Say your company is looking to invest in an African country, specifically in a firm based in Namibia making and exporting shower curtains. Research on the market potential for this relatively obscure product yields little information. Your options are to hire a management consultant to conduct due diligence, which can take months and be very costly, or go without. Should your company proceed with the investment?

This is where generative AI comes in. Algorithms can generate business intelligence reports in a manner analogous to a consultant analysing markets and offering advice – except for a fraction of the fee and almost in real time. INSEAD’s TotoGEO AI lab, for example, created algorithms that have generated 1.2 million reports covering world market outlook and trade forecasts across virtually all known product categories (especially obscure ones) for all national markets.

What does a GPT chatbot do or not do?

Here’s how the process works. When you ask a chatbot – the application for which generative AI is best known – a simple question that you might ask a consultant, it kickstarts a process known as “question decomposition”. The decomposition algorithm typically deconstructs the question (or prompt) using various approaches such as natural language understanding. This involves ignoring some of the text, including “a” and “of”, identifying the subject (e.g. shower curtains, copper oxychloride, six-inch copper nails), finding the user’s intent and requirements. For instance, the user may ask for a poem consisting of six lines, making “six lines” a constraint on the output.

The system passes these chunks of information to a trained model such as ChatGPT, which then generates an answer that passes a filter before it is displayed to the user. This is a layer of programming to avoid violating rules established by the model’s owner (OpenAI in the case of ChatGPT). This filter avoids responses that might appear racist, politically controversial or, perhaps, unprofitable to the owner of the algorithm. In these cases, users may receive a boilerplate response: “I’m sorry, but I am not programmed to XYZ” etc. Here’s an example.

While ChatGPT falls short in this area, alternative approaches have come of age, mostly without people noticing. One of the earliest forms to hit the business scene was co-authored by John D.C. Little at MIT. Little developed algorithms that mimicked PhD econometricians in the way that they detect trends, variations and anomalies in vast data sets such as those generated by optical scanners at grocery stores.

To make the intelligence actionable, an “authoring” or “generative” layer of algorithm writes a memo to the marketing manager. An example: “Your competitor is promoting in Cleveland and it seems to be a tactical experiment. You have three useful options: 1. Watch and learn; 2. Mess up their experiment by increasing your prices four percent; or 3. Gain share in Cleveland by moving your product to the end of the aisle.”

The final layer is distribution. The memo is sent to the manager in real time, via email, text message or other means. This article by Little highlights how the technology was deployed at Ocean Spray Cranberries, a fruit processing cooperative.

INSEAD’s journey into AI-powered business intelligence

Full disclosure: I am Little’s academic “grandson”; his student Leonard Lodish was my dissertation chairman at the Wharton School. At INSEAD, the former Dean of Executive Education, Arnold De Meyer, gave our TotoGEO AI lab a small budget to create executive education materials that were tailored to each individual attending a two-week programme. The common subject covered was strategic planning but a participant from the semi-conductor industry, for example, would receive course materials focused on that industry while another participant in the same room might receive materials on the toothpaste industry.

It worked. No matter how obscure the participant’s industry (e.g. copper oxychloride), the course programme had maximal relevance and impact. Feedback from the participants included “can I meet the analyst who prepared the materials?” With this encouragement, we set about putting our AI-powered approach on steroids.

The idea is simple. Prior to my academic career, my work involved estimating the market potential for cellular telephone networks across granular geographies. These estimates proved useful for cell site optimisation modelling. I also worked in the Caribbean, Asia, Africa and the Middle East, estimating the export potential of firms, some of which made rather obscure products like shower curtain rings or toilet seats. Turns out, unsurprisingly, that the more obscure the product, the less anything is published on it – just try Googling the market potential for shower curtain rings in Sri Lanka.

Foreign direct investment to such countries is hampered by the lack of data required to conduct full due diligence. Indeed, information asymmetries between buyers and sellers have long been cited as a reason why companies fail to sign all-important contracts. This problem is especially acute for small, underserved communities, especially those in emerging economies. By focusing on the long tail of products across traditionally remote geographies, AI algorithms can help reduce these asymmetries, thereby increasing investments, employment and value-creation opportunities within these regions.

At INSEAD’s TotoGEO AI lab, we set about creating algorithms leveraging various economic theories (proposed by the likes of John Maynard Keynes, Franco Modigliani, Milton Friedman and Irving Fisher, among others) to extrapolate from sparse data sets. This involves accurately estimating the consumption of a specific product category in one country and applying those consumption patterns in other countries after making the necessary adjustments for local conditions. Once estimates are generated, the algorithm takes care of the entire value chain of content creation, including all meta data, marketing collateral and distribution. MAID plc was an early distributor. Others soon followed.

Reports generated by our algorithms were priced under US$1,000, no matter how obscure the product, covering markets across all countries and cities. Even if data are not readily available online – the algorithm mimics the economist facing a “data desert” (i.e. where only sparse data are available, or too “dirty” to use in their raw format).

The Uber pitch deck and other reactions

Over the years, hundreds of Fortune 500 companies have either purchased one-off studies or subscribed to entire catalogues generated by TotoGEO’s algorithms. Perhaps the most interesting is Uber (then known as UberCab), which cited one of our reports in its now famous 2009 pitch deck:

In fact, most consulting firms now have internal research departments, such as McKinsey’s Research and Information Group, that purchase such reports rather than have human consultants create them at higher costs to their clients.

But other stakeholders’ reactions to AI-generated reports have been decidedly mixed. Some lament the death of human authorship. Others, such as Chris McManus at the University College London, have expressed concerns about the metaphysical and philosophical implications of automated writing.

Many disparage generative AI without adequate understanding of the underlying algorithms. In Wikipedia, for example, the editors write that one title results from “Computer-generated combination of boilerplate text and public-domain data related to… a type of cheese”. In fact, no public domain data exists in that report; all the data are AI-generated since no public source of the data exists.

Finally, some have responded with good-natured humour. American comedian and chat show host Jimmy Fallon poked fun at one of the research reports in his TV show for its droll title, “The 2009-2014 Outlook for Wood Toilet Seats in Greater China”. Another report, a study on fromage frais, won the Diagram Prize for Oddest Book Title of the year. Meanwhile humorous reviews have been written on the obscurity of some products.

Business consultants are still in business

What are the implications of these AI-generated reports? Does it mean that strategic consulting is doomed? Of course not. It just means consultants need to focus on non-formulaic tasks. This begs the question: Which parts of consultants’ jobs are formulaic and which are not?

Long gone are the days where consulting companies could crank out a 500-slide deck, leave it on the CEO’s desk and ride off into the sunset. I have concluded that some 70 percent or more of any given strategic consulting presentation is simply a subset of around 600 PowerPoint slides and therefore ripe for commoditisation.

In fact, INSEAD's TotoGEO AI lab is working with promising generative AI start-ups to help businesses and consulting companies be more efficient and spend less time on slides. For many consulting firms, a large chunk of their fees relate to recommending, not doing, things. With the proliferation of AI tools, the value of consultants will lie in their people skills as opposed to their analytical skills.

Of course, this focus on people skills and value-added thinking started in tandem with the development of generative AI back in the 1990s. Internal strategy teams are often composed of ex-consultants who nudge consulting firms to focus on implementation or new areas like sustainability, agility, eco-system management, automation or growth in slowing markets.

Finally, since many consultants are shifting to performance-based compensation, they need to spend less time tweaking PowerPoint slides and more on measurable value creation. Rapidly generated strategic reports will facilitate better alignment of time allocation for both the consultant and the client.

If an algorithm can create high-end, customised research reports and PowerPoint decks, why stop there? What about videos? Games? Online newspapers? In my next few articles, I will describe how generative AI will shape the future of a host of sectors, from education and gaming to global search engines.