A P R I L 2 0 1 8

Tax Technology RoundtableGerman MNEs eye AI dividends

Roundtable

www.tpweek.com 1

Tax Technology Roundtable:German MNEs eye AI dividendsSonja Caymaz, April 2018

TP Week hosts a roundtable withGerman multinationals AUDI AG andHenkel and consultancy WTS todiscuss progress in implementingartificial intelligence andautomation in tax departments. Theclear message is that machinelearning is happening now and willamplify tax capabilitiesexponentially in the future, muchbeyond our current purview.

Automated data processing is rapidlyspreading in the tax world and the sub-

sequent step of implementing machinelearning to apply rule-based processes totax is what frames artificial intelligence (AI)as one of the biggest opportunities for busi-nesses.The TP Week roundtable gives insights

into what is required by corporate taxdepartments to make the leap and fore-shadows many of many of the changes taxdepartments face with the advance of AI. The German Research Centre for

Artificial Intelligence, or DeutschesForschungszentrum fuer KuenstlicheIntelligenz (DFKI), and global tax consul-tancy WTS, are in a joint collaborationexploring the potential for practical applica-tion of AI in tax and the use of new soft-ware prototypes by partnering withGerman multinationals AUDI AG, Bosch,E.ON and Henkel.Prof. Dr. Peter Fettke, who spearheads

the DFKI’s collaboration on AI and tax,believes that significant improvements thathave recently been made in machine learn-ing will revolutionise business and taxdepartment processes.“Complex processes with mass transac-

tions like VAT or customs will quicklyadopt AI technologies to efficiently handlelarge amounts of data and gain valuableinsights. Examples include the identifica-tion of anomalies in mass transaction datathrough self-learning systems and theexploitation of further data sources liketexts or images for instance by means ofoptical character recognition,” Fettke says.

As a starting point, and as the minimumrequirement to keep up with evolving tech-nologies, companies need to get their datainto the cloud. Recognising the potentialfor enhanced human capabilities and rele-vant skills training are also key to reapingthe benefits.Fettke foresees the advent of tax tech-

nology, a new discipline that combines theknowledge domains of tax professionalswith that of AI and data specialists, andapplies this to tax processes. This begs the question: what sort of

high-value tasks humans will be performingand what kind of skills will they need? Ittakes an incredible amount of foresight atthe beginning of a process to understandthe possibilities. Most practitioners agree that algorithms

can be harnessed to ease complexity andtake on a support function in tax depart-ments, but as implementation of machinelearning progresses new areas of opportuni-ty are beginning to surface exponentially. As machine learning becomes more

ubiquitous, venture capital investments inAI in 2017 reached a peak of 309 deals inEurope, worth a collective €1.2 billion($1.5 billion), according to PitchBook’s2017 European Venture Report.“While many AI base technologies come

out-of-the-box and deliver decent resultson standard applications, the real challengeis to select technologies and adopt them tofit a specific use. In most cases, additionaltraining and configuration is necessarywhich requires both expert knowledge andappropriate amounts of high-quality data,”Fettke explains.

Transfer pricing synergies Transfer Pricing and customs are twofunctions that stood apart but now stand

to gain from being aligned throughautomation. Vanessa Just, project lead for AI at

WTS, tells TP Week about some of the keyfindings of the collaboration: “The studyclearly revealed indicators that show AIhas the potential to tremendously supportrepetitive tasks that only require a slightextend of social intelligence, creativity andphysical interaction. The question of lia-bility is more nuanced as traditionalmachine learning approaches deliver theirresults in a black-box-like way, i.e. theresults are not explicated or justified. Thisissue makes it hard to determine liabilityfor results unambiguously and that is whyself-explanation is still an ongoing topic inmachine learning research.”The capabilities of predictive analysis on

a large scale will augment cost-savings andenhance the influence of the tax function inbusiness strategy and decision-makingacross the supply chain. Just predicts the next stage of AI evolu-

tion will focus on a much deeper integra-tion of today’s dedicated, individualtechnologies into a comprehensive systemacross all business functions in the nextfour to five years. “By integrating data across functional

borders, AI will enable more profounddecisions and leverage potential for globaloptimisation,” she says.

AI & Tax RoundtableTo better understand the challenges andrequirements to make the leap to AI in taxand transfer pricing, TP Week invited WTSand German multinationals Henkel andAUDI AG to discuss their collaborationand ongoing efforts to integrate andenhance automation and machine learningin their tax departments.

Fritz Esterer, CEO, WTS Germany; Prof. Dr. Robert Risse, Corporate VP Tax & Trade, Henkel; Axel Dewitz,Head of Taxes, Customs and M&A, AUDI AG

Roundtable

www.tpweek.com 2

What are the basic requirements forimplementing artificial intelligence in tax? Fritz Esterer: You have to fulfil a few pre-conditions before AI can bring value to yourcompany, and cloud technology or homog-enous databases are very important to gath-er data. With Henkel we learned that onlywhen all data is available in one cloud, youcan make use of this data with AI. WithoutAI you would never structure a data lake,but with AI you can use data that is totallyunstructured. If you use many clouds inmany countries it is very difficult to com-bine the clouds to merge the data. A coun-try or company that is very far from adigitalised process they can still make use ofAI if they fulfil these few preconditions. You can extract the data from Excel or an

SAP or Oracle system; you can extract alldata from a big data lake by using AI andstructuring it. Companies do not have toabolish Excel sheets or the different taxtools they have in place, they only have tomake use of the data that is already in thosetools in an intelligent way. With more dataquality a company can reduce compliancerisks and have better tax planning.

How did you get started on theintegration of machine learning androbotic process automation?Robert Risse: The most important andfirst step: You should have all company datalike finance or controlling data available ina cloud. In the course of artificial intelli-gence, machine learning is an emergingtopic, but if you do not have the companydata at hand, you won’t be able to make useof machine learning or process mining.Therefore, it is most important to work ondata availability in the beginning. We – atHenkel – integrated most of our companydata in a Microsoft cloud, but of courseother cloud solutions like those fromGoogle or Amazon can be used as well.Due to our agile entry in a new field oftechnology and innovation, we areadvanced in the process and to some extentone and a half years ahead of others. Thebandwagon effect will reduce the head-start, but it is most important to be upfront with the others.

Axel Dewitz: We have not implemented AIor modern automation (RPA) in the taxand customs function at AUDI AG. Itrequires deep and thorough preparation.

Before implementing AI you have to havefull control of all processes and you need tohave an analysis of all relevant processes.Otherwise there is a risk to digitalise aprocess that is not intelligent. Our first con-tact with AI for the tax and customs func-tion has been the approach by Mr Estererand WTS in cooperation with the DFKIand the study they started. This was thekick off for us to start digitalising the taxfunction. Today I can say it is an integralpart of our tax, customs and M&A strategy.This will be a three-step approach – first

processes, then robotic process optimisa-tion and then AI.We have started a comprehensive

processes analysis that is also the base for asound tax compliance system. Then we aregoing to kick off robotic process automa-tion, which is a macro, an algorithm that isnot only closely linked to one system butcan also interlink different systems. AUDIAG as a company has set up an RPA archi-tecture that will be hosted in India in a‘robotics farm’. We have a very specific roofof concepts that we want implemented inthat architecture.Implementation of AI for certain

processes should not be done only for thetax and customs function because ourvalue-adding service is that we have fullprocesses in our analysis and those start inmarketing, then procurement and control-ling and so on.

Where are you seeing the immediatebenefits of machine learning? Risse: We started automation and process-es leading up to AI usage in the EuropeanVAT area. Now we get benefits out of thisprocess documentation and start to imple-ment robots doing this manual work.Downloading SAP reports and filling outtax declarations, this is now automated bythe use of five programmed robots. Basedon this exercise, we are going to usemachine learning as a further step to com-bine the robots’ activities.For transfer pricing we have introduced a

tool by a small start-up, Optravis LLC, aspin off from a Swiss chemicals firmClariant. The tool provides an automatedset up for the mark-ups in cost-plus arrange-ments. There are machine learning exerciseson how to determine the cost-plus amountsin a well-designed process format, but this isprogrammed by the start up.

At Henkel we apply the residual profitsplit method, and for the routine functionremuneration we use the cost-plus system.This is now automated by the tool and bymachine learning so that we can testwhether we have used the right or thewrong mark-up. The first test calculationfor all Henkel entities for 2018 is done bythis machine. For example, today we are clustering the

mark-up calculation in groups. This meanswe have sub-business units to calculate thecost-plus mark-up amount. With the toolwe are able to calculate on a product basis.This benefits us because there might beproducts in the sub-business unit that aredeficit areas. If they are loss-making, itmeans that the market price is lower thanthe transfer price calculated by this basket.In a basket there may be to some extentproducts that are loss-making productsbecause the all-over basket is positive, butsome products are negative. By using thetool and machine learning, we can really seethat the machine is considering that weeliminate these negative products ratherthan using a different mark-up. Themachine is telling us to which extend this isfeasible. The outcome of these analyses is really

brilliant and very detailed. We were not ableto do this beforehand in Excel or with otherSQL tools, but with this machine learningexercise we are able to do so. This does notonly saves efforts, but also creates benefitslike some additional tax savings for this year.

How great is the potential for transferpricing and customs optimisation withAI? Esterer: We have been familiar with bigdata technology for many years but welearned from our partners in the AI study,especially Henkel, that before they used AI,the analysis of customs data could not havebeen done in such an efficient way as it iswith AI. AI is currently the most excitingand most important area in tax because withAI we can potentially solve a lot of the prob-lems we currently experience. With AI we can identify processes and

anomalies, find mistakes and deviations,which we would not have been able todetect with older existing technologies.Human beings could never identify devia-tions with fair trade agreements the wayAI can.

Roundtable

www.tpweek.com 3

We see a lot of potential for analysis ofbig data in taxes in the customs, VAT, trans-fer pricing and RegTech areas.

Risse: A simple example is the combinationof transfer pricing, the customs burden andVAT. These are three areas that were neverever put together before. You need a lot ofresources, time and money, to pull data outof the systems for these three areas to com-bine them. Now it is really feasible to opti-mise the customs burden by considering theright transfer prices, and to optimise thecash-positions for VAT purposes in terms oftransfer pricing and customs. Now we canoptimise all three areas.

Dewitz: There are classical conflicts ofinterest that may arise from the differentfunctions if they are not under one roof.Customs is interested in having low customsvalues to not pay too much for customs andexcise duties. Transfer pricing may have adifferent rationale. Transfer pricing is inter-ested in finding the right balance of profitattributions to the countries involved, andthat may sometimes be inconsistent. Forexample, if a company has low costs abroad,that is good for customs but it might shiftyour profit margin to different countries,and that is not in the interest of German taxauthorities. That is a classical conflict and AIcould help to resolve that. For example,when looking at the relevant benchmarkanalysis needed for transfer pricing reasonsin order to say, ‘ok, what is the right riskprofile of my counterpart abroad and whatdoes that imply for the acceptable or correctprofit margin?’ That is where AI could be oftremendous help to extract, collect andanalyse data, and then come up with pro-posals for the right profit margin systems. In the customs area we see another

aspect that is of strong interest. There aremany international free trade agreements,and they partially allow for preferential cus-toms treatment. You could have extraordi-nary low or zero custom rates, but thatdepends on the local content that yourgoods are composed of. Sourcing and pro-curement decisions within AUDI AG couldbe different if the information of the prefer-ential free trade agreement customs rate is athand in the moment you are taking thatdecision. There might be a supplier that hasan inferior price offer but that would comeup with the mandatory local content quali-

fication. In the total analysis of the valuechain this specific supplier might be benefi-cial to Audi when buying certain goods forthe composition of cars. In those fields AIcould be of tremendous help, even if it isjust in the first stage of analysis of free tradeagreements and correlations between thoseagreements.

What are the challenges in implementingmachine learning and AI? A lot of patterntype of functions can be automated butwill you still need human oversight? Dewitz: We are currently working with het-erogeneous IT systems in the AUDI AG,and that also holds true for the VolkswagenAG. We do not have one enterprise resourceplanning system, we have many, so it is dif-ficult to have a uniform approach to extractdata from those systems. In Germany we have grown to roughly

58,000 people. You need to identify andthen speak with a lot of different processparticipants and owners in order to get aprocess digitalised. Our focus is on those stages, the pure

digitalisation work that comes after that Iam not afraid of. That will be the easier partof the process. In the German market, the profile of

people who can do both – understand cus-toms and taxes and are quick in implement-ing and digitalisation of things – are very,very rare. The requirement for qualificationsin the future will be for people that areexperts in their specific areas but they willalso need to get along with new technolo-gies.

Risse: We have to learn what the machine isdoing. If you give the wrong data to themachine then you have garbage in, garbageout. We are still learning how to work withmachines, starting with robots and machinelearning and then deep learning. We are notat the level of deep learning yet; we are stilltrying to understand what machine learningmeans. This level however will offer manybenefits we would not have been got if wewere not to implement AI.

Tax authorities are rapidly upgrading inbig data analysis and machine learningsystems. What will be the impact of AIon tax audits and disputes? Esterer: AI increases tax planning in anon-aggressive way, a very legal and legiti-

mate way, while also creating compliancesecurity and certainty in a way you cannotachieve with existing technology. Togetherwith blockchain technology as example wecan create a lot of additional compliancecertainty that we did not have before.With a dashboard solution you can sim-

ulate different transfer pricing systems inthe various countries and look at what theoutcome is when you use TP model 1. Thatcould be a cost-plus model, resale minusmodel or a profit split model. So these arevery helpful instruments for tax planning intransfer pricing. Not just documentation, compliance or

benchmarking, but AI can also do tax plan-ning much faster much more efficient. Andit can prepare a decision on which TP sys-tem might be the appropriate one for thatcompany and which TP system leads to atax-efficient structure. It is important tooptimise your tax burden in a very transpar-ent way. You can show it to the tax author-ities. You can compare different pricingsystems and then use the right one for yourspecific business. That brings a lot of addi-tional value to the company and createshigh certainty when you talk to the taxauthorities. You do not need a person onbenchmarking now, it can be done by amachine on a global basis. Mid-sized, globally active companies

who have limited access to that data can dobenchmarking analysis much easier with AI.For those companies, the use of AI maybring additional value. As companiesbecome more international, more jurisdic-tions and tax authorities get involved.Globalisation drives that. What that meansfor companies is tax risk management,country-by-country reporting (CbCR),advance pricing agreements and transpar-ent TP systems. For example, CbCR can bedone with AI in a different way and it cancreate a real time information exchange.The goal is to improve tax risk managementthrough AI.

Risse: Specifically looking at TP documen-tation, with a good AI solution you canprove how the business model in a compa-ny is applied for tax purposes. Previously,when the tax auditor used to come in andsay ‘I would like to challenge you becauseyou applied the wrong transfer price, youwere not able to do this nitty-gritty docu-mentation, qualitatively yes but not on

Roundtable

www.tpweek.com 4

quantitative side’. Now we are capable ofdoing that. It will be significantly morecomplex for the tax authority to create adifferent fact pattern because the data treat-ment is so accurate that it will be very diffi-cult for them to conclude other findings.Transfer pricing is not a science, it is a

practical approach. We are driving thisback to science with high-quantitativedata accuracy.With the support of AI we can now tell

them that this is our business model andhere is all the data, and we can drill downinto the materiality of the data. We werenever able to do so before. The auditorcannot say ‘I do not trust your materialnumber’. Everybody now, to some extent, likes

to identify intangibles. At Henkel, weprint shoes with a 3D printing machine.Where is the intellectual property for that?Is it a software, is it in the machine or inthe ideas of printing? With these new busi-ness models – printing is a part ofGermany’s Industrie 4.0 – new questionswill occur. If you then look to the debateof BEPS Action 2, with this type of indus-try, where should the intangible be allo-cated to? I think this will changesignificantly. But what we are able to dowith AI is, we can tell the authoritieswhere our IP is situated in such a detailed,accurate way that we were never able tobefore. That is the big advantage.To some extent it seems that tax

authorities like to appoint as much com-pliance as possible to the companies to getthem busy with the intention to reducetime for tax planning. I guess we willcounteract with AI capabilities i.e. withcompliance automation.

Dewitz: Our understanding is that the taxauthorities in Germany certainly need tokeep up with state of the art technology butthey have already started. We as corporatesalso need to kick off these processesbecause if we do not do anything in thenext five years, we will be in a much weakerposition. They will be able to extract data,and through data pattern analysis theauthorities or auditors will be able to detectthe anomalies in the data lake and identifythose things that are not accounted for inthe correct manner from a tax perspective.So we need to be at least on par and weshould be better.

How will increasing automation impacton the trend of outsourcing of servicesand employees’ skill requirements? Risse: Six or seven years ago we started todedicate big chunks of our work to a sharedservice centre in Bratislava, Slovakia, whichnumbers roughly 100 FTEs today, and issplit between trade, customs and tax work.For the time being, they have standardisedprocesses for taxes and transfer pricing, andtrade and customs. They would be the firstwe could replace by machineries but notreally. AI tools are going to create newtypes of new abilities or capabilities for usthat will be incredible. So the work contentwill change in SSC. The SSC will help us toperform what is called Tax compliance infuture.Meanwhile, the tax group here at the

Dusseldorf headquarters will have to changetheir skills more towards technical expertise,too. Looking back at the example of com-bining transfer pricing, customs and VAT,people at the headquarters have to under-stand that with their AI enhanced capabili-ties and sub-processes, they will have anability to go beyond what we have today.

Esterer: In a few countries, some morethan others, research and preparatory workis already being done by AI rather thanhumans, for example in Australia, the UK,Canada and the US. There is partialreplacement of humans with computers interms of research because research can bedone much more efficiently with AI. AIcan also help you with preparation optimi-sation and in research. However, weshould view AI as an assistant system ratherthan a total replacement system. Thereplacement of more repetitive work will ofcourse take place and is already ongoing,not just in the tax area, but also in legalarea and the R&D replacing engineerswith machines. As a result, tax profession-als will have more time for specialist con-sulting and organisational tasks.

Dewitz: In general, the first field of imple-mentation of these new technologies willbe the mass data relevancy for VAT, wagetaxes and customs. In those areas youwould be in the perfect position to inte-grate compliance issues combined withefficiencies in the processes themselves. Wewould not need that many humans onthose work packages as we do now.

If they’re doing their job right and havethe processes right at the beginning, thenthe momentum that is partially seen in dif-ferent industries to have shared service cen-tres, will be driven down. There is a certaintrade-off between digitalisation and a sharedservice centre that you might want to havein India or the eastern part of Europe thatmay be cost driven. If certain processes aredigitalised the cost gap will not be thattremendous in the future. There will be lessneed to shift certain services to low-costcountries if you do digitalisation in the rightmanner. You would have the efficienciesthat you need at hand already.

What type of competencies do youexpect to achieve with AI in the future?Dewitz: In about five years we should havedone our homework regarding processes.These processes will have been analysed andreshaped to an ideal process that is then, atleast partially, digitalised. You reduce man-ual work and have those processes automat-ed. If I could make up the future, I wouldlike to have robotic process automationalready installed with AI helping to supportthe pure customs and tax analysis work.And all those things where you need toascertain a certain level of quality and com-pliance that is something machine learningor algorithms can do much better thanhumans. We would not like to havemachines do all of that work but if themachine is coming up with a proposal ofwhat the right accounting treatment wouldbe, that would tremendously drive downprocess time and shift up quality.

Risse: We are at the early beginning oftaking this machine learning into accountbut we see in the first results of machinelearning that the machine is able to learnhow procedures and processes are treatedin the accounting world and what the out-come should be. Everything that is rule-based can be

digitised. Structured data, this you cantake very far, which means the learningcurve at Henkel is significant. We are cur-rently programming 20 use cases. Whenwe started with the first three we werereally wondering what the benefit shouldbe. Now the benefits coming out of themachine work are so accelerated that peo-ple are really thinking differently aboutwhat is feasible.