Climbing the analytics ladder honestly: food and beverage analytics at Marriott
Food and beverage was the hardest analytics domain in lodging, assembled from point of sale, inventory, and manual systems that did not talk to each other. Properties could see what happened last night. They could rarely explain why, compare across the portfolio, or act ahead of demand.
The situation
I was asked to define the requirements for, and lead the deployment of, an enterprise food and beverage analytics capability across property and above-property levels for the Americas. The temptation with a new analytics platform is to promise prediction on day one. The reality is that most organizations have not yet earned the right to trust their own descriptive numbers.
What I did
I architected the requirements around the analytics maturity ladder: descriptive, then diagnostic, then predictive, then prescriptive, in that order and without skipping. Skipped rungs are where analytics programs quietly die. A property that cannot yet trust its descriptive numbers has no business acting on a forecast.
Two requirements followed directly from that. The rollout was gated on a data quality audit, so the analytics would sit on numbers worth analyzing. And the diagnostic layer explicitly incorporated weather and promotions data, so we could begin to separate what caused a result from what merely accompanied it.
Then we made the insight coachable. The clearest diagnostic finding, which servers underperform and why, became a Single Server Mentoring program: managers coaching specific behaviors, with performance measured before and after.
What happened
92 percent of pilot properties passed the data quality audit. Usage reached 250 active users across 13 properties at 3.52 views per user per day, against a 3.0 enterprise standard, which told us the platform was becoming part of the daily routine rather than a monthly report. At one property, an open item cleanup cut mis-keyed sales from $62,500 to $3,000 a month, restoring measurement integrity on roughly $714,000 of annualized sales that had been attributed to the wrong products. The mentoring program measured a 3.83 times return and roughly $827,000 in annualized incremental sales across 15 properties.
Those last figures deserve their asterisks, and I will supply them. The mentoring results came from before and after comparisons, not a controlled test, and because we targeted low performers, some of the improvement was regression to the mean. Today I would design it with a matched control group or a staged rollout. The return figures are the vendor's measured numbers, annualized and net of cost of goods, and I present them as exactly that, never as an independent audit. I am proudest of the data integrity work, because it is the least exciting number in this study and the one everything else depends on.