For instance, imagine you're a car manufacturer. Your big data sources and tools tell you that certain vehicle models have a flaw that may cost a few cents to repair on vehicles yet to be manufactured, but would cost significantly more to repair in vehicles that have already been purchased by customers and are in production use. The data, and thus your data scientists on staff, might recommend fixing the issue on cars still on the assembly line but not bothering to fix the cars already out there in the world, simply because the data might have shown the cost exceeded the likelihood of damages across the board.
(Note that this scenario may sound familiar to you if you have been following the General Motors ignition switch saga. However, this is only a hypothetical example, and further, there is no evidence big data played into the GM recall.)
Say your company has a value statement that quality is job 1 and safety is of paramount importance. Though the data suggests a recall isn't worth it, you make the call as an executive to start the recall. You're informed, but you're not controlled by big data.
Above all, it's vital to remember that sometimes the right answer appears to be the wrong one when viewed through a different lens. Make sure you use the right lens.
Big Data Can't Solve Non-Quantifiable Problems
Behold the old saying: When you're a hammer, everything looks like a nail. Once you begin having some success using big data to predict and solve business problems, there will inevitably be a temptation to "ask the data" every time you have an issue or an item about which a resolution is unclear.
As mentioned before, data can present you with more and better choices and, perhaps, make clear what may happen with each of those choices. Sometimes, though, data is no good at all — and that's when it's used with individuals.
Why? It's nearly impossible to quantify an individual's behavior. People have their own sets of circumstances, their own little universes, their own reasons and contexts. It's impossible to apply math to a single individual. Rather, you have to look at a group of individuals, a cohort of subjects with similar characteristics. Only then can you observe the trends of behavior that apply to the whole group.
This actually isn't a big data problem. It's a statistical problem. The easiest example that comes to mind is credit scoring, which purports to break consumers into groups and analyze the repayment and borrowing history of the individuals in each group in the aggregate.
If someone has, say, a 720 credit score, what that score actually means is that their repayment history puts them into a like statistical group — X percent (depending on which particular credit score, and which variant of that credit score, you look at to determine the actual percentage) of the persons in that grouping of individuals (in other words, a percentage of borrowers that had a score in that range) went on to either become seriously delinquent or actually entered default.