Skills Gap Analysis With AI: A Step-by-Step Method

Somewhere in your shared drive there’s probably a skills audit. It lives in a spreadsheet that took a quarter to build. Half the people in it have since changed roles, and nobody has opened it since the day it was presented. Now the business wants to launch something new, and someone in a leadership meeting asks whether your people have the skills to deliver it. That spreadsheet can’t answer the question anymore.

That’s the normal fate of a skills gap analysis. The method isn’t hard, but the data goes stale faster than a manual process can refresh it. AI can help with that, because it can read signals that already exist in your systems and keep the picture current. It can also guess wrong with great confidence. So this guide does two things. It gives you a six-step method you can run, with a filled-in example. It also shows you how to check the AI before you act on anything it tells you.

Before you start a skills gap analysis: data, privacy and scope

My advice is to check your data first, because a gap analysis can fail before the first step when nobody knows what they have to work with. A short check at the start saves you from building a matrix on sand.

Check the data you already have

AI can only infer skills from records that exist. Before you buy anything or build anything, find out whether these are in place:

• An up-to-date list of roles that maps each person to one role
• Job descriptions or role profiles that someone has reviewed in the last year
• Learning records that show completions and dates
• Performance reviews or manager notes kept in a system rather than in email
• Some record of what people work on, such as project assignments or tickets.

If two or more of these are missing, fix that first. Our HR data readiness guide walks through how to check each one.

Tell people what you’re doing, and collect less

When AI infers skills from work signals, it is processing personal data about employees. In many places that also counts as monitoring. So the first rule is to tell people before you start. Say what you’ll look at, why you’re looking at it and who will see the results. You should also give people a way to see their own profile and correct it. A skills profile that the person can’t see is the fastest way to lose trust in the whole exercise.

Collect less than the tools allow. You don’t need private messages, calendar content or keystroke data to know if someone can run a renewal negotiation. Set a retention period for the inferred profiles and delete old versions when you re-run the analysis.

If you operate in the EU, two parts of the EU AI Act matter here. Since 2 February 2025, Article 4 has asked organisations that use AI systems to take measures on their staff’s AI literacy. That includes the HR team running this analysis. Separately, Annex III of the Act lists AI used to evaluate workers or to make promotion and termination decisions as high-risk, with much heavier duties. That’s one reason this guide keeps the gap score away from those decisions. Our post on AI governance in HR covers the wider picture.

Pick a small first scope

Don’t start with the whole company. Pick one or two roles where three things are true. The role matters to a business goal this year. It has at least 10 people in it, so the numbers mean something. The managers also have to be willing to spend an hour checking results. A first run on one team teaches you more than a company-wide survey that nobody finishes.

The six steps of an AI skills gap analysis

Here is the method in order. Each step says where AI helps and where it shouldn’t be trusted.

Step 1: Define the skills taxonomy and role profiles

A skills taxonomy is simply an agreed list of skills, with a name and a short definition for each one. A role profile then says which of those skills a role needs, and at what level. This is the step I’d spend the most time on, because it’s where I see analyses go wrong. Teams copy skills straight out of job descriptions, and job descriptions are often wish lists written for hiring. If the target is wrong, every gap you calculate later is wrong too.

• Start from a public list such as ESCO, so you aren’t inventing names from scratch
• Keep it short. Six to ten skills per role is enough for a first run
• Use a 1 to 5 scale, and write what each level looks like in behaviour
• Ask two or three strong performers in the role what the job really needs

AI is useful for a first draft of the list. A human must set the required levels, because that’s a business decision about what good looks like.

Step 2: Collect evidence on current skills

Next you gather what you know about each person’s current level. Manatal’s guide suggests rating each skill from three sources: the person, their manager and an objective test. That’s a sensible default, because each source corrects the others. Manatal notes that self-ratings tend to run high. Managers can also rate the people they like more highly. A test or a work sample is harder to argue with, so it settles the disagreements.

AI isn’t used in this step. The job here is deciding which evidence you trust, and that’s a human call.

Step 3: Infer skills with AI

Skills inference means the AI reads records such as work output, learning history and project roles, then estimates each person’s level on each skill. This is where AI earns its place. It can do in an afternoon what would take your team weeks of interviews. It also keeps working, so the next run doesn’t start from zero.

• Feed it only the sources you told people about in the privacy step
• Ask for a confidence level with every estimate, not just a score
• Keep the inferred score separate from the self and manager ratings

Step 4: Validate before anything else happens

Treat every AI score as a draft until a person has checked a sample of them. The full check is described in its own section further down. The short version: managers review about 20 profiles, and you count how often the AI was wrong. If the miss rate is high, you stop and fix the inputs. No AI is used in this step, because the whole point is a human check on the AI.

Step 5: Calculate and rank the gaps

The gap is the required level minus the current level. That part is arithmetic. The judgment comes in the ranking. A large gap in a skill that barely affects results matters less than a small gap in a skill tied to revenue. So score each skill for business impact as well, and rank by both.

AI can do the sums and sort the table. People must set the impact scores, because only the business knows what hurts.

Step 6: Decide what to do, and set the re-run date

For each gap you have three options. You can train the people you have. You can redeploy someone who already has the skill from another team. Or you can hire. The choice depends on how big the gap is, how many people have it and how quickly you need it closed.

Finally, set a re-run date and write down what “closed” means. Manatal suggests a gap is closed when the person meets the target and holds it for one review cycle. That’s a good test, because it rules out a skill that spiked after a course and then faded. My advice is to re-run every six months for the roles in scope. That’s our judgment, not a research finding, but it fits most review cycles.

The six steps at a glance (copyable template)

Step	Input	Who does it	AI used?
1. Define taxonomy and role profiles	Public skills list, top performers, business goals	HR with role managers	Yes, for a first draft only
2. Collect evidence	Self ratings, manager ratings, work samples	Employees and managers	No
3. Infer skills	Work output, learning records, project data	AI tool, set up by HR	Yes
4. Validate	Sample of about 20 profiles	Managers	No
5. Calculate and rank gaps	Validated scores, impact scores	HR, with AI doing the sums	Partly
6. Decide and set the re-run	Ranked gaps, budget, hiring plans	HR and business leaders	No

A worked example: one team, six skills

This is a composite made up for this guide, and the figures are invented. Claire is Head of L&D at a software company of about 1,200 people. The company is about to release an analytics product, and the 14 customer success managers will have to sell renewals that include it. Her CEO wants to know if the team is ready by the next quarter.

Claire ran the six steps for that one role. Here is the skills matrix she ended with. The current level is the team average after validation, on the 1 to 5 scale.

Skill	Required level	Current level	Gap	Business impact	Action
Renewal negotiation	4	2.6	1.4	High	Train
Technical setup for customers	3	1.2	1.8	Medium	Hire one specialist
Reading account data	3	1.8	1.2	Medium	Redeploy and train
Knowledge of the new product	4	3.5	0.5	High	Short course
Handling escalations	4	3.8	0.2	High	No action
Pricing and contract terms	3	2.9	0.1	Low	No action

The biggest gap is technical setup, but Claire didn’t put it first for training. Teaching 14 people a technical skill they’d use a few times a quarter would cost a lot and fade fast. So she recommended hiring one specialist that the whole team can call on. That’s the kind of call the table can’t make for you. It shows the size of a gap, but not whether everyone needs to close it.

Renewal negotiation became her top training priority. The gap is large and it sits right on revenue. When she looked behind the average, she also found the gap wasn’t spread evenly. Four people already scored at level 4, and the other ten were around 2. That changed the plan. Instead of buying a course for everyone, she paired each of the ten with one of the four for live renewal calls. Averages hide this kind of spread, so always look at the individual scores before you choose an action.

The gap in reading account data had a different fix. She found two analysts in the operations team who already worked with the same dashboards. One of them wanted a customer-facing role. Moving that person into the team, and having them run short sessions for the others, closed most of the gap without a new hire.

The last two rows needed no action. A gap of 0.2 or 0.1 is within the noise of any rating method. Chasing it would waste the team’s time. Claire set the re-run for six months later, with “closed” meaning the team average meets the target and holds through one review cycle.

Where AI gets skills inference wrong, and how to check it

AI-inferred skills scores look precise, which makes them easy to believe. But each score is only as good as the data behind it, and every source has blind spots.

What each data source is good and bad at

Data source	What it shows well	Where it misleads
Work output (documents, tickets, CRM notes)	What people actually do, week to week	Rewards people who write a lot. Misses work done in calls or in person
Learning records	What people have studied, and when	A finished course isn’t a skill. Completion often means a quiz was passed
Project data	Which roles people held on real work	Being on a project doesn’t mean doing the skilled part of it
Performance notes	A manager’s view of results over time	Inherits manager bias. Notes are often vague or missing

The pattern to watch is that AI over-credits whatever is easy to see. Someone who documents everything looks more skilled than someone who quietly solves problems on the phone. That’s not the AI being careless. It’s the AI being faithful to uneven data.

The validation check

Run this before any score goes into a decision or a report.

1. Pick about 20 people from the roles in scope. Include a mix of strong and weaker performers, and people from different teams.
2. Show each manager the AI’s scores for their people, without the person’s self rating next to it.
3. Ask the manager to mark each skill score as right, too high or too low.
4. Count the misses. A miss is any score the manager marks as off by one level or more.
5. Look for patterns in the misses. If one skill or one data source is behind most of them, fix that source of data and run again.

Decide your acceptable miss rate before you look at the results, so you aren’t tempted to move the line. Our judgment is that the inference isn’t ready to use if managers disagree with more than one score in four. Managers can be wrong too, of course. If a manager and the AI disagree, a short work sample settles it better than either opinion.

Bias, and where a gap score must stop

The AI can carry bias in from its data. People who work part time, people on leave and people whose work is less visible tend to get lower scores. So check your validated results by group. If one group scores lower across the board, look at the inputs before you accept that the gap is real.

A gap score is a planning tool for training and staffing. It should never be the reason someone is not promoted, put on a performance plan or let go. Those decisions need direct evidence about that person, reviewed by people who know their work. Using AI that way can also put the system in the EU AI Act’s high-risk category, with duties that a planning exercise doesn’t carry.

When not to use AI for this

AI isn’t always the right tool. In these cases a manual method will give you a better answer for less effort.

• Small teams. If the role has fewer than about 10 people, a one-hour session with the manager and a simple spreadsheet will be more accurate than any software.
• No usable data. If work happens mostly in meetings and the learning system is empty, there’s nothing for the AI to read. Use self and manager ratings plus a work sample instead.
• High-stakes decisions about individuals. Promotion, redundancy and performance cases need direct evidence and structured interviews. Keep AI out of it.
• A brand-new role. If nobody has done the job yet, there’s no history to infer from. Define the profile with the business and assess against it by hand.

What it costs and how long it takes

We don’t have sourced figures for the cost of a skills gap analysis, and none of the guides we reviewed publish any. So rather than invent a number, here’s what drives the effort. The biggest driver is the state of your data. If your role list is clean and your records are in systems, setup is mostly switching things on. If they’re scattered across spreadsheets, the cleanup can take longer than the analysis.

The second driver is manager time. Steps 2 and 4 both need managers, and their hours are the scarcest input you have. Plan for that before you start, and get their leaders to agree to it. The third driver is whether you buy a tool or build something on your own systems. A tool brings its own skills list and data connections. A custom build fits your roles and data more closely, but someone has to build and maintain it.

What we saw building Risely’s skills model

We built Risely, our own AI coaching product. Its coach is called Merlin. It works inside Slack and Microsoft Teams in 40 languages, and it has coached more than 5,000 people. Underneath it is a defined set of 83 workplace skills across two tracks, one for managers and one for individual contributors. Risely reports an average 26% skill improvement in 12 weeks.

The lesson that carries over to gap analysis is about the list itself. Coaching only works when every skill has a clear name and a clear meaning, so that progress on it can be tracked. The same is true of a gap analysis. Splitting the list into separate manager and individual tracks also mattered, because the same skill name can mean different things at different levels. A gap analysis that uses one list for everyone hides those differences. If you take one thing from this guide into step 1, make it this: spend more time on the skills list than feels necessary.

Next step: a tool, a custom build, or a call

If your data is in good shape and your roles are fairly standard, a skills platform may be the quickest route. Our guide to skills intelligence software compares the main options. If your roles are unusual, or you want the analysis to run inside the systems you already use, a custom agent may fit better. You can see what that looks like in the learning and development area of our HR Agent Map, which includes a skills gap analysis agent. We also wrote about how to weigh HR AI software against custom agents if you’re unsure which way to go.

If you’d like to talk it through for your own roles, book a 30-minute call with us. We’ll look at what data you have and tell you which route we’d take.

Frequently asked questions