The business of Artificial Intelligence

Question

I'm putting together a presentation aimed towards entrepreneurs on the present state of industrial AI development, titled "The business of AI"; however, what little resources I have found on Google seems awfully outdated.

So I turn to the nice folks on Stackoverflow: Of the present day used systems, which products do you consider good business cases of applied AI? I'm looking for concrete product examples (not just wide technologies), in either consumer, or business everyday usage, with high profitability, and/or impact factor.

Edit: I understand, that SO might not be the best place for such question; pointing me to relevant discussions / forums would be greatly appreciated.

Edit2: Please also name the most prominent example of the technology (such as GMail for spam filtering, etc)

Results so far:

The following is an overview of AI systems in current business use, ordered by the strength of correlation between improvements on the problem-solving capability of the system, and improvements on the bottom line:

-Adsense, and adwords:

Problem: given a list of classifieds, and a list of website placements, select the highest value (probability of click through X price of clickthrough) advertisement for given placement.

Used AI technologies: clustering, and similarity search

Currently providing approx 30% of all revenues of Google

Method of capitalization: directly on the feedback loop

-Google search:

Problem: given a list of keywords, select the most relevant websites

Used AI technologies: (back-link counting), field-specific spam filtering, (genetic-inspired) duplicate-filtering, ...?

Method of capitalization: full text search is centralizing the web -> attentionware.

-E-mail classification:

Problem: given a large amount of incoming mail, classify it on properties -such as "spammyness", "business mail", etc

Used AI technologies: machine learning, (usually) Bayesian classification

Implementation:

• GMail (best spam classification as of 2008)
• Fogbugz (1) (2) (multi-classification)

Method of capitalization: not directly; used as a competative advantage

-Consumption pattern recognition:

Problem: given a list of historical baskets, predict the outcome of a given deal/offer, best product placement, and minimize loss leadership

Implementation:

• Tesco's Clubcard system (by Dunnhumby, see answer below)
• amazon's recommendation system, ?

Used technologies: data mining? (decision trees? machine learning?)

Method of capitalization: reducing costs, and improving conversion rate

-"Expert systems":

Problem: for some domains of expertise, demand for expert decision making is much higher, than availability of a competent expert. Thus, these kind of systems act as a limited decision proxy, by being programmed to solve a subset of the problem domain. Usually very hard-wired, with little-to-none "intelligent" behaviour.

Implementation samples: Quicken (for personal finance), Mycin (historical, medical diagnostic tool)

Method of capitalization: historically as commercial software; SAAS nowdays

-Voice and speech recognition:

Problem: Given an audio input (sample: menu navigation, voice mail, phone order), determine either the speaker ("who"), and/or the plain text ("what") of the audio; subproblems involve removing background noise, tone/voice recognition, etc

Implementation:

• Google Voice
• Nuance's speech recognition

Method of capitalization: Off-the-shelf-software; potentially will increase the relevancy of advertisement for video ads (via text recognition)

-hand writing, and optical character recognition

Problem: Given an image of either scanned textual pages, or hand-writing, determine the plain text (ref)

Implementations: Wikipedia OCR software list, and Handwriting recognition

Method of capitalization: direct software

-Sentiment analysis:

Problem: given a large bunch of plain text, determine the most prominent topic of the conversation, along with it's carried sentiment

-Fact extraction:

Problem: extract object-property-value trios from plain text

-Quote extraction:

Problem: extract topical quotes, along with it's source, from plain text

Implementation: Google labs inquotes

Method of capitalization: not directly; used as a competative advantage

-Machine translation:

Problem: given a plain text in one language, provide a translation in another language, with highest accuracy possible.

Method of capitalization: not directly; used as a competative advantage (also, ads can be translated for transparent localization)

-Netflix:

Problem: given a list of previous user ratings on movies, maximize the accuracy of predicting future ratings of non-rated movies

Used: meta-algorithm: outsourced public/open research with a high price for breakthrough-performance

Method of capitalization: used as a competative advantage, also basis of new video recommendation

-Lingpipe:

Providing basic NLP tools for research (named entity extraction, POS tagging, etc), free for academia, paid-for by commercial usage

Method of capitalization: direct software (although academic license is available)

-Image similarity:

Problem: given a single image, find images with matching similarity, OR content

Method of capitalization: Not directly (-ads)

Implementation:

• TinEye
• Google Similar Images

-Further reading:

Most recent known survey from 2001:

• Artificial Intelligence in Business

Wikipedia provides a couple of good overviews, although not business-oriented:

• Applications of artificial intelligence
• AI behind the scenes

Answer 1

Most if not every single game in the world has some level of AI. The whole thing about gaming is defeating a computer at some level.

I personnaly work for a research lab building robots. We aim for different levels of automation depending on time/budget constaints but this field is rapidly evolving mostly with computer vision domain becoming increasingly cheaper.

Answer 2

I would consider Bayesian filters to be a form of AI. Used very widely to filter spam.

Answer 3

How about the iRobot vacuum cleaner. It works pretty good

Also, AI is used by credit card companies and banks to detect fraud, automate loan applications, etc...

Answer 4

Using neural networks to build diagnostic systems.

Answer 5

Modern cars ar full of AI xD

Answer 6

Stockmarkets and retail-sector consumer patterns

“Now it’s an arms race,” said Andrew Lo, director of the Massachusetts Institute of Technology’s Laboratory for Financial Engineering. “Everyone is building more sophisticated algorithms, and the more competition exists, the smaller the profits

AI is used in many stock-market software systems for:

• Analysis and learning of trends and trend triggers
• Optimization of investments based on non-direct events (Neural Networks)

and tons more. StockMaster from MIT was a good example, but that has gone walkies ... There are also tons of 'standard algorithms' IIRC.

• Some Info on Goldman's algorithms
• Info on Banks use of algorithmic trading
• Wikipedia's Algorithmic Trading Article, references common systems

Many solutions are roll your own and highly proprietary so that banks can cut the number of traders and increase margins. Many banks have algorithmic trading departments :)

I think this highlights AI's uses in the analysis of human behaviour. What makes it more interesting is that human behaviour, which often defies logic, is presented in a purely numeric environment ... making it an ideas source of data for AI, predictive algorithms and AI on data graphs that account for or can guess at a human reaction by looking at a human's view of the data and prior behaviour.

Another, subtly similar area is in which AI is combined with statistical analysis is buying patterns in the retail sector such as Dunnhumby's data mining of Tesco Clubcard data in the UK and the Nectar points scheme that can match buying patterns across many companies. This data is often used to guess the future success of deals and offers, and which areas of the country they would prove most popular ... allowing loss leadership with minimal loss and maximum exposure to the consumer.

The problem is, when AI starts predicting AI, things get messy (or is that cleaner) in mathematical terms.

Response to comment Examples of buying patterns outside the online market, as mentioned above are data mining of credit-card and clubcard systems. These are used mainly for regional and multi-store retailers to better categorize their customers. They may find that store A and B both have a high population of 20-30 year-old wealthy individuals with male buying patterns, and so will weight the stores products to that demographic. Store C, in close proximity to A and B may have a mixture which also encompassing A and B's customers when they commute. The AI in systems like these may then specify store layout, deals and new-products based on the data from the clubcards. The AI will make these decisions based on previous successes of similar products, demographics, exposure in the store through location and many other factors. Questions such as:

1. Which stores should 'new product X' be trailed in to give a good approximation of it's desirability to a given demographic?
2. Where should the product be placed? Premium 'eye-level'?
3. Which stores should deal-X be launched in, to ensure loss leadership with minimal loss but maximum marketing effectiveness for the brand/store/company?
4. Should a petrol station be built at store Z, does it see many commuters?

Although these decisions are not 100% AI, in that some person has to green-light suggestions, weighting and iterative analysis of sales and demographic movement data provides input into the likelyhood of success or failure for a strategic movement. Much like the stock market.

I should note, I have not been actively involved in the development of these systems, but gather this is what is in place. Dunnhumby is an example of a company contracted for analysis of tesco-clubcard data. There will be others and most examples will probably come from Europe as we have a high store density per square mile.

Answer 7

google maps, classic AI pathfinding algorithm.

Answer 8

Hmmm, not sure if it's "AI" per se but when I was in Property Preservation we realized that our decision making processes in determining if an Invoice was paid and quality assurance etc was so proceduralized we could set up "Robots" to do it for us. So we had these little scripts that we use to check this website, that report, this database, hit the mainframe and so on and do the work for us. The only reason we really needed them was "glue". If the systems could all talk to each other then the whole thing could have been automated. But it wasn't so we basically botted the individual clients to bring it all together. I know one thing I wrote (which in hindsite was horrible, horrible, code) freed up 3 and half full time positions (although they just found more work for em:)). And they stayed in use for several years.

So I would say you can easily use automation to replace the human, if all the human is doing is interfacing between systems that should have been integrated.

Answer 9

Recommendation systems (Amazon etc.) use AI techniques such as clustering to find similar users/items.

Answer 10

If you are looking for software applications that use AI such as Genetic Algorithms, Neural Networks, etc then google these:

MatLab, Matematica, GeneXproTools, KXen and SQL Server Analysis Server.

There are of course many others. I am one of the GeneXproTools developers and I think that the field is still very much in its infancy. Success is usually determined by factors external to AI such as vertical market presence, bundles with other software, etc.

Answer 11

Let's not forget Netflix, who are interested in AI (mostly supervised machine learning) enough to offer $1,000,000 to anyone who could improve their recommendation algorithms. And it looks like one of the teams is about to win.

Wikipedia has a short summary of the competition, and Wired did a recent piece about it as well. One of the better teams (BellKor) has published a variety of papers about their efforts, including this recent tech report describing their approach -- it's about ensemble learning, clustering, and all of that good stuff.

It's a very recent and news-worthy example of the commercial sector's interest in AI.

Answer 12

My company uses some AI to extract market prices from dealer quotes (they are sent to us in a wide variety of formats, mostly text files).

Answer 13

I'm surprised no-one has mentioned industrial robotics yet. This is a huge current area for the application of AI.

Answer 14

There are a few companies using it in license plate recognition (for security systems and cameras in police cars)

Answer 15

Nuance.com is doing a pretty decent business on voice technologies, such as speech recognition.