16 Artificial Intelligence Projects From Deloitte .

16 Artificial Intelligence projects from Deloitte Practical cases of applied AIThe aim of this was to draw attention to artificialintelligence in an accessible way, explainsRoovers. “And it worked. Staff were surprisedand intrigued. They started conversationswith AIME and took selfies with her. Sheturned out to be a real conversation starter.”Our employees could then voluntarily signup for our “AI for dummies” course to learnmore about the subject. “The impetus wasconsiderable – even our CEO took part in thecourse,” continues Roovers.As far as the company is concerned, Rooversregards artificial intelligence as an opportunity,not a threat. “It’s true that artificial intelligencecan take over tasks previously carried out bypeople – and can even do them much morequickly and accurately. But the most importantpart of our business, the work that sets usapart from the others, lies in the contact wehave with our clients and in providing a tailormade service and those are things you simplycan’t outsource to an algorithm.” He goeson to point out that artificial intelligence canactually make our work more interesting. “Thedull, more repetitive work can be outsourced,leaving more time for the creative work thatenables we humans to make a difference.”As Roovers says, you can’t stand in the wayof change. The important thing is actually tounderstand how the world is changing andhow you can benefit from that as a company.07“In order to keep ahead of the rest, you haveto stick your head above the parapet. By usingartificial intelligence, we can continue to offerour clients the best possible service – and that’swhat it’s ultimately all about.”What types of solutions is AI able to offeryour company?The case studies in this book give animpression of the possible applications ofartificial intelligence. Amongst other things,AI technologies are used to improve serviceby means of chatbots, to avert cyberattacks,to trace potential fraudsters, to generatebenchmarking reports automatically, toestimate the risks that new customers pose toinsurance companies, to sort large quantities ofdigital evidence and much more besides.In short, the potential of artificial intelligenceis huge. Roovers: “The trick is to find outhow we can make it useable within our ownorganisation. On the one hand, we can do thatby creating smarter processes, but we can alsouse artificial intelligence to set up entire newproducts or services.”But where do you start? The first thing youneed is a knowledge of the technologies. Afterthat, it is possible to explore the possibilitiesthat exist within your own organisation andidentify opportunities and threats. After that,it’s a case of developing proofs of concept – andif those turn out OK, they can then be scaledup.Deloitte is able to assist with all of these steps:from exploring the possibilities to developingproofs of concept and long-term collaborationprocesses and co-creation. “Artificial intelligenceis no panacea”, warns Roovers. “It’s a caseof looking very carefully to identify preciselywhat problem you wish to solve and whattechnologies are available. In some cases, lessadvanced technologies are already sufficient tosolve the problem.”Thanks to the availability of APIs and cloudcomputing, however, developing a proof ofconcept is a relatively simple matter. “The nicething is that you can start small and if it works,you can quickly scale it up”, says Roovers.“There is such a lot you can do if you are smartwhen deploying artificial intelligence. Right now,we’re just at the beginning.”Using AI, companiesmay be able to get adecisive head start

16 Artificial Intelligence projects from Deloitte Practical cases of applied AIDefinitionsThere is a whole host of technologies that are associated with artificial intelligence.Here are just a few: Machine learning is a research field that is capable of recognising patterns in data anddeveloping systems that will learn from those. Supervised machine learning trains systems using examples classified (labelled) byhumans – for example: these transactions are fraudulent; those transactions are not fraudulent.Based on the characteristics of that classified data, the system learns what the underlyingpatterns of those types of item are and is then able to predict which new transactions are highlylikely to be fraudulent. Unsupervised machine learning is able to discover patterns in large quantities of unlabelleddata. It attempts to discover an underlying structure of its own accord, such as by clusteringcases that resemble one another and making associations. For example, retail companies areable to use purchasing data to recognise what products are often bought together and canadjust their offer to reflect that, or can even provide personalised offers. A neural network is a machine learning technology that mimics a structure resemblinga human brain (consisting of neurons and connections), and is capable of adapting its ownstructure to perform the task it has learned more effectively. The more complex neuralnetworks become and the more often they consist of several “layers”, the more we can makeuse of the term: ‘deep learning’. Natural language processing is an area of research that focuses on training artificial modelsto process a human language. Computer vision is an area of research that focuses on the processing of digital imagematerial.08“ Artificial intelligencewill enable us tosolve problems thathumans are unable,or hardly capable, ofsolving”

16 Artificial Intelligence projects from Deloitte Practical cases of applied AITax & Legal09

16 Artificial Intelligence projects from Deloitte Practical cases of applied AITAX & LEGALTAX-I:your virtuallegal researchassistantWhether searching for relevant case law, analysing rulings,or assessing whether a tax case is likely to be successful,tax lawyers have a lot on their plate. And the result is byno means always accurate.10

16 Artificial Intelligence projects from Deloitte Practical cases of applied AIWhat if you could automate this preliminarylegal work? Not only would it enable a taxcase to be settled more quickly, but also moreefficiently and accurately. That was the ideabehind TAX-I, a virtual legal research assistantdeveloped by Deloitte. TAX-I is able to useartificial intelligence to analyse thousands oftax cases of the European Court of Justice,relate them to similar cases, summarise them,and even predict how a court would rule in acase.TAX & LEGALBulk quantities of data“EU Member States are required by lawto publish their rulings and to make themeasily accessible to third parties,” says MarcDerksen, Consultant in Data Analytics & A.I.at Deloitte Indirect Tax, and a member of theteam that developed TAX-I. “So there are bulkquantities of data available.” Working in twosprints of six weeks, the TAX-I team examinedwhat they were able to do with it, and itsoon turned out that the possibilities werenumerous.First of all, the tool visualises the way in whicha new case corresponds to earlier cases. Aninteractive chart of lines shows how casesrelate to one another, and the size of a dotindicates the relevance of the case basedon the number of references. The tool alsoproduces a summary of all cases, basedon how often sentences or parts of themrepeatedly occur in a ruling. “The summariesstill aren’t well-written, flowing texts,” remarksDerksen. “We intend to improve that inthe future. In any case, the summaries arecurrently at a level that enables you to quicklyassess what the case is about.”Finally, TAX-I is able to predict how theEuropean Court of Justice is likely to rule ina case, based on facts that the user of thetool can enter. It uses a machine learningalgorithm that is trained to recognise patternsin tax cases and to draw conclusions. All1153 tax cases of the EU Court of Justice havenow been entered into TAX-I. The team alsohopes to analyse all cases from individualcountries later on. “In the Netherlands alone,11that already gives us 2.5 million cases,” saysDerksen. “This much data will enable usto use deep learning, and produce moreaccurate insights.”Continued developmentThe team has been working for a fewmonths now on making TAX-I simplerand more powerful. The analyses of TAX-Iare being evaluated in the meantime byDeloitte’s scientific office and in a study byVrije Universiteit Amsterdam and TilburgUniversity.In the long term, TAX-I is intended to becomeaccessible for all Tax & Legal consultants inEurope. “It will enable them to save time andimprove the quality of their services,” assertsDerksen. But the current prototype hasalready proven its worth. “I regularly speakto colleagues who found related cases usingTAX-I that they themselves had missed. Thatis a valuable achievement in itself.”All 1153 tax cases ofthe European Courtof Justice have beenentered into TAX-I

16 Artificial Intelligence projects from Deloitte Practical cases of applied AITAX & LEGALAn AI benchmarkstudy of transfer pricingIf a company forms part of an international group, the prices and conditionsapplied to the sale of goods and services within this group must be similar tothose of third parties. This is intended to prevent improper diversion of profitsbetween countries. However, it saddles companies with a problem, says MartijnKrassenburg, Manager Transfer Pricing at Deloitte. After all, which prices ofwhich companies are deemed ‘similar’?12

16 Artificial Intelligence projects from Deloitte Practical cases of applied AITAX & LEGALA benchmark study is required in order toanswer that question. One variant of suchstudies looks at what companies undertakeidentical activities in a similar industry, and thereported margins are checked against those ofthe company in question. That is quite alaborious process, in which many repetitivetasks are carried out manually. “This is why wedecided to find out whether it was possible toautomate this work,” explains Krassenburg.The technology is now beingtrained using large quantitiesof data from previousbenchmark studies13Thousands of screenshotsA benchmark study first examines whichcompanies are similar. An initial selection of afew hundred companies that are possiblysimilar is made by applying filters to aninternational database. They have to bescreened manually, such as by searching for thewebsites of all of those companies and takingscreenshots of them. Krassenburg: “RoboticProcess Automation is the automation ofsimple, repetitive tasks, which in this case aresearching websites and saving screenshots, andit has enabled us to speed up that processconsiderably.”But the team’s ambitions extend further. “Weare busy developing artificial intelligence thatwill automatically estimate the extent to which acompany is similar,” he adds. The system isalready able to perform rough screening, andits self-learning ability means it will becomeincreasingly accurate the more it is used. Thetechnology is now being trained using largequantities of data from previous benchmarkstudies. and technologies such as naturallanguage processing, neural networks andultra-precise entity recognition algorithms arebeing employed for this purpose.Soon, a percentage will appear next to thename of each company, indicating the likelihoodof that company being similar. This will meanthat an ever-smaller number of companies willneed to be checked manually. Ultimately, theaim is to complete part of the task fullyautomatically.The finishing touchThe project is still in the testing stage, butKrassenburg believes that the technology can berolled out widely once it is up and running. “Inthis way, you can test how the intercompanyprices relate to the market, and substantiate thiswith detailed documentation. All internationallyoperating companies that sell products orservices are affected by it. In this regard, AIBenchmark helps them maintain consistencyand is already saving them time.” Deloitte Globalhas already expressed an interest in this tool. Assoon as AI Benchmark can be implemented, itwill be employed more widely.When it comes to the question of whether AIBenchmark will render his own job surplus torequirements, Krassenburg is not concerned.“Performing a benchmark study is not whatcreates our added value. It is the finishingtouch, one part of a much larger and morecomplex process. The really interesting work,understanding exactly how a companyfunctions and how it relates to tax rules, is notsomething that can be simply outsourced.”

16 Artificial Intelligence projects from Deloitte Practical cases of applied AITAX & LEGALSONAR:find labellingerrors indatabases14

16 Artificial Intelligence projects from Deloitte Practical cases of applied AIJust under a year ago, Deloitte was approached by a majorretailer. Its range consisted of over 30,000 products, andthe commodity codes provided by suppliers had to bechecked manually for around 600 new products everymonth. In addition, information had to be entered relatingto the VAT rate and any local levies, such as the battery taxthat applies in Belgium for products containing batteries. Itwas not unusual for something to go wrong when it cameto this labelling. The retailer asked Deloitte for assistancein checking the information entered by human staffmembers.“SONAR allows you to checkthe information entered byhumans far more quickly andaccurately”TAX & LEGAL“Previously, we would have carried out spotchecks,” says Gerhard Smit, informationarchitect and data analyst at Deloitte.“But then we thought, can’t we automatethe checks?” Within one week, he and histeam members created a proof of concept:Similarity Observant Network AnalyticsReport, or SONAR for short. It is a tool thatpredicts the likelihood that the informationrelating to VAT, the commodity code and locallevies entered into in a product database iscorrect.Comparing dataIt works like this: a client supplies a data filecontaining as many details as possible – thecommercial product description, the VATrate, the commodity code and an indicationof whether or not each local levy applies. Butit also contains, for example, the barcodeand other information that can assist withunderstanding the nature of the product.SONAR compares this information against acustoms database containing all commoditycodes, a textual description for eachcommodity code, and the applicable rate ofVAT. The comparison results in a percentageto indicate the likelihood that the label addedby the client is correct. If a label is more than80 percent likely to be incorrect, for example,the product can be checked by a person.A great deal of label-related work is simple,but new, innovative products often requireadditional attention. “Legislation often failsto keep up with reality,” remarks Smit. “Takesmartphones. Should we classify them15as a phone, or as a navigation system, forexample?” Such cases need to be assessed byan expert. SONAR allows the checking of thevast majority of products to be automated,so that additional attention can be paid to thedifficult cases.Bicycle lightsThe SONAR team went to a shop togetherwith the client to test the tool, and carriedout a random check on a shelf of bicyclelights. In the case of one bicycle light, SONARindicated that something was likely to beincorrect regarding the battery tax. “Uponcloser inspection, it turned out that therewas indeed a small battery included in thepackaging, although that wasn’t included inthe description,” recalls Smit. “We thought itwas highly amusing: something we had builtwithin a week had an immediate impact.”SONAR was developed for a client, but Smitbelieves the technology is generic enough tobe implemented for other problems. It worksparticularly well with databases containingat least 2,500 products, and a referencedatabase must be available. “SONAR allowsyou to check the information entered byhumans far more quickly and accurately,”asserts Smit. “And the best part about it is,the more often you use the technology andthe more product information that becomesavailable, the more accurate the results willbe.”

16 Artificial Intelligence projects from Deloitte Practical cases of applied AITransaction detector withregard to the Dutch workcost regulationsThe work-related expenses scheme in the Netherlands,known as WKR, causes a world of problems for manycompanies. Deloitte has developed a clever solution to assistits clients in implementing this tax scheme correctly.Unfamiliar schemeThe WKR permits companies to spend 1.2percent of taxable wages on tax-free allowancesand benefits in kind for their employees.“When the scheme became compulsory forall employers back in 2015, it soon revealed anumber of problems,” explains Guy Thien, taxconsultant at Deloitte. “In practice, for example,there is a major lack of clarity regarding whois responsible for implementing the scheme.It requires data from HR, salary and financialrecords, to name a few, but no one takes therole of managing it.”16What is more, it is still a fairly unfamiliarscheme, and consequently many expensesare incorrectly excluded from it. “If a directorgives ten employees an iPad each, he is oftenunaware that this has implications under theWKR,” continues Thien. “He fails to inform theTAX & LEGALappropriate parties, and the expense fallsthrough the cracks.”when I go for lunch with a client. WKR Analyticsunderstands that distinction.”Categorising and learningIn order to overcome these problems, Deloitte’sFinancial HR Analytics team developed anintelligent system that reads descriptions ofexpenses and is able to categorise them: WKRAnalytics. “You enter all your expenses, and thesystem then picks out what is covered by theWKR. Without our tool, a client might achieve aquality level of 20 percent in its implementationof WKR. We increase it to 95 percent,” assertsThien.The advantageWKR Analytics is now being used by over fortyclients. “I have noticed that the time is ripe forsmart solutions,” says Thien. “Back in 2014,when I told employers about WKR Analytics,they scarcely believed that it would reallywork. Now, they are actually curious as towhat we can offer.” Not only is the number ofusers growing, WKR Analytics is continuing todevelop too. “We are continuing to improve thedashboard, as well as adding new informationand constantly increasing the degree ofintegration with VAT analytics,” says Thien. “Weare now also demonstrating what the level ofdata quality is. In just two hours, we can knowwhether a company’s data are good enough toenable WKR Analytics to be used.”A tax expert is called in for the remaining 5percent. The expert wastes far less time onWKR than was previously the case, and is ableto focus his attention on the doubtful cases.“If WKR Analytics categorises an expenseincorrectly, you can correct it and the systemwill learn from it,” explains Thien. “Everycorrection makes the system smarter and moreefficient. The longer you use it, the less time youwaste on it.”“Once you have gained anadvantage with AI, you don’tlose it very quickly”WKR Analytics comprises a combination ofmachine learning technologies and naturallanguage processing – in other words,knowledge of language and use of words.“The WKR is relevant in specific situations,”continues Thien. “Take going for lunch. WhenI go and have lunch with a colleague, that mayhave different implications under the WKR toThe success of WKR Analytics has not goneunnoticed. There are even a few companiesthat have attempted to copy it. Thien considersit a compliment: “That doesn’t scare me. Onceyou have gained an advantage with AI, you don’tlose it very quickly.”

Audit

16 Artificial Intelligence projects from Deloitte Practical cases of applied AIAUDITGRAPA:assistancewith riskstrategiesWhen auditors determine a risk strategy, they partly baseit on knowledge that they gained during previous audits.Deloitte is now developing a smart personal assistant thatsupports auditors using the pooled expertise of all theirfellow professionals.18

16 Artificial Intelligence projects from Deloitte Practical cases of applied AIKnowledge gainedDetermining the risk strategy forms animportant part of an audit. Twan van Gool,Director Innovation & Analytics within the Auditdepartment at Deloitte, explains: “The auditordetermines which sections of the annualaccounts are high-risk, based on marketdevelopments, new legislation and regulationsor events within the company, for example.The chosen risk strategy determines thesubsequent audit method.”AUDITWhen formulating the risk strategy, anauditor’s knowledge gained from previouscases will be very valuable. “Every business isdifferent, of course, but an audit is an audit,”notes van Gool. “While lingerie and cyclingmight ha

1 Artificial Intelligence projects from Deloitte ractical cases of applied AI 05 According to some, artificial intelligence is the most promising development for the future. From curing cancer to resolving the global hunger crisis, artificial intelligence is being presented as