only required learning is assigned
A sponsor approves it first
The owner is responsible for the audience being right

From the slides for Clean Up Your LMMess Using Learning Governance (DevLearn 2026, session 204), presented by Amanda Archer, AVP, Learning Experience & Transformation, Advocate Health.

Heard in other sessions

Four more ways to keep content healthy

Caleb Burrill, eGain · session 116Find what the AI is answering from

Check the knowledge center first, and make sure it is accurate.
Connect real conversations to find the most-asked questions.
Group the variants and find the root cause: a documentation need or a training need?
Rank by volume and complexity, then refine with SMEs.

Brie Cohen Vlach, Solventum · session 108Build content that is easy to update

Write standalone snippets, so one change doesn’t break a course.
Ask SMEs to flag what will go out of date quickly.
Be an awareness gate, not a stop gate: minor releases get quick fixes, major ones get a review.
Keep a snapshot of content health: what needs an update now?

Gwen O’Neil, U-Haul University · expo talk A101Start from what you already have

Catalog the existing content before building anything new.
Use a tool that reviews the library and points out what is out of date.
Track what people search for, and shape learning paths from it.
Build new content from the catalog.

Beyond the LMS panel · session 214Ask more of the platform

The shift is personalization: where to focus, not how many hours.
Learning happens whether or not a course is completed. Decide how to see it.
Ask each vendor what makes them different. Compliance is the test of “enterprise-ready.”
Make sure the knowledge and verification are there before opening things up to an LLM.

From my notes. Points are paraphrased.

AI tools

The tools that came up, and where they land in the work

Every tool named in the sessions, placed at the step of the design process where it helps. Names link to the tool.

Step 1

Understand the tech

Before you design anything: build your own sense of how these models work, and start a prompt library.

Prompt library on Notion

A Notion library of prompts to borrow from. Start your own, then turn the best prompts into skills.

Josh Cavalier, AI Intensive

Embedding Projector

Shows words as points in 3D space, a quick way to see what “meaning” looks like to a model.

Josh Cavalier, AI Intensive

LLM Visualization

A small GPT you can watch step by step, to show teams how tokens move through a model.

Josh Cavalier, AI Intensive

Step 2

Design and write

Analysis, objectives, outlines and scripts, where the designer’s judgment matters most.

DeepL

Translation for localizing courses and job aids, with a human check on terms.

Garima Gupta, Artha Learning Inc. · AI-Supported Learning Design

Simplified Technical English

A controlled-language standard to point AI at when writing for global and non-native readers.

Garima Gupta, Artha Learning Inc. · AI-Supported Learning Design

Step 3

Create media

Images, characters, voice and music,
made to one consistent look.

Adobe Firefly

Adobe’s AI assistant for images and edits inside the tools many design teams already use.

Garima Gupta, Artha Learning Inc. · AI-Supported Learning Design

Image generators

Gemini (Nano Banana), ChatGPT or Copilot, Magnific, Flora, Leonardo and Canva, for icons, illustrations and consistent characters. Use the ones your org approves.

Christy Tucker, Cohesive by Design

Cohesive by Design Padlet

Her session board with examples of styles, prompts and character consistency.

Christy Tucker, Cohesive by Design

ElevenLabs

Natural voices for narration and audio role-plays.

AI Intensive workshop

Suno

Turns an idea into music, for intros, beds and mood.

AI Intensive workshop

Synthesia

AI video presenters, used at Penske for course openings and level-setting.

Penske, Less Grind, Better Design

Step 4

Build practice

Role-plays, coaching and conversation practice that respond to the learner.

BTS Verity

Conversation simulations where learners practice and get scored.

AI Intensive workshop

Dessa

A speech-to-speech model that listens and talks at the same time, for natural spoken role-play.

AI Intensive workshop

Synthesia

AI avatar role-plays, so learners can practice a conversation with a realistic video character, not just watch one.

Penske, Less Grind, Better Design

AIReady

Adds AI coaches, role-plays, document-grounded answers and audio inside Storyline and Rise.

Garima Gupta, Artha Learning Inc. · AI-Supported Learning Design

Step 5

Author and build

Turning the design into a course, a help article or a structured content library.

Mindsmith

AI-native eLearning authoring.

Garima Gupta, Artha Learning Inc. · AI-Supported Learning Design

Parta.io

No-code, on-brand course authoring with AI drafting, real-time collaboration and one-click translation. Exports to SCORM, PDF and HTML.

DevLearn Expo

MadCap Flare

Structured authoring for help content and open documentation, published from one source.

Brie Cohen Vlach, Solventum

Oxygen

Structured XML authoring for documentation teams.

Tools to check out

Intellum

An LMS with built-in content authoring.

Tools to check out

Step 6

Deliver and support

Help at the moment of need, inside the product or the workflow.

WalkMe

In-app guidance that walks people through a product as they use it.

Brie Cohen Vlach, Solventum

Step 7

Automate the work around it

The handoffs, file moves and reminders. Start with what your IT team already approves.

Make

Visual automations that connect your apps.

Garima Gupta, Artha Learning Inc. · AI-Supported Learning Design

Zapier, n8n, Power Automate, Copilot Studio

Other automation and agent builders. Plan each one as trigger, input, logic, output, then a human gate.

Josh Cavalier, AI Intensive

CrewAI, LangGraph and agent SDKs

For building your own agent. Or skip the build and run one in an agent workspace such as Claude, Copilot Cowork, ChatGPT or Gemini.

Josh Cavalier, AI Intensive

Before you adopt one

Run a quick tech audit

EfficiencyDoes it make us faster?
Cost and ROIDoes it save or make us money?
IntegrationDoes it work with our systems?
Future-proofingWill it grow with us?
Data privacyWhat does the vendor do with our data, how long is it stored, and who are the sub-processors?

From A Framework for a Smart AI Tech Audit in AI-Supported Learning Design (DevLearn 2026, workshop P06), presented by Garima Gupta, Founder and CEO of Artha Learning Inc.

Change and adoption

How other teams handled mixed reactions

Penske was asked directly about pushback and job fears. Their answer, the customer education buzz discussion and the keynote panel shared a thread: build skills together, start small and show results.

Penske

When people worried about their jobs

They brought in outside experts to talk to the team.
They built skills together and made sure everyone was aligned.
They created a shared checklist for reviewing AI output, so every designer reviews the same way.

Customer Education Buzz Discussion

Start with the work nobody wants to do

Ask what we really don't want to do, and try AI there first.
Keep it optional, not mandatory. Give people other projects.
Use AI to check work against standards.
Capture tribal knowledge so it isn't lost.

Keynote 2 panel

Prove it before scaling it

Be clear on the purpose of the tool and the business problem it solves.
Find a safe place to pilot, with volunteers and champions.
Train the trainer, then put KPIs in place to show it works.

What panelists said to prioritize

Move past the idea that we just create content.
Upskill the team.
Redesign how work and roles are set up.
Realign with what leaders are looking for.

What panelists said to stop

Saying yes to every single request.
Using AI for everything.
Forgetting that AI is a lever. Know when you need to be the strategic thinker.

HP · Launching an AI agent

Roll out in phases, and manage it as a change

HP’s launch plan for its agent: start with a high-impact group, use ADKAR throughout, and clear the reviews before anyone sees it.

Awareness

Why is the change necessary?

Desire

What’s in it for me?

Knowledge

Do I know what to do and where to go?

Ability

Do I have the skills to do it?

Reinforcement

Can I go back to my old way?

Before launch

Data privacy Avoid personal data.
Cybersecurity No memory, no servers, no external APIs.
Legal Disclaimers in place.
Works councils Optional, with no effect on pay or promotion.
AI/ML review Approved models, nothing trained on company data, guardrails in place.
Employee comms Core messages drawn from ADKAR.

Redrawn from The HP Blueprint: Scaling High-Stakes Performance with Agentic Workflows (DevLearn 2026, session 810), presented by Brian Ciccotelli, Learning Technologist at HP Inc.

Key takeaways

What I'm taking home

AI is part of L&D now. How we use it is still our choice.Two in three teams use it. The research and the speakers agreed that change has to be chosen, not assumed.
Think in tasks.Some steps are worth handing to AI and some are not. Risk, difficulty and the need for judgment decide which.
Our expertise is the point.Behavior change, clear outcomes, story, practice and judgment are what turn a draft into learning.
Quality comes from standards and review.AI output is only as good as the sources and standards behind it, and the people who check it.
Put saved time back into design.The teams doing this well spent the hours AI saved on iteration, practice and testing with learners.

If you want to start somewhere

Five things to try this week

My own suggestions, drawn from the sessions. Each is small enough to do without a budget or a new tool.

Rate five of your tasks. Score each for impact of failure and difficulty, and see where a person should lead. AI Delegation
Turn your most-reused prompt into a skill. Add the rules, an example and the source it should read from. Frameworks · Prompting
Agree one review checklist. So every designer checks AI output the same way. Frameworks · Quality and review
Rewrite one rubric line. Split a bundled criterion, write three level anchors, and test it on real work. Frameworks · Quality and review
Set one retirement rule. For example: no owner, or no use in 12 months, and the content comes down. Frameworks · Content health

AI gets us to version one. We decide what version is right.Penske, Less Grind, Better Design

The human side

The PMI Team

The best part of the week was the human connections we made with so many smart people in our field. Here’s the PMI team at DevLearn 2026.

Two team members smile in a selfie in the keynote hall, with the DevLearn stage glowing behind them. — Keynote morning

A team member takes a selfie while two colleagues strike playful poses in front of the #DevLearn sign. — Striking a pose at the #DevLearn sign

Three PMI Learning team members smiling in a selfie in front of the large DevLearn letters in the conference hallway. — At the big DevLearn letters

Three PMI Learning team members in a selfie in front of the DevLearn photo wall. — Photo wall, MGM Grand

People we met

Facilitators, speakers and new friends

Joan stands with three workshop facilitators wearing speaker badges in a hotel meeting room. — With the AI Intensive facilitators, Mon 2 Nov

Joan and a smiling exhibitor in a cap take a selfie in front of the Parta booth in the expo hall. — Exploring new learning tools with Parta in the expo hall

From the stage

Keynotes and workshops

Select a photo to enlarge it.

A packed ballroom faces a wide stage screen reading Welcome to DevLearn. — **Welcome to DevLearn**Opening morning, Wed 4 Nov

Dan Pontefract on the big screen beside a yellow slide reading: AI is a companion to our thinking, to our responsibility to help people to learn, not a replacement for it. — **"AI is a companion to our thinking"**Dan Pontefract, opening keynote

Dan Pontefract on stage beside his Open Thinker Cycle slide: creative thinking, critical thinking and applied thinking. — **The Open Thinker Cycle**Dan Pontefract, opening keynote

The keynote screen introduces the podcast guests: Dr. Karl Kapp, Dr. Tessa Forshaw and Brandy Ford. — **Meet the podcast guests**Keynote 2, Thu 5 Nov

Garima Gupta presents a slide titled You're probably overprompting Claude to a workshop room. — **"You're probably overprompting Claude"**Garima Gupta, AI-Supported Learning Design

The closing slide of the AI-Supported Learning Design workshop reads Maintain your momentum. — **"Maintain your momentum!"**Closing the AI-Supported Learning Design workshop

After hours in Las Vegas

The conference didn't stop at 5

The Wizard of Oz end credits glowing across the curved screen inside the Sphere. — The Wizard of Oz credits, filling the Sphere

Three team members smile in their seats inside the Sphere, glowing red and orange before The Wizard of Oz. — Inside the Sphere before The Wizard of Oz

Three team members take a selfie at night outside the Sphere, beside giant striped witch legs from the Wizard of Oz show. — Outside the Sphere after The Wizard of Oz

5 days2–6 Nov in Las Vegas

2 full-day workshopsAI Intensive and AI-Supported Learning Design

16 sessionsKeynotes, breakouts and two expo talks

1 favoriteClone Your Best Evaluator, Dakota Duffy

Three key takeaways

AI is not replacing L&D. It is changing it.

Learning teams are not content engines.

AI can accelerate the work, great or sloppy.

AI is changing how L&D works. What hasn’t changed is the judgment that makes learning stick.

Who said it

Mon 2 Nov · Pre-conference workshop

Tue 3 Nov · Pre-conference workshop

Wed 4 Nov

Thu 5 Nov

Fri 6 Nov

Also referenced

Where L&D is with AI

What the research says

Voices in the room

What gave people pause

What gave people energy

Where most speakers landed

The skills AI can't supply

We know what changes behavior

Clear outcomes are what make a good prompt

AI can say what to teach. We know how.

Knowing isn't the outcome. Doing is.

Useful friction

Choose fewer things and check they work

Four shifts for an information-rich, learning-poor era

Her sequence for behavior that sticks

The useful question is which task, and how much judgment it needs

Start from when people need help

Rate each task on two scales

Ask before handing a task to AI

Not every step needs AI

Is this task a good fit for AI?

Human–AI delegation

Agentic doesn’t mean autonomous

Name the tool before you build it

Six questions before you build

Where AI fits in a design workflow, phase by phase

Analysis

Objectives

Design and storyboard

Development

Review and quality

Evaluation

The guardrails teams described

AI output review checklist

Clone Your Best Evaluator: Build AI Rubrics That Actually Capture Quality

Ask for evidence before a score

Decide who rules when AI is wrong

Start from standards and sources

Four accessibility checks AI can run first

Less magic phrasing, more context and intent

What to stop, and what to do instead

Four formulas worth keeping

Seven ways to get a better answer

Ask for the shape you want, not just the content

Consistent characters, believable photos

Trying it in the workshop

From prompt to skill to agent

What’s inside a skill

A 30/60/90-day start

AI answers are only as current as the content behind them

Four steps to govern a learning library

Four things the rules cover

Four more ways to keep content healthy

The tools that came up, and where they land in the work

Understand the tech

Design and write

Create media

Build practice

Author and build

Deliver and support

Automate the work around it

Run a quick tech audit

How other teams handled mixed reactions

When people worried about their jobs

Start with the work nobody wants to do

Prove it before scaling it

What panelists said to prioritize

What panelists said to stop