As companies navigate the unpredictability of COVID-19 and finance continues an accelerated phase with Big Data and predictive analytics, AI is making its case for inclusion in business forecasting.

Like CRMs, which have evolved from being a "nice to have" to a "need to have" over time, forecasting is becoming essential to a company's bottom line. It's what helps a business cover operating expenses, buy more inventory, market new products and attract new investments. But traditional models can no longer keep up with changing macroeconomic conditions, and AI-based models add a level of predictability and adaptability that is now crucial.

Traditional forecasting models fall short

On any given day, traditional forecasting might achieve 75% accuracy (compared to AI's potential to reach much higher). But why is this method more limiting? Traditional models lack adaptability. They can only take in a limited amount of data and variables compared to AI and machine learning models, which can take unlimited internal and external data sets.

Traditional forecasting also requires an extended amount of time to evaluate projections versus actuals and adjust assumptions. The assumptions these models use are often high-level metrics that don't tie into the operational realities of the business. AI, however, offers deeper granularity and updates its forecasts at a daily rate, allowing businesses to react and adjust much faster.

Unfortunately, many companies avoid AI as a mere "nice to have," and many that adopt AI abandon their models or don't realize its full ROI. As the AI/Data/BI Product Director at Paro, where we've operationalized AI into our forecasting models, I have come away with some important lessons for implementing AI forecasting and bringing it to its full potential.

Lesson one: Go granular

Many companies currently use a top-down approach for forecasting, in which they focus their model on top-line financial projections. I recommend doing the opposite.

Instead, each company project should get its own AI forecast. The sum of those forecasts equals the top-line performance. This approach increases forecast accuracy and makes it easier to identify a culprit when performance doesn't meet expectations.

Lesson two: Take advantage of AI's data capabilities and retrain your model

Imagine if your forecast considered the call transcript of every customer touchpoint with a client. The sentiment of those calls may indicate if a client will churn. In turn, that indicator would allow you to more confidently forecast growth. The same unique data capabilities can be applied, for example, at an agriculture company that forecasts growth based on weather patterns and global crop yield. A lot of this data would be considered noise in the traditional model.

But AI forecasts are still only as good as the data you provide. If you don't continuously provide the most up-to-date data, the model's performance will degrade over time. It's best to establish a cadence of adding new data and retraining the model.

Lesson three: Define metrics first

Before you build a model, spend time debating the metric, all the edge cases for the metric and how the metric is calculated. That way, what the metrics are and how to use them will be clear to both the data engineers building the model and the key stakeholders interpreting

Lesson four: Manage day-one expectations

The hardest day in any AI implementation is day one. When you launch, you will not get the performance that you are expecting, and if you do, you should be suspicious of it.

It's extremely common for AI forecasts to only have 70-80% accuracy at first. However, they have potential to improve to 97% accuracy over time with fine-tuning and additional real-time data. Don't give up on your model if it's not perfect right away, and don't set an unrealistic expectation with company stakeholders.

Lesson five: Avoid building models in a silo

Building and maintaining AI forecasts doesn't just involve one data scientist. A team of engineers build and adjust the model, analysts interpret the forecasts, company executives use the forecasts for business decisions and so on. To avoid future confusion, define each member's role in building, maintaining and interpreting the AI forecast at the outset to maintain accountability, governance and overall success.

Ultimately, the possibilities of AI forecasting are accessible to any business willing to put in the time and resources to properly invest in growth. But implementing AI is not a game of quantity. In general, focusing on one main model and adding sophistication to it will lead to better results than working on multiple half-baked models. By managing expectations and ownership, defining important metrics and supporting your model with the right data, your business can realize ROI with AI forecasting.