Pandemic disruption pushes companies to rethink AI training
AI systems built on historical data face new strain as businesses confront patterns that no longer match pre-pandemic conditions.
By Maya Lindqvist · Senior Technology Correspondent
3 min read
The coronavirus pandemic is exposing a weakness in many business AI systems: they often rely on old data to make decisions in conditions that have changed sharply. Ahmer Inam, chief AI officer at Pactera Edge, told Fortune that the crisis is likely to speed adoption of AI methods built to handle a wider range of scenarios.
Fortune’s Jeremy Kahn reported in the Eye on AI newsletter that many AI tools are trained on historical patterns. That approach can falter when lockdowns, travel limits, altered consumer behavior and other disruptions make current conditions look unlike the data those systems learned from.
The problem may not end when restrictions ease. Fortune reported that a return to business with some distancing rules or travel limits could create another set of patterns, different from both the pre-pandemic period and the height of lockdowns.
Human oversight remains central
Inam told Fortune that one safeguard is keeping people involved in AI decisions. He said AI systems perform best when they assist human judgment rather than fully replacing it, especially when conditions are changing quickly.
He also said companies should closely monitor the data being fed into AI software. According to Inam, “model drift,” the gradual shift in data over time, is a common issue even in ordinary business operations because companies and markets change.
Pactera Edge, based in Redmond, Washington, provides technology consulting and services and helps businesses, including many Fortune 500 companies, implement AI, Fortune reported. The company spun out from Chinese IT company Pactera Technology International in January, according to Fortune.
Simulations offer another path
Inam pointed to reinforcement learning as another way companies could build more durable AI systems. Unlike supervised learning, which trains algorithms on historical examples, reinforcement learning teaches software through experience, often inside a simulator.
Fortune noted that researchers have used reinforcement learning in recent years to build AI capable of beating humans at games including Go and poker. Businesses, however, have been slower to adopt the technique.
According to Inam, the barriers are practical. Building a dependable simulator can take time and money, reinforcement learning requires scarce technical expertise, and training those systems can demand costly computing power.
Inam said the investment can pay off because simulated environments let companies expose algorithms to many possible situations before the software is used in the real world. Fortune reported that he has built simulators to study how climate change could affect coffee supply reliability and pricing for a coffee shop chain, and how a hurricane could affect regional sales for an automotive retailer.
More recently, Inam told Fortune, he built a simulator for a logistics company seeking better routing. The project took six months and ultimately helped the company save millions of dollars in fuel and labor costs, according to Inam.
Inam told Fortune he expects the pandemic to accelerate corporate use of more sophisticated AI techniques. The shift reflects a broader lesson for companies using AI: systems trained only on yesterday’s data can struggle when the business environment changes abruptly.
This story draws on original reporting from Fortune.