Using AI to Design Incentive Compensation Plans: Power, Pitfalls, and Practical Reality
A new HR employee at a long-standing Prosperio client recently called asking for advice on how to design an incentive plan for their carrier sales role. Of course, I first asked if she had a copy of Taming the Compensation Monster - and she did. She said there were several copies around the office! Then we talked about some of the important nuances in compensation design for this role – such as using GM$ instead of revenue, considering how to handle negative loads, how this role interacts with the customer sales side, etc. And I left her with a few ideas for putting together a simple GM% and Load Count Matrix, which she could either payout using a dollar per load (thereby avoiding a problem with negative loads), or as a commission against GM$ (but then she’d have to decide how to handle negative loads).
A few days later, she came back with a draft plan that she asked me to review, and to let her know if there were any problems that I saw with it. A quick perusal made it evident that she’d plugged her notes from our conversation into an AI program and had copied out what it told her to do and sent that to me. I’ve no doubt this is just the first in what will be many such interactions with clients. I don’t blame her for doing this, and I was curious what result the AI would give her. We all know (by now) that this burgeoning technology can be simultaneously very confident and very wrong. The recommendation it gave wasn’t so much wrong, as it was just…off. Plan weights were assigned to modifiers and the approach was not something that could have been administered (for example, it suggested taking a deduction for negative loads after the first 2 in a month – loads change value all the time!), and it said that all GM$ from negative loads should not be included in the GM% calculation which would only have the effect of artificially INFLATING the GM%.
I gave her my response, and a screen shot of a generic matrix that would require considerable tuning to get to where she wanted it to be. But it got me thinking, as certainly other HR professionals or small brokerage owners are going to want to take this path. If it saves money, of course they will! But there are risks that need to be carefully considered before skipping down this yellow brick road.
Before I go into what I see as the pros and the cons, you should understand where I am with my own AI usage. I started really using it in earnest in the summer of 2025 and I currently use both ChatGPT and Claude as an Excel add-in and for its new Cowork feature. Both are Professional subscriptions and I’ve found them to be tremendously useful for quick research on a prospective client, running an error check on a complex incentive calculator, helping me develop time-saving macros for things like expanding and collapsing a group of tabs, consolidated data from multiple files, etc. However, as we all should know by now, it ABSOLUTELY makes mistakes and so you need to check everything carefully and not just assume that it is correct (and some of the mistakes are just “weird” – no other word for it).
I have noticed that it fails miserably at “jumping to conclusions.” The app seems to have been programmed like an eager puppy ready to do whatever trick we can think to ask, but in so doing it often grabs the wrong bits of information or makes the wrong connections. I will use AI for market research (keeping up with M&A in this space can be a full-time job) and I’ve noticed that AI is so very eager to show me what it has learned about Prosperio, that it can confuse industry segments in a very annoying way – it doesn’t really (at least as of yet) understand the nuanced differences between freight broker, SaaS company serving the industry, 3PL, 4PL, etc. This eagerness to “sound smart” can be both dangerous and devaluing. When I notice that a blog was AI written (too many emojis, double lines breaking paragraphs, overuse of “That’s not XYZ, it is ABC” framing) then I immediately devalue the source. I suspect that employees will come to recognize the same when it’s coming from management and question, if nothing else, authenticity.
When it comes to compensation, AI will never be able to fully understand the ecosystem your incentive compensation plan must live within, which includes:
Manager and employee history with prior plan designs
Current business growth cycle stage
External economic conditions
Employee strengths and weaknesses
The myriad of options available for comp design beyond the “prompt” which may be far better solutions to the problem at hand
How to negotiate an acceptable plan among your leaders who have different expectations and needs from the design (and internal politics)
Trade off decisions between simplicity and complexity, individual and team, long-term and short-term focus
You should also be very careful that it understands fully what the role is as many organizations use different terms for the same role (for example a Business Developer could be new customer sales or Account Management, and don’t get me started on the word “Broker”). Just trust that what it means to YOU is very likely not what it means to everyone in this world.
There are some serious financial risks as well, as poorly designed plans can pay too much (or too little), may be easily gamed, may generate unintended consequences or not be flexible enough to allow for easy changes to swings in the market (this is a complexity trade off, as the simplest plan – a straight commission – is also the most inflexible).
No doubt the world is changing, and AI is likely here to stay. Prosperio is using it to improve OUTCOMES for our clients, such as helping speed the development of enhanced programming features for our calculators and as a final step in our QC process. But AI will not replace our own knowledge (and instincts) built on nearly 30 yrs of developing incentive plans for over 500 clients. In that time, we made, and learned from, many mistakes. The personal experiences we have watching companies evolve and grow over decades, isn’t replicable through AI.
