AI in Learning and Development: Personalized Learning Paths and Content

By Kevin Kim · Updated Oct 11, 2026

AI in learning and development (L&D) is the use of machine learning and language models to find where employees lack skills, recommend what each person should learn next, produce training material faster, and coach people between formal courses. The aim is to replace one-size-fits-all curriculum with learning that fits the person and the job, while L&D staff keep control of what gets taught and what counts as "good."

This guide is for L&D leads, HR managers and department heads who are deciding where AI belongs in their training. It covers the four jobs AI does in L&D, what research says about whether it works, a step-by-step rollout, and the mistakes that sink most projects. It sits inside the AI in HR section, so it stays on the learning function and points elsewhere for the rest of the employee lifecycle.

Key Facts

• In LinkedIn's 2025 Workplace Learning Report, 71% of L&D professionals say they are exploring, experimenting with or integrating AI into their work (LinkedIn Learning).
• In the same report, 49% of L&D and talent professionals agree that their executives are concerned employees lack the skills to execute the business strategy (LinkedIn Learning).
• A meta-analysis of 50 controlled evaluations of intelligent computer tutoring found a median gain of 0.66 standard deviations in test scores, enough to move a typical student from the 50th to the 75th percentile (Kulik and Fletcher, 2016). Those were students, not employees.
• In a field experiment with nearly 1,000 high school math students in Turkey, an unrestricted chatbot raised practice scores by 48% but left students 17% worse on the exam without AI, while a tutor-style bot with guardrails left exam results about level with the control group (Wharton).
• Among 5,179 customer support agents given a generative AI assistant, productivity rose 14% on average and 34% for novice and low-skilled workers, with minimal impact on experienced workers (NBER).
• The EU AI Act lists emotion recognition in workplaces and education institutions as a prohibited practice, a ban that took effect in February 2025 (European Commission).

The Four Jobs AI Does in L&D

Traditional L&D runs on a catalog and a calendar: a list of courses, and a schedule of when people take them. AI changes the unit of work from "course" to "person." In ACE Framework terms, most L&D AI ingests learner records, job data and content, analyzes them to find gaps, predicts what a person should learn next, generates material and feedback, and executes small actions such as enrolling someone or sending a nudge.

Read the last column as the real design. Every job above has a person who owns the result. If you can't name that person for a given use, you aren't ready to deploy it.

Job 1: Skills Gap Analysis

A skills gap is the distance between what a role needs and what the people in it can do. Measuring it used to mean surveys, manager ratings and an annual training needs analysis. AI shortens that by reading evidence that already exists: job descriptions, performance notes, project histories, certifications, and course completions.

The practical flow looks like this:

1. Define the target. List the skills each role requires, at what level. If you don't have a skills framework, start with your ten most important roles, not the whole company. The groundwork is covered in skills-based talent strategy.
2. Gather the evidence. Pull from HR records, learning platform history and, where appropriate, self-assessments and manager input.
3. Let the model infer and compare. It estimates a skill profile per person and measures the gap against the target role.
4. Aggregate to the team and company level. The useful output is rarely one person's gap. It's the pattern: "Eight of twelve account managers lack negotiation practice."
5. Validate with humans. Managers confirm or correct the findings before anyone acts on them.

The pressure to do this well is real. In LinkedIn's report, half of L&D and talent professionals say executives worry about whether employees have the right skills (LinkedIn Learning). Skills inference is a good use of AI because the output is a hypothesis to check, not a verdict on a person. It becomes risky the moment an inferred skill score feeds a promotion or layoff decision. Keep it in the development lane. Clean inputs matter too, and HR data foundations for AI explains what to fix first.

Job 2: Personalized Learning Paths

A personalized path is a sequence of learning activities matched to a person's role, current skill level, goals and pace. The AI part is the matching and the re-matching: someone who passes the pre-assessment skips the basics, and someone who struggles gets a different explanation.

What the evidence says is encouraging for the design but narrower than vendor pitches suggest. The strongest research base is on intelligent tutoring systems, software that adapts instruction step by step. Kulik and Fletcher reviewed 50 controlled evaluations and found a median gain of 0.66 standard deviations, with the size of the gain depending heavily on whether outcomes were measured on locally developed or standardized tests (Kulik and Fletcher, 2016). That is a strong result, and it describes students in structured courses. It does not prove that a recommendation engine suggesting LinkedIn-style courses to adults improves job performance. Treat it as support for adaptive practice with feedback, not for "AI picks your courses."

A path worth building has these parts:

• A clear target (a role, a skill, a project), not an open-ended "learn AI."
• A baseline check so people don't repeat what they know.
• Short activities mixed with practice on real work.
• A manager touchpoint. In LinkedIn's data, only 15% of employees say their manager helped them build a career plan in the past six months (LinkedIn Learning). AI can prepare the conversation, but it can't replace it.
• A visible link from learning to opportunity. Paths feel pointless if finishing them changes nothing. Connecting them to internal moves is the idea behind the AI internal talent marketplace.

If you want to see the build side, the AI learning and development agent blueprint shows how an agent recommends paths, enrolls learners and nudges people who stall.

Job 3: Content Creation

Content is where most L&D teams start, because the benefit is easy to see: a first draft in minutes instead of days. A model can turn a policy document into a lesson outline, write quiz questions, create role-play scenarios, summarize a recorded workshop, and translate material for another region.

Used well, this frees subject matter experts and instructional designers to review and improve rather than start from blank pages. Used badly, it produces a library of confident, generic material that nobody checked. Three rules help:

• Ground it in your sources. Give the model your actual process documents and ask it to work only from them. Generic output is how wrong procedures end up in training.
• Keep an expert in the loop. Every compliance, safety or technical module needs a named reviewer who signs off before release.
• Test with learners. A module can read well and still teach nothing. Check that people can do the task afterward.

The broader skills for HR teams working with these tools are in generative AI for HR teams.

The Brynjolfsson, Li and Raymond study is a useful reference here, with a caveat about what it measured. They studied 5,179 customer support agents who got a generative AI assistant, and found productivity rose 14% on average and 34% for novice and low-skilled workers, with minimal impact on experienced workers (NBER). That was on-the-job assistance, not a training program. But it points to a design idea: AI that supports people while they work may help newer employees most, which makes a good argument for putting guidance inside the workflow instead of only in a course.

Job 4: Coaching and Practice

Coaching is the newest and least settled use. It covers chat assistants that answer "how do I handle this?" at the moment of need, simulations where a learner practices a difficult conversation, and feedback on written work.

The most important evidence here is a caution. In a field experiment with nearly 1,000 high school math students in Turkey, students given an unrestricted chatbot did 48% better on practice problems than the control group, but 17% worse on an exam taken without AI. A tutor-style version that gave hints instead of answers did 127% better in practice and about the same as control on the exam (Wharton). The students were teenagers learning math, not employees learning a job, so don't carry the numbers over. The lesson does carry over: if the AI does the thinking, performance during the exercise rises while learning may not.

For workplace coaching, that suggests a few design choices:

• Make the AI ask questions and give hints before it gives answers.
• Use practice simulations for skills people will perform alone later, such as feedback conversations and customer calls.
• Assess without the AI sometimes, so you know what people can do on their own.
• Keep sensitive topics (performance problems, conflict, health) with a human.

Be careful about tools that claim to read emotion from voice or video during practice or assessment. The EU AI Act prohibits emotion recognition in workplaces and education institutions (European Commission), and the science behind such claims is contested.

Real Examples by Function

L&D team. A learning lead feeds job descriptions and course history into a skills tool, finds that a third of the support team has never practiced de-escalation, and commissions a short scenario-based module. The AI found the pattern; the lead decided what to do about it.

Sales enablement. A manager uses a role-play bot so new reps can rehearse discovery calls before talking to a prospect, then reviews three recordings a week for coaching conversations. The practice volume goes up, and the manager's time goes to the conversations that matter.

Compliance and safety training. Instructional designers draft refreshed modules from updated policies, and a compliance owner reviews every line. The drafting is fast, and the sign-off is slow on purpose.

Frontline operations. A shift supervisor gets a short, role-specific refresher recommended after a process change, instead of a company-wide course nobody finishes.

Individual contributors. An analyst asks an assistant to explain a statistical method using the company's own reports as examples, then tests herself with generated questions. She owns her learning; the tool shortens the route.

How to Roll Out AI in L&D, Step by Step

1. Pick one business problem. "Reduce time to productivity for new support hires" beats "adopt AI in learning." A narrow target makes success measurable.
2. Audit your learning content and data. Know what courses exist, who owns them, whether they are current, and what the learning platform records. AI can't recommend what's missing or mislabeled.
3. Define skills for the target roles. Even a short, agreed list of skills and levels gives the system something to match against. A formal training needs analysis is a structured way to find the gap first.
4. Start with content drafting or skills inference. These carry the least risk to individuals and show value quickly.
5. Name an owner for each AI use. Write down who reviews output, what they can override and how often.
6. Set privacy rules before launch. Decide what employee data the system sees, who sees the results, and whether individual skill data is visible to managers. Tell employees plainly.
7. Pilot with one group for a quarter. Measure against a comparison group where you can, not just satisfaction scores.
8. Add paths and coaching next. Once the content and skills data are reliable, personalization has something solid to stand on.
9. Connect learning to outcomes. Link completion data to performance and mobility metrics through people analytics, watching for correlation being mistaken for cause.
10. Review and retire. Remove courses that nobody finishes or that no longer match the work. A recommendation engine amplifies whatever catalog you give it.

For the larger question of rebuilding skills across a workforce, reskilling programs that actually work goes deeper.

What to Measure

Satisfaction scores are the weakest signal in L&D, and AI doesn't change that. Pair them with measures tied to work.

• Adoption: the share of eligible employees who start and finish assigned paths.
• Time to proficiency: how long new or transferred employees take to reach the standard.
• Skill movement: change in assessed skill levels, checked without AI assistance.
• Business results: the metric the training was meant to move, such as resolution quality or error rates.
• Manager follow-through: whether development conversations are actually happening.
• Fairness: whether recommendations and access differ across groups for reasons unrelated to the job.

Common Pitfalls

• Treating completion as learning. AI can raise completions by nudging. That says nothing about whether anyone can do the task.
• Letting the tool do the thinking. The high school math experiment above is the reminder: assistance can lift performance in the moment while learning stalls.
• Using inferred skill scores for high-stakes decisions. A model's estimate of someone's skill is a development prompt. Promotion and pay decisions need evidence and human judgment, and may trigger legal obligations.
• Skipping the skills framework. Without clear definitions, "personalized" means "recommended by popularity."
• Publishing unreviewed content. Generated training can contain confident errors. In safety and compliance, one wrong step has consequences.
• Ignoring bias. Recommendation systems can steer certain groups toward lower-visibility opportunities if the underlying data reflects past patterns. Test outcomes by group.

Where This Fits in the Bigger Picture

L&D touches the whole employee lifecycle, from onboarding to succession. If you're mapping where AI belongs across HR, AI across the employee lifecycle shows where learning sits next to hiring, engagement and retention.