You can assess this without being technical
Comparing machine learning proposals feels like it requires expertise you do not have. Mostly it does not - the strongest signals are about reasoning and completeness rather than technique.
A proposal that is vague about what it will predict and precise about which algorithms it will use has the emphasis backwards, and you do not need to know the algorithms to notice.
The absences that matter
- No baseline. If nobody proposes measuring the current method, there is no way to prove improvement later.
- No data assessment stage. A supplier confident without looking at your data is either very experienced with an identical problem or not thinking carefully.
- Nothing after launch. No monitoring, retraining or support means the model degrades quietly and the cost arrives later.
- No risks section. Every project of this kind has real risks; a proposal with none has not examined them.
- No mention of who uses the output. Adoption decides whether it produces value.
Accuracy promised in advance
A supplier guaranteeing a specific accuracy before seeing your data is making a claim they cannot support. Achievable accuracy depends on your data, and nobody knows it yet.
The reasonable version is a commitment to measure honestly against a baseline and to report what is achieved, with a decision point if it falls short. That is a supplier managing risk with you; the guarantee is one transferring it to a number they will later reinterpret.
Solution-first proposals
| Proposal says | Concern |
|---|---|
| We will build a deep learning model | Technique chosen before the problem was understood |
| Our platform handles this | Product being sold, problem being fitted to it |
| Similar to what we did for another client | Reasonable if the problem matches; check that it does |
| We will start with the simplest approach that could work | Good sign |
| We may find machine learning is not the answer | Very good sign |
Questions that separate proposals quickly
- What would make you recommend we do not proceed?
- What is the current process, and how well does it perform today?
- What happens to this model in eighteen months?
- Who in our business needs to use the output, and how will it reach them?
- What data would change your estimate most if it turned out to be poor?
The first question is the most revealing. A supplier with a considered answer has thought about failure, which is the best available evidence that they have thought about the project properly.
A proposal with no risks section has not looked for any.