General enquiries :
+44 (0)20 7602 6000

Why artificial intelligence and machine learning can't replace location planners

Friday 11 January 2019 Data Insight & AnalyticsRetail Consultancy


James Stevens's picture
By James Stevens

It seems every article I read is pointing to the imminent demise of this job or that career. The “rise of the machines” means we’ll soon all need to find something else to do with our time.  Nobody is safe from this sort of speculation –doctors, lawyers, artists and even location planners.  Well I just don’t buy it.

Yes, most jobs will change over the next ten or twenty years with advances in technology, just as they have over the last twenty years, but I don’t believe many of them will become obsolete.


So why can’t AI replace location planners? 

1. All models are NOT created equal

The process of building a strong predictive model is non-linear.  It requires judgement, experience, knowledge of the business and the industry.  It needs to be a collaborative process, where initial assumptions can be constantly questioned and revisited.

Being able to build a statistical model that fits the data using a standard software package is the easy part.  It’s all very well being a whizz with Python, SAS, Alteryx or R, but without real-world experience and ability, tenacity and drive to create the best possible solution for the client, you’re going to end up with something that falls short of the mark. 

Here is a common approach that I’ve seen many times
Throw all the data you have into a stats package, run a step-wise regression and create a brilliantly accurate model. 

This gives us a model that fits the data perfectly when we look backwards; unfortunately, it will be much less accurate for predicting future sites. 

I’m sure there are instances where regression modelling can work perfectly with minimal input required from a skilful practitioner, but this assumes a number of pre-requisites that are rarely true in our field.  Even if we are fortunate enough to have a large enough sample size (which we rarely are), there is an implicit assumption that site locations are currently randomly selected.

In other words, up to this point in time (when we are building the model), the property agents have taken a completely scattergun approach to picking new sites!  Now this clearly isn’t true of course – even if all decisions have been based on “gut instinct” and experience, with no “science” or data used to back them up, they are certainly never random.

Now the student of statistics might recognise the limitation described above as the assumption of homoscedasticity (or homogeneity of variance).  They might even realise that the assumption doesn’t hold for what they are doing, but by then they will probably crack on and apply a regression analysis anyway and take the output as a “scientific fact”.

The client might well be impressed with the scientific rigour employed, and feel confident in the model ...until they check the results against their knowledge of the UK retail landscape. At that point, the statistically “correct” answer can start to look a bit silly.

This is just one of numerous potential pitfalls which the Artificial Intelligence will always fall foul of.


2. Locations are unique

No matter how many different variables we can collect about a location, there will always be some elements we can’t capture or quantify.   The micro-locational factors, for example, are particularly important where convenience is the main driver.

There is also the human element.  The difference of a good, bad or average manager can make a significant difference to performance – easily 10 or 15% in most cases and much more for some sectors.  We see this even more in leisure than retail, where a customer can be easily turned off completely by one bad experience or one rude member of staff. 

These factors would never naturally surface through statistical analysis, so can only be unearthed through collaboration with the client and building a true understanding of the business.


3. We already use Machine Learning

Machine Learning is a branch of Artificial Intelligence, which sounds incredibly advanced, but all it really means is deriving algorithms that can learn from and make predictions from data.  This is what we do every time we build a turnover forecast model, build a customer segmentation or rank potential site locations.  Machine Learning is simply another word for predictive analytics and covers techniques which include regression, clustering, classification and optimisation.  This is our “bread and butter” in location planning; this is our toolkit.



We can already buy the tools to build a house but that doesn’t mean the house will build itself.  It doesn’t make the architect, engineer and builder redundant – some things are best left to the experts.

Here at CACI, we have a team of highly skilled consultants who specialise in solving geospatial problems across a wide range of sectors, including retail, leisure, automotive, grocery, housing and many more besides – essentially any organisation which has physical outlets.  Typically, we help formulate expansion strategies, optimise estates, build predictive models to forecast turnover and impacts for potential new sites, create dashboards and innovative data visualisation to aid faster more effective decision making, advise on store formats and market capacity across UK and internationally.

We have detailed and accurate knowledge, powered by mobile app data on your customers and how they interact with your physical and digital environments. Find out how we can help you optimise your network.

It seems every article I read is pointing to the imminent demise of this job or that career. The “rise of the machines” means we’ll soon all need to find something else to do with our time. Well I just don’t buy it.

Why artificial intelligence and machine learning can't replace location planners