Today's work in artificial intelligence is amazing.We've taught computers to beat the most advanced players in the mostcomplex games.We've taught them to drive cars and create photo-realistic videos andimages of people.They can re-create works of fine-art and emulate the best writers.Yet I know that many businesses still need people to, e.g., read PDFdocuments about an office building and write down the sizes of theleasable units contained therein.If artificial intelligence can do all that, why can't it read a PDFdocument and transform it into a machine-readable format?Today's artificial intelligence algorithms can recreate playableversions of Pacman just from playing games against itself.So why can't I get a computer to translate my colleague's financialspreadsheet into the format my SAP software wants?
It’s worth trying to understand these positions, and in the absence of being able to have a frank discussion with the decision-makers, we’ll try to analyze their decisions with a model of human behavior. This model comes from the combination of two ideas: one about beliefs and social structures responding to economic conditions and the other about how political change happens. “All models are wrong, but some models are useful” – let’s hope that this model leads us to some useful and not harmful insight, but beware the dangers!
Despite two decades of advancements in artificial intelligence, it feelsthat the majority of office work consists of menial mental tasks.We should expect that artificial intelligence would automate this workin much the same way that past machines automated physical labor.Indeed, many writers have been sounding the alarms about the coming joblosses.Though I believe that artificial intelligence would make both existingjobs more interesting and create more jobs in yet-unthought-of fields.In practice, however, I still see many people doing jobs that computersshouldbe able to do but justcannottoday.Why is that?
I think part of the problem may be in the way we interact with computers.Computers are based on an architecture that requires explicit, preciseinstructions on how to manipulate data.Even with voice-controlled virtual assistants on our smartphones, westill interact with them by giving them explicit, precise (albeithigher-level) instructions.Artificial intelligence algorithms likely can infer many of thoseinstructions implicitly.Perhaps we are awaiting a second information revolution - maybe usingExcel for modern business tasks is like writing software in machine codewhen higher-level programming languages are available.This might be true, but I think we face two more immediate problems:a lack of data and a lack of awareness.
Today's artificial intelligence is powered by data.And the bulk of today's data comes from the internet - text, images,videos, and our interactions with them.If a group of software engineers wanted to create a model that could,for example, identify the make and model of a car in a picture, theycould start with a pre-trained model from other researchers that detectsobjects in pictures and then "top it up" by training on a smaller set ofexamples that just includes cars.This is calledtransfer learning.But there is no existing "document-understanding" model that we couldadapt to our specific business processes via transfer learning.The excel spreadsheets, marketing brochures, legal contracts, and otherdocuments that make up the business world are hidden in email inboxesand other silos within various companies.No group of researchers can train a "document-understanding" modelsimply because they don't have access to the relevant documents orappropriate training labels for them.
What's more, artificial research teams lack an awareness of the specificbusiness processes and tasks that could be automated in the first place.Researchers would need to develop an intuition of the business processesinvolved.We haven't seen this happen in too many areas.The big successes have happened where the problem is easily understoodand has many publicly-available examples (machine translation), wherethere is a promise of a massive ROI (self-driving cars), or where alarge company arbitrarily decides to throw enough resources at theproblem until they can crack it (AlphaGo).
This means, however, that we can expect artificial intelligence tosucceed in automating business processes when 1) researchers are able tofocus on a specific problem, and 2) they are able to accumulate enoughdata to train a workable model.(Another criterion for success is that should aim to empower the peopleinvolved in the process, not replace them, but that is for a differentdiscussion.)And where they succeed, people who work in those industries can expectto be spending more of their time doing interesting, creative work andless time doing dull, time-consuming tasks.A great example of this isProda, a London-based startup that canautomatically standardize commercial real estate data.Startups or initiatives that promise more general solutions are likelyto run into difficulties.
The upside of this is that there is a near-endless stream of businessopportunities with significant moats around them.Each business process is a chance for automation, and therefore anopportunity for a business to save an entire industry much time andmoney and pocket some of the profits.If one can understand the AI research well enough, understand thebusiness process thoroughly enough, and the gather enough data to traina model, then one can build a profitable company out of it.
Thank you to Gabriella Abraham and Jagna Feierabend for reading earlier drafts of this essay.