My friend Dave owns a small plumbing company with eight employees. For three weeks last summer, he tortured himself agonizing about purchasing a second work van. He made a spreadsheet. He talked to his accountant. He talked to his wife. He talked to me. He lost sleep.
He used a method from AI research called "regret minimization" (explained below) and made the decision in twenty minutes. While the idea behind the method is easy to understand – it’s basically a very simplified concept – it changed how he framed the question. It helped cut through all the noise.
Below is what I’m talking about, and why I believe every small business owner needs to know about it.
The AI framework that actually translates
Researchers from the University of Alberta created an algorithm in 2007 designed to make decisions based on incomplete data. The researchers’ AI did not use forecasting methods; it evaluated the level of regret it would experience if it had done something differently. After making this evaluation, the AI simulated countless potential scenarios, modified the strategy based on minimizing regret, and arrived at strategies that worked well no matter which option was chosen by the opponent.
Eventually, the AI beat top human experts in competitions. However, the underlying principles are universal.
Dave’s Van Decision:
He was concerned as to whether or not purchasing the van was the “correct” decision. Wrong question. Much better question: “If I decide to buy the van and business slows down, how much will I suffer? If I do not buy the van and business increases, how much will I pay in lost jobs?”
The second scenario suffered more. Dave would lose customers to competitors who could respond quicker. The first scenario was somewhat manageable – he could either sell the van or rent it.
Decision made. Regret minimized.
Three things I borrowed from AI research
Link Your Decision Speed to Reversibility: AI systems typically allocate almost none of their computational power toward decisions that are reversible quickly and relatively little toward irreversible ones. I borrowed this idea directly.
Deciding on a new supplier for paper towels? Five minutes. You can simply change suppliers within a month. Deciding on a five-year commercial lease? Take three weeks. The length of time you need to deliberate should correspond to the severity of the consequences of making a poor decision.
Dave took three weeks deciding on a van. A van is a large but reversible decision – you can always sell it. Therefore, he should have taken three days and not three weeks. Meanwhile, he signed a three-year service contract with a parts supplier after just one telephone call. Such a serious commitment deserves significantly more reflection than it received.
Do Not Evaluate Decisions By Outcomes
One difficult point to accept: we judge our decisions primarily by their success or failure rather than by the quality of the decision process itself.
AI researchers have proven mathematically that optimal decision-making leads to undesirable outcomes a high proportion of the time. Libratus --the Carnegie Mellon AI that defeated $1.77 million worth of bets against top human experts--lost approximately 45% of individual decisions. Those decisions it won produced marginally larger returns than those it lost, overall, across 120,000 decisions. That is what optimal performance looks like.
For Dave: Even if Dave hired a plumber who interviewed well and checked references thoroughly, and that plumber quit after only two months --that is not a poor hiring decision. That is variance. If Dave stopped hiring plumbers from that talent pool due to one bad hire, he would be optimizing for luck and not for process. That is actually a bad decision.
Compounded Small Improvements: AI systems do not generate vastly superior individual decisions. Rather, they produce 3-5 percent better decisions consistently across thousands of decisions. These benefits may appear invisible individually. However, collectively they result in dramatic improvement over time.
For a small business: reviewing during fifteen minutes of weekly downtime what goes right and wrong and checking competitor prices on a monthly basis rather than never and asking one customer for candid feedback each week. Each activity seems insignificant alone. Each contributes to cumulative growth.
Six months ago, Dave began conducting a fifteen minute review of his week each Friday. Six months later, he said: "I don’t feel like I am making better decisions. But for some reason things are going better." That is cumulative growth. You cannot sense it occurring, but it shows up in results.
The information advantage you’re not using
A fundamental aspect of competing in any arena according to AI research is that the party extracting more information while providing less to others enjoys structural advantages.
Many small business owners fail to develop systematic knowledge regarding their competitors’ activities. As such, they generally become aware of competitive changes via passive means — “Bob’s Plumbing added a new sign.” — without linking these observations together.
Competitive activity indicating increased hiring of additional personnel? Competitive activity likely booking up fast. Competitive activity offering discounts? Competitive activity may be slow. Competitive activity experiencing decreased Google reviews? Competitive activity possibly decreasing in quality. In isolation, each observation represents a minimal amount of information individually; collectively they provide insight into direction in which the market is trending.
None of this requires software or consulting services. Simply take ten minutes each week writing down everything you’ve observed about your competition and examine this material on a monthly basis. You’ll clearly recognize patterns that individuals failing to record competitive developments will be unable to identify.
The actual lesson
While there isn’t a specific technique from AI decision-making research that I learned – although I think many readers will find several useful techniques – the most profound realization I developed from studying AI decision-making research is that uncertainty is forever. You will never have sufficient information available so that you can be sure about any significant business-related decision — regardless of how extensively you seek information — simply because this information doesn’t exist yet — it relies upon future events by customers, competitors, and markets that have not occurred.
Individuals able to manage uncertainty well aren’t attempting to remove uncertainty altogether. They develop practices that allow them to achieve good results despite uncertainty. Fast decisions regarding reversible choices. Thoughtfully-made decisions regarding non-reversible ones. Evaluations of processes as opposed to evaluations of outcomes. Small differences that grow exponentially.
Dave purchased the van. Business increased. Luck favored Dave in terms of timing — however, if luck had been absent — if business had declined instead — the process utilized to arrive at that decision was correct. This is what he would tell you truly mattered.
