With version 3 of our skill and resource management solution decídalo, we made a radical change. We replaced the skills catalog with a Large Language Model (LLM). For over 15 years, the skill catalog was a central part of decídalo. Now it is gone (mostly). Here is what we learned from migrating our customers to the AI world of skills.
The essentials in brief
1. No one misses the skills catalog, but some kind of structure is still needed.
2. Three skill levels are enough (in principle).
3. Interfaces to legacy systems require compromises.
What is a skills catalog?
A skills catalog is the list of all skills that are relevant for a company. Ok, what is a skill, and which ones are relevant? That is unclear. In practice, 90% of a skill catalog consists of products, technologies or tools. Since there are thousands of these, they are grouped in a folder structure – the skill tree.
The idea is to record skills in a structured form: skill plus skill level.
In theory, this simplifies reporting and analysis of capabilities.
In practice, skill catalogs are too detailed to be useful for reporting. Skills must be aggregated. Doing this manually is a daunting task.
What replaces the skills catalog in the AI age?
The simple answer: ChatGPT – or another Large Language Model (LLM).
LLMs have much more information about products, technologies or whatever is considered a skill than could ever be included in a skill catalog.
If the input and output are texts, you don’t need a skills catalog. But free text is not always the best choice.
Structured data is the better choice for reporting and analytics. But here you are not interested in thousands of skills, but in a set of maybe 50 to 100 core skills. Maintaining a company-specific list of core skills is feasible with reasonable effort.
The question remains how to determine core skills if your input is largely text-based – typically in the form of an employee profile. This is where the LLM helps again. In practice, various pieces of information such as project experience, certificates and training must be evaluated to obtain a reliable mapping of employee profiles to core skills.
Do you need a skill level?
It depends. In the previous decídalo versions, all skills had to be specified with level. This was too cumbersome for many companies.
No levels are needed to find people by skill. With AI and LLMs, you can use text information to determine how well someone fits a requirement.
When showcasing skills externally on consultant CVs, this is mostly done without levels. For the recipients of the CVs (our clients’ customers), skill levels are not verifiable. They only look at project experience
In decídalo V3, we therefore started without skill levels.
But some customers did have skill levels in their consultant CVs. So, we introduced a three-level scale as an option. Three turned out to be the maximum number of levels used on CVs.
The biggest problem was that our beautiful new AI solution had to integrate into old system landscapes. Some customers export skill data from decídalo to their HR systems. And they usually use five-level scales. For these cases, we eventually made five- or more-level scales possible again.
What we learned
In the age of LLMs, many use cases can do without structured skill data. Free text as input and output is more easily available and often more accessible to humans.
Structured skill data is still required for reporting. But here we only need a subset of 50 to 150 core skills. Lists of core skills can be easily generated with LLMs such as ChatGPT.
Core skills can be determined dynamically from employee profiles using AI. The advantage is that analyzing skills is decoupled from assessing skill this way. If core skills change, just run the AI mapping again. No need to bother employees with updating their skill profiles.
Our customers have quickly learned to appreciate the advantages of AI-based skill management. No one misses the old skill catalog.