An individual asking why they keep stalling on price, and a sales director asking which of their fifty people to coach first, are not the same question. So we stopped trying to answer them with the same build.
Welcome to post 2 in the weekly series focused on building MyAcumen from the ground up using AI-First principles. The decisions we made, why we made them, and how we went about it. Aimed wider than my usual sales audience.
One chat box over all the data sounds efficient and we could have implemented this much ore easily, saving lots of time. … but in practice it gives everybody slightly worse answers than they should get. The model is handed either too little context to be specific, or so much data that it can’t tell which part matters. Context isn’t a detail of an AI build. It is the build.
Take the individual first. One person, around 500 data points across four assessments and their 360 feedback, and a set of questions that are nearly always about themselves. Good looks like this. They ask about business acumen (for example), and the answer names the sub-competencies and the specific answers in their own results that are impacting the score. Not their score read back to them in a friendly voice.
Now put fifty people through the same assessments. The questions change shape completely. Which role is weakest and why. Where two functions see the same thing differently. Whether the pattern in one team shows up in the other. A twenty-five person diagnostic alone runs to somewhere between 11,000 and 13,000 rows of raw responses. Throw all of that at a question about one competency and the answer is buried inside it.
So we made the decision to invest in two builds, and they work differently on purpose.
Individual Ask Mya pulls that person’s full record, works out what the question is about, and feeds the AI just the slice of data that fits it. Report pages load their own context first, so nobody has to explain which assessment they mean.
Project level Ask Mya covers one team diagnostic. The finished team report becomes the fixed context, and each question is a short tail on the end of it. Two things follow from that, both worth knowing if you’re building something similar. It forced us to make the report stand on its own, because a chat layer that keeps reaching for more context loses the benefit of holding that context steady. And every privacy rule is applied before anything reaches the AI, so what’s being asked about is already safe to answer from.
None of it worked first time. The early version went back to the database for extra context on every single question, which was slow, and the answers came back worse because the AI was reading a slightly different slice of the data each time you asked. Rebuilding the report so it carried everything Ask Mya needed was weeks of work that produced nothing a customer could see.
What it means is that now you get a straight answer. A seller asks why they are failing at a step in the sales process and gets the two sub-competencies and the limiting belief sitting behind it. A director asks which of their fifty to coach first and what on, and gets six names with a reason attached to each. All this in place of a fortnight trawling a spreadsheet.
Right context gives you two things at once. Answers that are more accurate, and answers you can act on. Wrong context gives you a confident answer about the wrong people, which is worse than no answer at all.
