I want to start with a line from Josh Bersin: compensation is one of the most powerful and least well understood communication functions in a company. I agree completely. Compensation moves money, and money drives motivation, the decision to stay or leave, and whether a person feels valued at all. Yet very few businesses use that power deliberately.
For a long time the reason was simple. To run compensation properly you either built a dedicated department or hired a consulting firm. Job evaluation methodologies cost money, salary surveys cost money, and the expert who could assemble all of it into a working system cost even more. Companies without a budget simply lived without grading, without salary ranges, without transparent rules — and collected the full set of consequences.
Today that has changed radically. Artificial intelligence knows every major job evaluation methodology, handles large data sets, holds your company context in mind, and delivers output at a consultant’s level. I spent years running compensation and benefits at the STB television channel and at Wargaming: grading, salary market surveys, benefits design, EVP audits, payroll budgeting and salary reviews, bonus scheme design. I can say honestly that work which used to take me weeks now takes hours. In this article I break down the first and most fundamental block: job architecture and grading with AI.
For a long time the reason was simple. To run compensation properly you either built a dedicated department or hired a consulting firm. Job evaluation methodologies cost money, salary surveys cost money, and the expert who could assemble all of it into a working system cost even more. Companies without a budget simply lived without grading, without salary ranges, without transparent rules — and collected the full set of consequences.
Today that has changed radically. Artificial intelligence knows every major job evaluation methodology, handles large data sets, holds your company context in mind, and delivers output at a consultant’s level. I spent years running compensation and benefits at the STB television channel and at Wargaming: grading, salary market surveys, benefits design, EVP audits, payroll budgeting and salary reviews, bonus scheme design. I can say honestly that work which used to take me weeks now takes hours. In this article I break down the first and most fundamental block: job architecture and grading with AI.
Why C&B got stuck in Excel and why that no longer works
The Compensation & Benefits function has historically been married to Excel. Salaries lived there, analytics were built there, review cycles were planned there, and grading lived there too. Anyone who has worked in Total Rewards remembers the endless VLOOKUPs, INDEX formulas and pivot tables that break the moment you add a column, plus the feeling that half your working time goes into making an error message disappear.
I am not telling anyone to abandon Excel. It remains an excellent carrier of information and a convenient way to store and share it. But living only in Excel belongs to yesterday. Artificial intelligence has to join the spreadsheets, because it takes over the thinking, not just the arithmetic.
There is another effect I have felt personally. Once you know a task can be done with AI at least as well as by hand, and you still do it by hand, you physically feel yourself wasting time. For a 50-person company you can still imagine life without AI. For a large company there is no choice left.
510 hours a year: what manual Total Rewards actually costs
Let us count how much time the C&B function eats over a year if everything is done manually. In my estimate, a full cycle of typical tasks is about 510 hours a year — roughly three months of one person working full-time.
And note this is not even the full list. It excludes designing bonus schemes, analysing how those schemes actually perform, reviewing off-cycle salary adjustments made over the year, and preparing conclusions and recommendations. Add those and the count comfortably passes 510 hours.
The Total Rewards Statement line deserves separate attention. Those 250 hours are half of the function’s entire time budget, spent on a document most companies never produce precisely because “there is no resource for it”.
The Total Rewards Statement line deserves separate attention. Those 250 hours are half of the function’s entire time budget, spent on a document most companies never produce precisely because “there is no resource for it”.
What a company loses without a mature Total Rewards function
The absence of a system is not a neutral state. It is a set of concrete losses I see in companies constantly:
•A pay gap of 20% or more. This is the difference in pay between men and women, on average across the industry, after controlling for job level and tenure. In other words it is not explained away by “they hold different positions” — it is a clean gap.
•One and a half times higher turnover of top performers compared with companies where Total Rewards is transparent and where employees understand what their package is made of.
•62% of salary reviews fail. After a badly run review, employees say they are undervalued and leave. Often they leave not because of the number but because of how the process was run and explained.
So the choice is simple: live with these losses, hire a large department and consultants, or learn to use artificial intelligence.
•A pay gap of 20% or more. This is the difference in pay between men and women, on average across the industry, after controlling for job level and tenure. In other words it is not explained away by “they hold different positions” — it is a clean gap.
•One and a half times higher turnover of top performers compared with companies where Total Rewards is transparent and where employees understand what their package is made of.
•62% of salary reviews fail. After a badly run review, employees say they are undervalued and leave. Often they leave not because of the number but because of how the process was run and explained.
So the choice is simple: live with these losses, hire a large department and consultants, or learn to use artificial intelligence.
25 hours instead of 510: the function on AI
When the same tasks are moved onto artificial intelligence, the picture becomes unrecognisable.
The point is not the saved hours in themselves, but what those hours convert into. When pay equity is calculated on demand rather than once a year, you can surface a problem before it becomes a legal and reputational case. When 500 personalised statements take an hour, they stop being a privilege of the executive team and start existing for everyone.
The TRACE framework: five maturity levels of Total Rewards
The AI for C&Band TotalRewards.From analyst — to partner course is built around a model I called TRACE. The logic is that each level rests on the previous one, and you cannot skip a step.
Let me unpack the logic. If you have no job architecture, benchmarking breaks immediately: you compare your senior people against something undefined, get a market median below what you actually pay, and have no idea what to do with it. Next comes market data, sourced from paid provider surveys, open sources, and what the model already knows. There is an important nuance here: AI is trained on such a vast body of material that some of what it knows is not findable in the open internet at all, and some of it sits in corners you would never think to look into.
Then comes salary review, then cost modelling — because you cannot set ranges or plan increases without understanding how they turn into a budget. And the end of the cycle is Total Rewards Experience: making rewards reflect the needs of different people so they actually change behaviour. That last part was practically out of reach before AI.
Then comes salary review, then cost modelling — because you cannot set ranges or plan increases without understanding how they turn into a budget. And the end of the cycle is Total Rewards Experience: making rewards reflect the needs of different people so they actually change behaviour. That last part was practically out of reach before AI.
Level T: what job architecture actually is
Job architecture is the systematic organisation of every position in the company inside a single structure. It has three layers.
Job families
Groups of roles by type of activity: engineering, product, design, sales, marketing, HR, finance, operations or production. The set depends on the industry — a product company will have one list, a manufacturer another.
Career levels
Levels inside each family, consistent across the whole company. For example IC1 through IC8 for individual contributors and M3 to M7 for managers. Consistency is the key word: level 5 in engineering and level 5 in finance should mean comparable complexity and impact.
Job profiles
A specific role, in a specific family, at a specific level, with its own description, scope and impact. A senior backend engineer at level 5 is already a job profile.
Many companies additionally split functional tracks: individual contributors on one side, managers on the other. This is the base of the entire Total Rewards cycle. Building it used to take months and required consultants who brought their own methodology and either matched your jobs to their catalogue or built a bespoke structure for you.
The dual career ladder in technology companies
Technology companies commonly run an eight-level model where the expert and management tracks sit in parallel. The idea is that at any point an employee can choose to go deeper into expertise or into people management, and neither path is considered higher than the other.
Data providers such as Radford use architectures of this shape. The important part: the model is well known to artificial intelligence, so building an equivalent structure for your own company is now a one-day job.
Job evaluation methodologies that Claude knows
Methodologies exist to evaluate a job, distribute positions into grades, and then attach salary ranges to those grades. Here are the main ones used worldwide.
There are also local and regional methodologies from less known providers, and they are perfectly usable if they are established in your market.
How the Hay methodology works
When you evaluate a job under Hay you assess three things:
1.The level of know-how required for the position.
2.The level of problem solving — how complex and unstructured the challenges the role handles are.
3.Accountability — what the role influences and at what scale.
Each factor earns points, the points are summed, and the total produces a grade and a career level.
When you evaluate a job under Hay you assess three things:
1.The level of know-how required for the position.
2.The level of problem solving — how complex and unstructured the challenges the role handles are.
3.Accountability — what the role influences and at what scale.
Each factor earns points, the points are summed, and the total produces a grade and a career level.
My own experience: $10,000 for a methodology and a two-day training
When I worked at the STB television channel we wanted to bring order to our job structure. There were two options: hire Hay, which was expensive, or find a way around it. This was right after the global financial crisis of 2008–2009, when a lot of laid-off specialists were on the market, including consultants in grading and job architecture.
We took the pragmatic route: instead of the firm, we hired consultants who used to work there. They delivered a two-day training for us, advised us a little, and after that we evaluated jobs ourselves. Even in that economical version we paid around $10,000.
I am not talking anyone out of hiring consultants today — if the budget exists, it is a good option. But if there is no budget, you now have an alternative. Claude knows the methodologies: the principles, the factors, the point logic. All you need to do is load your data, frame the task properly and understand what result you are expecting.
Another live example is Preply, today a technology unicorn. When it was small, building job architecture, salary ranges and evaluating positions took a great deal of time — part of it done by me, part by my people operations manager. The same volume of work now takes a fraction of the time, and the accuracy is no lower.
Why Claude rather than ChatGPT for C&B work
I registered with ChatGPT the moment it appeared and used it for several years. Then I moved to Claude, though I still use ChatGPT occasionally. Here is the comparison on the parameters that matter for Total Rewards.
On simple tasks the difference is barely visible: ask both tools to write a job advert and the outputs will be comparable. The gap appears on complex tasks where a lot of data and nuance must be held at once.
Building the “C&B Job Architecture” Claude project in 90 minutes
A project in Claude is a workspace holding your knowledge base, your instructions and all the chats on the topic. In effect you are creating a personal AI agent, an expert in Total Rewards.
1.Create the project. In the left menu, Projects → New Project. Name it, for example, Job Architecture.
2.Write the system prompt (Project Instructions). This defines the agent’s role, expertise, working principles and constraints.
3.Load the knowledge base. These are the files the agent will rely on.
4.Set the output style and format. How you want answers delivered: a table, an evaluation card, a short summary.
5.Test and tune. Run a few positions and compare the output with your own judgement.
A useful detail: whenever you discuss something valuable inside a project, Claude stores it in the project memory and reuses it in other chats. Insights carry over from one conversation to the next.
What to load into the knowledge base
•Methodologies, if you have them. Hay or Mercer documents are their intellectual property, so I cannot distribute them within the AI for C&Band TotalRewards.From analyst — to partner course and can only refer to them. If you have nothing, that is fine — the model is trained on these methodologies, and I have tested that across different scenarios.
•Your job descriptions — at least 10 to 20 of them. The more you load, the better the model understands your specifics and the way you describe roles.
•The organisational structure — the list of roles, departments and reporting relationships. Surnames can be left out: that is a question of confidentiality and GDPR requirements if you are in Europe. The decision about personal data always stays with you.
•A salary survey, if you have one. If not, it is not critical — open sources are covered in the second session.
•Internal specifics: your naming conventions, terms, established labels. If none of it is documented, simply type or dictate it and save it into the project knowledge.
The structure of a job evaluation prompt
A good C&B prompt rests on four elements:
1.Role — how the agent behaves and who it is in this conversation.
2.Context — company description, industry, specifics. This can live in the knowledge base.
3.Instruction — exactly what you are evaluating and under which methodology.
4.Answer structure — the format in which you want the result.
Every typical task builds on that frame: calculate Hay points for a senior backend engineer and produce copy-ready text, evaluate ten roles in a single prompt, run a calibration, generate a list of job families for your company, or build a career architecture complete with behavioural indicators.
A seven-day cycle to bring the project to working condition
The first result is almost never perfect, and that is normal. Here is the sequence that gets a project into working shape:
1.Run five roles at different levels — different on purpose, so you can see how the model distinguishes complexity.
2.Compare the evaluations with how you would grade those positions yourself and with the data you already have.
3.Fix the instructions wherever the logic diverged from yours.
4.Add the missing context to the knowledge base.
5.Run difficult, borderline cases — that is where the quality of the setup shows.
6.Test with your team and let colleagues check the output on their own roles.
7.Freeze the final version as the project’s working standard.
One technique I use constantly: when the model produces something other than what I expected, I ask it for a retrospective. “Why did you arrive at this?” It explains. “What needs to change so the result looks like this instead?” It formulates the answer. I add that answer to the project knowledge, and the next run comes out correct.
Tools and templates
Beyond prompts, ready-made files do a lot of the lifting. The evaluation file is straightforward: the first tab holds instructions, then come the methodology tables, the mapping, and the job card template where you select options and the points roll up into a total score, a hay grade and a career level. There is a filled example for a senior backend engineer and a mass comparison sheet for large numbers of positions.
You can work with these files in two ways: fill them in by hand, or hand the template to Claude together with your job descriptions and ask it to complete the file for your company.
If you are looking at ready-made paid solutions, the market offers Pave, HR Signal and beqom. They are expensive, and my argument always favours the “Claude plus templates” combination: the result is no worse and the cost is incomparably lower.
What to do next
1.Gather the source material. Export 10 to 20 job descriptions and your organisational structure without personal data.
2.Create a Claude project named, for example, C&B Job Architecture, and write a system prompt casting the agent as a Total Rewards expert.
3.Define job families and career levels for your company: how many IC levels, how many management levels, and where they align.
4.Run five different roles through evaluation under your chosen methodology and compare with your own expert judgement.
5.Run a retrospective on the gaps together with the model and add the conclusions to the project knowledge base.
6.Assemble the full architecture — families, levels, job profiles — and lock it into the template.
7.Set up metrics to track the quality of your grading and job architecture management over time.
1.Create the project. In the left menu, Projects → New Project. Name it, for example, Job Architecture.
2.Write the system prompt (Project Instructions). This defines the agent’s role, expertise, working principles and constraints.
3.Load the knowledge base. These are the files the agent will rely on.
4.Set the output style and format. How you want answers delivered: a table, an evaluation card, a short summary.
5.Test and tune. Run a few positions and compare the output with your own judgement.
A useful detail: whenever you discuss something valuable inside a project, Claude stores it in the project memory and reuses it in other chats. Insights carry over from one conversation to the next.
What to load into the knowledge base
•Methodologies, if you have them. Hay or Mercer documents are their intellectual property, so I cannot distribute them within the AI for C&Band TotalRewards.From analyst — to partner course and can only refer to them. If you have nothing, that is fine — the model is trained on these methodologies, and I have tested that across different scenarios.
•Your job descriptions — at least 10 to 20 of them. The more you load, the better the model understands your specifics and the way you describe roles.
•The organisational structure — the list of roles, departments and reporting relationships. Surnames can be left out: that is a question of confidentiality and GDPR requirements if you are in Europe. The decision about personal data always stays with you.
•A salary survey, if you have one. If not, it is not critical — open sources are covered in the second session.
•Internal specifics: your naming conventions, terms, established labels. If none of it is documented, simply type or dictate it and save it into the project knowledge.
The structure of a job evaluation prompt
A good C&B prompt rests on four elements:
1.Role — how the agent behaves and who it is in this conversation.
2.Context — company description, industry, specifics. This can live in the knowledge base.
3.Instruction — exactly what you are evaluating and under which methodology.
4.Answer structure — the format in which you want the result.
Every typical task builds on that frame: calculate Hay points for a senior backend engineer and produce copy-ready text, evaluate ten roles in a single prompt, run a calibration, generate a list of job families for your company, or build a career architecture complete with behavioural indicators.
A seven-day cycle to bring the project to working condition
The first result is almost never perfect, and that is normal. Here is the sequence that gets a project into working shape:
1.Run five roles at different levels — different on purpose, so you can see how the model distinguishes complexity.
2.Compare the evaluations with how you would grade those positions yourself and with the data you already have.
3.Fix the instructions wherever the logic diverged from yours.
4.Add the missing context to the knowledge base.
5.Run difficult, borderline cases — that is where the quality of the setup shows.
6.Test with your team and let colleagues check the output on their own roles.
7.Freeze the final version as the project’s working standard.
One technique I use constantly: when the model produces something other than what I expected, I ask it for a retrospective. “Why did you arrive at this?” It explains. “What needs to change so the result looks like this instead?” It formulates the answer. I add that answer to the project knowledge, and the next run comes out correct.
Tools and templates
Beyond prompts, ready-made files do a lot of the lifting. The evaluation file is straightforward: the first tab holds instructions, then come the methodology tables, the mapping, and the job card template where you select options and the points roll up into a total score, a hay grade and a career level. There is a filled example for a senior backend engineer and a mass comparison sheet for large numbers of positions.
You can work with these files in two ways: fill them in by hand, or hand the template to Claude together with your job descriptions and ask it to complete the file for your company.
If you are looking at ready-made paid solutions, the market offers Pave, HR Signal and beqom. They are expensive, and my argument always favours the “Claude plus templates” combination: the result is no worse and the cost is incomparably lower.
What to do next
1.Gather the source material. Export 10 to 20 job descriptions and your organisational structure without personal data.
2.Create a Claude project named, for example, C&B Job Architecture, and write a system prompt casting the agent as a Total Rewards expert.
3.Define job families and career levels for your company: how many IC levels, how many management levels, and where they align.
4.Run five different roles through evaluation under your chosen methodology and compare with your own expert judgement.
5.Run a retrospective on the gaps together with the model and add the conclusions to the project knowledge base.
6.Assemble the full architecture — families, levels, job profiles — and lock it into the template.
7.Set up metrics to track the quality of your grading and job architecture management over time.
The 10/20/70 formula: how much expertise you actually need
Previously, to work in Total Rewards you needed close to one hundred percent expertise — the field is complex, methodology-heavy, full of nuance. The proportion is different now:
1.10% — your own C&B and Total Rewards knowledge. Enough to understand the logic of the field and what result you want at the end.
2.20% — your ability to work with artificial intelligence. Prompting, projects, knowledge bases, and knowing how to verify and tune the model.
3.70% — the work the AI does for you. It knows the methodologies, remembers market practice, calculates, structures and formats.
That is exactly why the AI for C&Band TotalRewards.From analyst — to partner course works for beginners and experienced specialists alike. Newcomers pick up the Total Rewards discipline along the way, while veterans immediately see how to move familiar processes onto new tools.
✅ The AI for C&Band TotalRewards.From analyst — to partner course uses Excel templates: a job architecture builder, a job evaluation tool based on the Hay methodology, and a usage guide.
1.10% — your own C&B and Total Rewards knowledge. Enough to understand the logic of the field and what result you want at the end.
2.20% — your ability to work with artificial intelligence. Prompting, projects, knowledge bases, and knowing how to verify and tune the model.
3.70% — the work the AI does for you. It knows the methodologies, remembers market practice, calculates, structures and formats.
That is exactly why the AI for C&Band TotalRewards.From analyst — to partner course works for beginners and experienced specialists alike. Newcomers pick up the Total Rewards discipline along the way, while veterans immediately see how to move familiar processes onto new tools.
✅ The AI for C&Band TotalRewards.From analyst — to partner course uses Excel templates: a job architecture builder, a job evaluation tool based on the Hay methodology, and a usage guide.
