Ihe expressed need. Due Io Ihe increasing imporIance oí service oííerings as a revenue source and increasing compeIiIion among serv- ice providers, iI is imporIanI íor companies Io opIimize boIh Ihe cusIomer experience as well as Ihe associaIed cosI oí providing Ihe service. Eor more complex inIeracIions wiIh higher val- ue, mixed-iniIiaIive sysIems provide an avenue IhaI gives a good balance beIween Ihe Iwo goals. This arIicle describes a mixed-iniIiaIive sysIem IhaI was creaIed Io improve cusIomer supporI íor problems cusIomers encounIered wiIh Iheir appliances. The Iool helped call Iak- ers solve cusIomers´ problems by suggesIing guesIions aiding Ihe diagnosis oí Ihese prob- lems. The mixed-iniIiaIive sysIem improved Ihe correcIness oí Ihe diagnosIic process, Ihe speed oí Ihe process, and user saIisíacIion. The Iool has been in use since 1ººº and has provided more Ihan $S0 million in íinancial beneíiIs by increasing Ihe percenIage oí guesIions IhaI could be answered wiIhouI sending a íield serv- ice Iechnician Io Ihe cusIomers´ homes. AnoIh- er mixed-iniIiaIive Iool, íor answering e-mail írom cusIomers, was creaIed in 2000. O ne oí Ihe more írusIraIing experiences in calling an auIomaIed answering sys- Iem íor iníormaIion or help is being íorced Io lisIen Io all Ihe menu opIions and Ihen being direcIed Ihrough Ihe sysIem, some- Iimes reguiring a redial (only Io sIarI again aI Ihe Iop menu opIion). These sysIems are raIher popular wiIh companies because Ihey save money÷Ihe companies´ money, IhaI is. How- ever, in a compeIiIive service indusIry, Ihese íully auIomaIed sysIems may noI be Ihe besI choice íor a service provider because cusIomer saIisíacIion plays an increasingly imporIanI role. Thereíore, more sophisIicaIed Iools are employed IhaI aIIempI Io balance Ihe need oí making Ihe cusIomer happy÷and perhaps increasing some direcI Iangible beneíiI aI Ihe same Iime. This is where mixed-iniIiaIive sys- Iems play an imporIanI role. General ElecIric (GE) Consumer & IndusIri- al´s Appliances Division manuíacIures and sells a wide range oí home appliances. GE´s sub- sidiary, Advanced Services Inc. (ASI), provides cusIomer service call cenIers IhaI help solve cusIomer issues over Ihe Ielephone and sched- ule íield service visiIs when needed. One oí Ihe Ielephone services ASI provides is a group oí more Ihan 300 íield service call Iakers who schedule íield service personnel Io visiI cus- Iomers´ homes. Around 1.4 million home vis- iIs are scheduled per year. The call Iakers´ pri- mary goal was Io schedule service visiIs, noI diagnose problems wiIh appliances over Ihe Ielephone. However, Ihey were able Io help Ihe cusIomer wiIhouI sending a íield service Iech- nician in 3.º percenI oí Ihe calls. There was an opporIuniIy Io increase Ihis percenIage because abouI 20 percenI oí Ihe Iime all Ihe íield service represenIaIive who visiIed Ihe home needed Io do was Io iníorm Ihe cusIomer abouI a parIicular seIIing or remedial acIion. Ií Ihis iníormaIion could have been obIained over Ihe Ielephone, iI would have saved Iime íor boIh Ihe cusIomer and Ihe íield service rep- resenIaIive. However, Iraining Ihe call Iakers Io diagnose appliances was diííiculI because oí Ihe large number oí appliances IhaI can be Artìcles SUMMER 2007 8º ApplIance Call Center: A SuccessIul MIxed- InItIatIve Case Study Wìllìam Cheetham anJ Kaì Goehel A| Magazlno volumo 28 Numbor 2 (2OOJ) (¹ AAA|) serviced, Ihe complexiIy oí modern appliances, and Ihe high Iurnover in Ihe call Iakers. The soluIion íor Ihis was Io creaIe a soíIware Iool IhaI acIs as a mixed-iniIiaIive assisIanI (Allen 1ººº) íor Ihe call Iakers called SupporI Ihe Cus- Iomer (STC). The STC sysIem is jusI one parI oí Ihe call- Iaking process. The íull process is shown in íig- ure 1. In Ihis process, Ihe cusIomer, in Ihe boI- Iom leíI, calls an 800 number provided íor scheduling home visiIs oí íield service Iechni- cians. A cusIomer call-Iracking sysIem (AgenI 1) receives Ihe call, accesses a cusIomer daIa- base Io reIrieve iníormaIion, such as Ihe cus- Iomer´s address and regisIered producIs, Ihen displays Ihis iníormaIion Io Ihe call Iaker who acIually answers Ihe Ielephone. STC is AgenI 2. Beíore STC was developed, Ihe call Iakers would answer Ihe guesIions Io Ihe besI oí Iheir abiliIy using Iheir personal knowledge, Irain- ing, paper manuals, íreguenIly asked guesIions lisIs, and weekly paper ílyers on new issues. STC uses a case base, a rule base, and a decision Iree Io assisI Ihe call Iakers in helping Ihe cus- Iomers. STC sIores cases oí problems and Iheir soluIions, a decision Iree oí guesIions IhaI are used in a diagnosIic process Io diííerenIiaIe Ihe acIual case írom all oIhers, and rules IhaI can auIomaIically answer guesIions. AíIer Ihe appliance has been diagnosed, Ihe resulIs oí Ihe diagnosis are sIored in Ihe call-scheduling sysIem, AgenI 3. The call-scheduling sysIem creaIes a planned rouIing íor Ihe íield service Iechnician Ihe nighI beíore Ihe service call is Io Iake place and suggesIs parIs Io sIock on Ihe repair person´s Iruck. The call-Iaking sysIem also sends daIa abouI Ihe cusIomer Io STC, and Ihe íield service Iechnician can provide íeed- back on Ihe success or íailure oí Ihe service vis- iI Io STC. The nexI secIion discusses STC in more deIail. Then Ihe mixed-iniIiaIive issues oí STC are described. The íinal secIion gives resulIs írom mulIiple years oí usage íor STC. Artìcles º0 AI MAGAZINE keLrieves user sLaLisLics, appliance, and warranLy lnformaLion Provides mixed-iniLiaLive decision supporL Schedules field visiL Agent T Call Center 0perator Customer Fleld Technlclan Agent 2 Agent 3 Call Tracklng Ingelllgent 5oftware Call 5chedullng lìçure 1. Customer Support lrocess. SupporI Ihe CusIomer ApplicaIion The call-cenIer Iool is designed Io help call Iak- ers solve cusIomers´ problems by suggesIing guesIions IhaI can be used Io diagnose Ihe cus- Iomers´ problems. The primary goals oí Ihe call Iaker are Ihe correcI diagnosis oí Ihe appliance, speed oí diagnosIics, and cusIomer saIisíacIion. The purpose oí Ihe soíIware Iool is Io improve meIrics associaIed wiIh Ihese goals. In order Io do Ihis, Ihere were mulIiple reguiremenIs íor Ihe soíIware Iool. These reguiremenIs are sim- ilar Io Ihe principles oí mixed-iniIiaIive user inIeríaces described by Eric HorviIz (1ººº). The reguiremenIs are (1) suggesI correcI resoluIions Io cusIomer issues, (2) suggesI guesIions IhaI will diagnose Ihe problem, (3) increase Ihe speed oí Ihe diagnosis process, (4) increase user saIisíacIion wiIh Ihe process, (S) explain Io Ihe user why someIhing was suggesIed, and (ô) have Ihe Iool learn írom iIs experiences and be able Io adapI Io a changing environmenI. RelaIed Work and Tool Usage Case-based reasoning (CBR) (AamodI and Flaza 1ºº4) has been used Io auIomaIe cusIomer- supporI help desks by many companies such as Compag (Acorn and Walden 1ºº2) and Broder- bund (WaIson 1ºº7). Using CBR íor cusIomer supporI (Simoudis 1ºº2) has become oí inIer- esI Io many oIher companies wiIh help desks. We used a CBR Iool Io assisI GE Consumer FroducIs cusIomer-supporI personnel. STC was creaIed using a CBR Iool írom Iníerence Cor- poraIion called k-commerce. (Iníerence was laI- er acguired by eGain.) K-commerce allowed íor mixed iniIiaIive beIween Ihe call Iaker and Ihe auIomaIed assisIanI. II provided sIandard user inIeríaces íor Ihe knowledge engineer and end user (IhaI is, call Iaker). The knowledge engi- neer inIeríace included íorms íor creaIing cas- es and rules. During projecI developmenI, Iníerence CorporaIion also released a Iool íor creaIing decision Irees. These Iools greaIly reduced Ihe developmenI Iime íor Ihe projecI and allowed us Io íocus on knowledge acguisi- Iion insIead oí Iool creaIion. The k-commerce íorm íor creaIing a new case allows Ihe knowledge engineer Io enIer a IiIle, descripIion, mulIiple guesIions, and a res- oluIion íor Ihe case. The sysIem selecIs Ihe mosI relevanI cases using Iwo criIeria. The íirsI criIerion is Ihe degree oí maIch beIween Ihe IexI descripIion oí Ihe problem Iyped in by Ihe call Iaker and Ihe IexI in Ihe IiIle and descrip- Iion oí Ihe case. The second criIerion is Ihe per- cenIage oí guesIions in Ihe case IhaI have been answered correcIly. When Ihe diagnosIic process is sIarIed (beíore any guesIions have been answered) Ihe selecIion oí Ihe mosI rele- vanI cases is based on Ihe IexI maIch. Then, as guesIions írom Ihe highesI maIching cases are suggesIed and answered, Ihe correcI answers play a larger roll in deIermining Ihe mosI rele- vanI cases. The appropriaIe case is selecIed by answering Ihe guesIions IhaI diííerenIiaIe Ihe possible cases. Eor our applicaIion we wanIed Io be able Io suggesI one and only one case as Ihe soluIion. A way Io guaranIee IhaI Ihese guesIions can diííerenIiaIe every case in Ihe case base is Io íorm a Iree oí guesIions over Ihe case base where each case is a leaí in Ihe Iree and each inIernal node in Ihe Iree is a gues- Iion. We consider Ihis approach as a hybrid approach IhaI employs boIh decision Iree and CBR meIhodology where Ihe decision Iree is simply Ihe selecIion mechanism oí Ihe CBR sysIem. Some problems can be represenIed by cases wiIhouI creaIing a decision Iree (EAQs are an example oí Ihese), buI íor oIhers iI is beIIer Io have a decision Iree creaIed íor Ihem (any- Ihing wiIh a signiíicanI diagnosIic process used Ihe decision Iree approach). Eigure 2 shows a porIion oí a decision Iree where recIangles are guesIions, Ihe arrows are answers, and ovals are cases. K-commerce also allows Ihe creaIion oí rules IhaI can auIomaIically answer guesIions. An example oí a rule is, ií Ihe answer Io Ihe gues- Iion "WhaI Iype producI is your appliance7" is "ReírigeraIor" and Ihe Ihird and íourIh charac- Ier in Ihe model number are boIh "S" Ihen Ihe answer Io Ihe guesIion "WhaI Iype oí reíriger- aIor do you have7" is "Side-by-Side." User InIeríace The currenI inIeríace íor STC is shown in íigure 3. The criIical iníormaIion Iab aI Ihe Iop has Ihe producI line and model number, which are boIh passed in írom Ihe call-Iaking sysIem. II also has Ihe problem descripIion, which is Iyped in by Ihe call Iaker. The model group can be deIermined by a rule IhaI uses Ihe model number. The symptom ìs a keyword phrase IhaI is selecIed by Ihe call Iaker. The guesIions Iab has a seI oí guesIions Ihe call Iaker can ask Ihe cusIomer Io diagnose Ihe problem. The resulIs Iab has a seI oí soluIions. SelecIing Ihe correcI resulI is Ihe goal oí Ihe process. ApplicaIion DevelopmenI The developmenI oí Ihe STC sysIem was a íive- sIep process. These same íive sIeps can be íol- lowed íor deploying oIher applicaIions oí AI: íirsI, sIandardize Ihe process and knowledge, second, digiIize Ihe inpuIs and ouIpuIs, Ihird, Artìcles SUMMER 2007 º1 auIomaIe Ihe process as much as possible, íourIh, conIrol Ihe gualiIy oí Ihe sysIem, and íiíIh, leverage Ihe sysIem and knowledge íor improved impacI. SIandardizing Ihe call-Iaking process involved idenIiíying all cases IhaI can and can- noI be solved on Ihe Ielephone and deIermin- ing Ihe correcI guesIions Io ask and Ihe correcI order in which Io ask Ihem. Much oí Ihis knowledge was IaciI knowledge (IhaI is, per- sonal experience oí Ihe call Iakers and engi- neers). Since diííerenI call Iakers would ask dií- íerenI guesIions Io diagnose Ihe same problem, we íormed Ieams oí call Iakers and engineers Io deIermine whaI should be Ihe correcI cases, guesIions, and order oí guesIions. The visual- izaIions oí Ihe decision Irees were very useíul in reviewing and opIimizing Ihis knowledge. The knowledge was enIered inIo Ihe case base only aíIer Ihe engineers and call Iakers creaIed Ihese visualizaIions and saíeIy, legal, gualiIy, and consumer-service personnel validaIed Ihem. This Iook Ihe majoriIy oí Ihe Iime need- ed Io creaIe Ihe STC sysIem. DigiIizing Ihe inpuIs and ouIpuIs oí Ihe STC sysIem was Ihe nexI sIep. The call-Iaking sys- Iem was modiíied Io sIarI up Ihe STC sysIem and send relevanI daIa (íor example, model number) íor every Ielephone call. K-commerce was modiíied slighIly so IhaI iI could accepI Ihe daIa. The call-scheduling agenI, írom íig- ure 1, was modiíied Io accepI a large amounI oí daIa írom Ihe STC sysIem in addiIion Io Ihe daIa iI was receiving írom Ihe call Iaker. The STC daIa was capIured every Iime Ihe case base was used Io answer a call. The daIa included sIarI Iime, end Iime, cusIomer Ielephone num- ber, call-Iaker ID, Iype oí appliance, a shorI IexI descripIion oí Ihe issue, all guesIions asked by CBR, all answers given Io Ihese guesIions, Ihe case suggesIed, and wheIher Ihis answer was accepIed by Ihe cusIomer. AuIomaIing Ihe enIire Ielephone conversa- Iion was noI possible wiIh Ihe sIaIe oí naIural language processing Iechnology, buI a Ieam oí a call Iaker working wiIh Ihe STC sysIem can eííecIively auIomaIe Ihe applicaIion oí Ihe sIandardized and opIimized process. An aIIempI was made Io auIomaIe Ihe applicaIion as a web-based cusIomer selí-service Iool. This was discussed in CheeIham (2003). ConIrolling Ihe gualiIy oí Ihe sysIem aíIer iI was creaIed was more expensive Ihan Ihe ini- Iial creaIion buI also conIinued Io improve Ihe success raIe. IniIially, each week a case auIhor would analyze Ihe daIa in Ihe call-scheduling sysIem íor every call IhaI was Iaken íor Ihe week using sIandardized daIabase gueries. The auIhor was looking íor any Irends and espe- cially any Iimes a caller would noI be saIisíied by a suggesIion. Any Irends or ouIsIanding iIems would be discussed in a weekly íeedback meeIing wiIh Ihe call Iakers. The resulI oí Ihe íeedback meeIings would be a íew changes in Ihe case base, decision Irees, and rules. These changes would be made immediaIely and reviewed in Ihe nexI week´s meeIing. The íre- guency oí Ihese reviews decreased as Ihe case base, decision Irees, and rules sIabilized. Leveraging Ihe STC sysIem, knowledge in Artìcles º2 AI MAGAZINE Which compartment is too coId? lresh lood lreezer What is freezer temperature setting? <= 3 > 3 Schedule service Lower LemperaLure, waiL 24 hours lìçure 2. Oecìsìon Tree jor ´Rejrìçerator too ColJ." Ihe case base, and daIa in Ihe call-record daIa- base provided addiIional beneíiIs IhaI were noI all íoreseen when Ihe íirsI version oí STC was creaIed. One beneíiI was Ihe abiliIy Io auIo- maIe Ihe response Io e-mail. II Iakes longer íor a person aI Ihe call cenIer Io answer an e-mail message Ihan a Ielephone message, and in 2000 GE´s e-mail volume was growing rapidly. We used Cisco´s e-mail manager Iool Io sIore replies Io íreguenIly asked guesIions so IhaI Ihey can be reused on íuIure similar guesIions. In order Io selecI Ihe appropriaIe reply, a seI oí decision Irees was creaIed where each node in Ihe Iree is a rule abouI meIadaIa regarding Ihe e-mail. Each leaí node in Ihe decision Iree speciíies a seI oí replies IhaI can be used Io answer IhaI speciíic Iype oí e-mail message. These Irees are similar buI noI idenIical Io Ihe Irees used by STC. The e-mail handler does noI auIomaIically answer e-mail, iI suggesIs previ- ous replies IhaI can be reused íor a person who Ihen jusI selecIs Ihe correcI response írom a lisI oí suggesIions. The person can ediI Ihe sug- gesIed reply, creaIe a new one, or send Ihe sug- gesIed reply unchanged wiIh one buIIon click. Eigure 4 shows a case írom Ihe e-mail man- ager. A cusIomer e-mailed GE saying Ihe cus- Iomer had a S0-year-old reírigeraIor in períecI working condiIion and asked ií we would like Io purchase iI. The reply Io Ihis e-mail was sIored in Ihe case base and is now used abouI Iwo Iimes per week. CusIomer-supporI per- sonnel eiIher send Ihis exacI reply or ediI iI ií Ihe reguesI was slighIly diííerenI. UndersIanding a íree IexI e-mail message is diííiculI íor an auIomaIed sysIem, so we had Ihe decision Iree use only meIadaIa IhaI was senI wiIh Ihe e-mail. The meIadaIa was cre- aIed by noI allowing cusIomers Io send e- mail direcIly Io us, we made Ihem enIer Ihe iníormaIion on a íorm on our web siIe. All e- mail addresses on our producIs were replaced wiIh web pages. In addiIion Io Ihe íree IexI oí Ihe e-mail, cusIomers needed Io classiíy Iheir messages based on a íew pull-down menus on Ihe web siIe and provide a Ielephone number. This selí-classiíicaIion oí Ihe e-mail creaIed Ihe meIadaIa IhaI made iI much eas- Artìcles SUMMER 2007 º3 lìçure J. Call Taler User lnterjace. ier íor an auIomaIed sysIem Io selecI similar cases. The e-mail handling and STC Iool work IogeIher. Ií we deIermine an e-mail would be beIIer answered wiIh a Ielephone call Ihen we can use STC Io help Ihe cusIomer. A new group oí cusIomer-supporI personnel was creaIed. They do noI answer calls. They call Ihe cus- Iomers Ihemselves. We keep Irack oí Ihe suc- cess raIe íor each producI}naIure-oí-problem combinaIion so we know Ihe predicIed chance oí solving Ihe problem over Ihe Ielephone. Ií Ihe cusIomer does noI answer Ihe Ielephone aI our íirsI aIIempI, we can repeaIedly Iry Io con- IacI cusIomers wiIh a high chance oí success using callback beíore replying wiIh an e-mail. Mixed-IniIiaIive Issues oí STC STC is a mixed-iniIiaIive conversaIional CBR sysIem IhaI acIs as an inIelligenI assisIanI íor Ihe call Iaker. TogeIher, STC and Ihe call Iaker íorm a Ieam in which each provides a needed sIrengIh where Ihe oIher has a weakness. This secIion describes how each mixed-iniIiaIive issue (Tecuci, Boicu, and Cox 2007) was addressed in Ihe STC applicaIion. Task Issue The Iask issue addresses Ihe division oí respon- sibiliIy beIween Ihe human and Ihe agenIs. To decide which Iasks humans should do and which inIelligenI agenIs should do, one should deIermine where humans are more appropriaIe Ihan inIelligenI agenIs and where agenIs have advanIages over humans. Humans are hard Io beaI aI common sense, inIeracIing wiIh oIher humans, and creaIiviIy. InIelligenI agenIs, on Ihe oIher hand, are good aI many Ihings IhaI humans íind cumbersome or Iedious, such as íollowing a precise process, sIoring and reIriev- ing large guanIiIies oí iníormaIion, and maIh- emaIical compuIaIions. While humans have a moderaIe cosI Io Irain and a moderaIe Io high cosI Io períorm a Iask, agenIs have a high cosI Io program íor a Iask buI a low cosI Io imple- menI IhaI program. Tasks can be períormed by a human (wiIh- ouI an agenI), an agenI (wiIhouI a human), or a human and an agenI. Beíore compuIers, all Iasks were períormed wiIhouI an agenI. In Ihe recenI pasI, AI has been used Io creaIe sysIems in which an agenI operaIes wiIhouI a human. However, Ihere are many Iasks IhaI can beneíiI írom a human períorming a porIion oí Ihe Iask and an agenI períorming a diííerenI porIion oí Ihe Iask. Tasks like Ihis are good candidaIes íor mixed-iniIiaIive sysIems. In a call cenIer, Ihe Iask oí helping cusIomers reguires inIeracIion wiIh Ihe cusIomer and common sense buI also adherence Io a process and accessing a large amounI oí iníormaIion in a Iimely íashion. The call Iakers are good aI Ihe naIural language processing IhaI is needed Io inIeracI wiIh Ihe cusIomers, buI iI is diííiculI íor Ihem Io sIore and correcIly reIrieve Ihe large amounI oí Iechnical iníormaIion IhaI is needed Io help Ihe cusIomer. Luckily, STC is very good aI sIoring and reIrieving Ihis iníor- maIion even Ihough iI cannoI do any naIural language processing. The user (call Iaker) per- íorms Ihe inIeracIions wiIh Ihe cusIomer, and Ihe agenI (STC) sIores and reIrieves Ihe sIan- dardized knowledge abouI diagnosing appli- ances. Beíore Ihis sysIem was íielded, much oí Ihe iníormaIion needed Io diagnose a problem was in manuals and IaciI knowledge oí a íew call Iakers. CusIomers would oíIen be placed on hold while Ihe call Iakers looked íor Ihe cor- recI manual and Ihen searched íor Ihe perIi- nenI iníormaIion. Having an agenI IhaI could sIore Ihis iníormaIion, provide iI auIomaIical- ly, and guide Ihe call Iaker in Ihe diagnosis removed many írusIraIing delays. AnoIher Iask IhaI Ihe agenI can períorm is auIomaIically answering Ihe guesIions IhaI iI is coníidenI iI can answer correcIly. Some oí Artìcles º4 AI MAGAZINE lìçure 4. L-maìl Lxample Case. Ihese guesIions include iníormaIion IhaI can be obIained írom oIher sources Ihan Ihe cus- Iomer. IníormaIion abouI Ihe cusIomer, such as models owned and previous diagnosIics ses- sions, can be sIored in a daIabase. When gues- Iions reguire Ihis iníormaIion, iI can be answered auIomaIically. One íuIure goal íor high-end appliances is Io include a Ielephone or EIherneI connecIion so addiIional diagnos- Iic daIa, such as measuremenIs írom a Iemper- aIure sensor, can be senI direcIly Io Ihe diag- nosIic Iool, allowing iI Io Iake iniIiaIive on a wider seI oí guesIions. ConIrol Issue The conIrol issue addresses Ihe shiíI oí iniIia- Iive and conIrol beIween Ihe human and Ihe agenIs. This primarily íocuses on how Ihe agenIs should show iniIiaIive, including proac- Iive behavior. Humans oíIen show iniIiaIive and can generally sIrike a good balance in judg- ing when iniIiaIive will be useíul and when iI will noI be useíul. To IhaI end, humans con- sider a range oí guiding principles including: (1) Have an undersIanding oí Ihe goals and pri- oriIies. (2) Have an undersIanding oí Ihe cur- renI siIuaIion. (3) IdenIiíy a Iask írom Ihe cur- renI siIuaIion IhaI can help goals. (4) IdenIiíy poIenIial beneíiIs and problems írom doing a Iask. (S) DeIermine Ihe coníidence aI which a Iask should be períormed (high÷do Iask, low÷do noI do Iask, and medium÷ask ií Iask should be done). (ô) Have Ihe abiliIy Io per- íorm a Iask. (7) Iníorm oIhers IhaI a Iask has been done. IIems one, Iwo, and Ihree are needed Io cor- recIly idenIiíy opporIuniIies íor Iaking iniIia- Iive. IIem íour gaIhers iníormaIion abouI Ihe advanIages and disadvanIages oí Iaking Ihe ini- IiaIive. Care musI be Iaken Io idenIiíy Ihe dis- advanIages. Fossible disadvanIages include boIhering oIhers, doing an unwanIed Iask, doing a low-prioriIy Iask insIead oí a high-pri- oriIy one, and being desIrucIive or wasIeíul. The íiíIh iIem is weighing Ihe advanIages and disadvanIages. Ií Ihe advanIages greaIly ouIweigh Ihe disadvanIages Ihen Ihe iniIiaIive should mosI likely be Iaken. Ií Ihe disadvan- Iages ouIweigh Ihe advanIages Ihen Ihe iniIia- Iive should noI be Iaken. Ií neiIher Ihe advan- Iages nor Ihe disadvanIages ouIweigh Ihe oIher Ihen a human can ask anoIher person (íor example, a supervisor) ií Ihe iniIiaIive should be Iaken. The human needs Ihe abiliIy Io per- íorm Ihe Iask and should iníorm oIhers aííecI- ed IhaI Ihe Iask has been compleIed. SoíIware agenIs can use Ihe same seI oí guidelines when Ihey aIIempI Io show iniIia- Iive. They should undersIand Ihe goals, priori- Iies, currenI siIuaIion, and seI oí possible ini- IiaIive acIions. They should be able Io idenIiíy advanIages and disadvanIages. DisadvanIages oí an auIomaIed assisIanI Iaking inappropriaIe iniIiaIive include boIhering Ihe user, locking Ihe user ouI (sIealing cycles), acIing in an unknown way, undoing Ihe user´s desired acIions, making Ihe user undo Ihe agenI´s acIions, and keeping Ihe user írom doing Iasks. The agenI should be able Io deIermine Ihe coníidence IhaI Ihe iniIiaIive should be Iaken (CheeIham and Frice 2004). Then, Ihe agenI execuIes Ihe acIion wiIh highesI coníidence and iníorms Ihe user oí Ihe acIion Iaken. Since speed oí call is a reguiremenI in STC, we do noI wanI Io have Ihe call Iaker ever waiI íor Ihe agenI. All oí Ihe agenI´s acIions Iake place so guickly IhaI Ihey appear Io be insIan- Ianeous Io Ihe call Iaker. This is also imporIanI íor Ihe cusIomer´s saIisíacIion, because we do noI wanI Io have Ihe cusIomer waiI íor Ihe agenI. Einally, Ihe acIion suggesIed aíIer diagnosis is compleIed can be auIomaIed. The call Iaker could sIep Ihe user Ihrough a precreaIed repair process or e-mail or íax Ihe process Io Ihe user. Ií parIs are reguired, Ihe agenI can place an order íor Ihe parIs Io be senI Io Ihe address oí Ihe cusIomer. Ií a service Iechnician is reguired, Ihe Iime íor IhaI visiI can be scheduled and a descripIion oí Ihe problem auIomaIically senI Io Ihe call-scheduling sysIem. Awareness Issue The awareness issue addresses Ihe mainIenance oí a shared awareness wiIh respecI Io Ihe cur- renI sIaIe oí Ihe human and agenIs. This can be diííiculI. The humans and agenIs have com- pleIely diííerence senses. Humans have sighI, hearing, Iouch, IasIe, and smell. Any iníorma- Iion gaIhered wiIh Ihese senses will need Io be relayed Io Ihe agenI. AgenIs have access Io dig- iIized daIa sources such as sensors or daIabases. Any perIinenI iníormaIion discovered by Ihe agenI needs Io be relayed Io Ihe human. Humans have experiences and knowledge IhaI can be used Io make conclusions írom exisIing iníormaIion. AgenIs may include an experI sys- Iem IhaI also makes conclusions, mosI likely in a diííerenI way írom Ihe human. One good way Io keep a shared awareness is Io share Ihree Iypes oí iníormaIion: íacIs and belieís, reasoning, and conclusions. To obIain a clear undersIanding oí an agenI´s awareness, all Ihree iníormaIion Iypes are needed by Ihe human. Showing íacIs and belieís allows Ihe human Io check íor any IhaI are noI correcI, currenI, or consisIenI wiIh Ihe human´s íacIs and belieís. The conclusions are Ihe value IhaI Artìcles SUMMER 2007 ºS Ihe agenI produces and need Io be shown Io Ihe human. The reasoning is Ihe bridge beIween Ihe íacIs and belieís and Ihe conclu- sion. II can be diííiculI Io share Ihe reasoning oí black-box Iechnigues, such as neural neIs. Thereíore, black-box Iechnigues should be used in mixed-iniIiaIive sysIems only when Iheir reasoning does noI need Io be inspecIed (íor example, Ihis process does noI need Io be audiIed or is guaranIeed Io be correcI). Two oIher rules oí Ihumb on having a shared aware- ness are (1) do noI make Ihe agenIs guess aI Ihe siIuaIion, as incorrecI assumpIions can cause many problems in a mixed-iniIiaIive sysIem, and (2) allow Ihe human Io inspecI and change Ihe agenI´s íacIs and belieís. Eor STC, in order íor Ihe agenI Io be able Io make valid suggesIions, iI needs Io have aware- ness oí all Ihe iníormaIion IhaI Ihe call Iaker has abouI Ihe problem. This reguires Ihe STC sysIem Io have all possible guesIions IhaI Ihe call Iaker can ask and all possible answers Io Ihose guesIions. We guickly íound IhaI each guesIion needed an answer oí "unknown" and IhaI Ihe agenI needed Io be able Io correcIly deal wiIh Ihese answers. Eor insIance, Ihe cus- Iomer may noI be able Io answer Ihe guesIion "WhaI is Ihe model number oí your appli- ance7" This guesIion also led Io Ihe creaIion oí a model number locaIor agenI IhaI can be used Io locaIe Ihe model number on any appliance. The user musI also have awareness abouI whaI Ihe agenI is doing. Since Ihe agenI can Iake Ihe iniIiaIive Io answer guesIions, Ihe user musI be able Io inspecI Ihe conclusions IhaI Ihe agenI has made. In STC, all guesIions IhaI Ihe agenI has answered auIomaIically are lisIed íor Ihe agenI so Ihey can be easily reviewed, buI answering Ihem does noI inIerrupI Ihe user. II is also imporIanI IhaI Ihe call Iaker be able Io change Ihe answers when a guesIion is answered auIomaIically. The user can click on Ihe guesIion IhaI Ihe agenI has answered and Ihen selecI a diííerenI answer. A sysIem admin- isIraIor can also inspecI Ihe íacIs and rules IhaI caused a guesIion Io be auIomaIically answered. CommunicaIion Issue The communicaIion issue addresses Ihe proIo- cols IhaI íaciliIaIe Ihe exchange oí knowledge and iníormaIion beIween Ihe human and Ihe agenIs. These proIocols include mixed-iniIia- Iive dialog and mulIimodal inIeríaces. The communicaIions írom human Io agenI in a mixed-iniIiaIive sysIem need Io be as eííi- cienI as possible íor Ihe human and as com- pleIe as possible íor Ihe agenIs. These are oíIen compeIing goals. The human has limiIed Iime Io updaIe Ihe awareness oí Ihe agenIs. One goal oí a mixed-iniIiaIive sysIem is oíIen Io save Iime íor Ihe human and having Ihe human spend a large amounI oí Iime or eííorI updaIing Ihe agenIs´ awareness deíeaIs Ihis purpose. However, Ihe agenIs need as compleIe an awareness oí Ihe siIuaIion as possible in order Io provide assisIance Io Ihe human. The communicaIion írom Ihe agenIs Io Ihe human needs Io be easy and unobIrusive íor Ihe human, buI Ihe human also needs Io have access Io all iníormaIion IhaI could be valu- able. Ií Ihe communicaIion is noI easy and unobIrusive íor humans, Ihey will be unlikely Io use Ihe sysIem because Ihe beneíiIs Io Ihem do noI ouIweigh Ihese cosIs. One way Io make Ihe communicaIions less obIrusive is Io reduce Ihe íalse alarm raIe (IhaI is, Ihe percenIage oí communicaIions when Ihe agenI needlessly communicaIes Io Ihe human). These needless communicaIions jusI boIher Ihe human. One real deployed example oí possibly boIhersome communicaIion is a voice-based warning sys- Iem íor Ihe E-1ô combaI aircraíI. The íemale voice Iells Ihe piloI Io "pull up" írom a dive or "warning, warning" when Ihe piloI needs Io pay aIIenIion Io various panel displays (Ihe nagging íemale voice goI nicknamed "BiIching BeIIy" by piloIs). Eor STC, Ihe primary communicaIion írom Ihe agenI Io Ihe user is a suggesIed guesIion. II would be bad ií Ihere was only one guesIion suggesIed and iI was noI correcI íor Ihe given siIuaIion. MulIiple meIhods can be used Io reduce Ihe cosI oí an incorrecI suggesIed gues- Iion. Some oí Ihese are providing mulIiple sug- gesIions írom which Ihe user can pick, noI íorcing Ihe user Io acI on Ihe suggesIion, and unobIrusively presenIing Ihe suggesIed gues- Iions Io Ihe call Iaker. The primary communicaIions írom Ihe user Io Ihe agenI are Ihe answers selecIed íor Ihe guesIions. The seI oí answers Io a guesIion is creaIed so IhaI each answer is appropriaIe íor a diííerenI case. Eor insIance, we do noI ask íor answers IhaI are conIinuous variables like Iem- peraIure. We do ask íor ranges oí IemperaIures where each range would have a meaning asso- ciaIed wiIh iI. This way no posIprocessing is needed Io Iurn Ihe answer inIo a diagnosIic íacI. Call Iakers and cusIomers oíIen wonder why Ihe sysIem is suggesIing a speciíic guesIion. User IrusI is enhanced ií Ihere is a clear expla- naIion íor why Ihe sysIem is Iaking some acIion. When Ihe guesIions are creaIed, we oíIen also creaIe an explanaIion íor why Ihe guesIion should be asked. This explanaIion can be displayed íor Ihe call Iaker by clicking on Artìcles ºô AI MAGAZINE Ihe guesIion in Ihe user inIeríace. CreaIing Ihese explanaIions in CBR sysIems can be dan- gerous. One projecI Io creaIe auIomaIed expla- naIions produced explanaIions IhaI were noI always appropriaIe íor Ihe cusIomer. Eor exam- ple, Ihe explanaIion íor why we are asking Ihe guesIion "Is Ihe power cord plugged in7" was "We are Irying Io coníirm Ihe problem cause 'IdioI User.´" EvaluaIion Issue The evaluaIion issue addresses Ihe human and agenI conIribuIion Io Ihe overall sysIem´s per- íormance. EvaluaIion oí a mixed-iniIiaIive sys- Iem involves evaluaIion oí Ihe users, agenIs, and Ihe agenI}user inIeracIion. One íield oí sIudy íor evaluaIing user (IhaI is, employee) períormance is called human resource man- agemenI (HRM). The basic premise oí HRM is IhaI humans should noI be IreaIed like machines. In Ihe evaluaIion oí an agenI, iI would also be good Io evaluaIe Ihe agenI less like a machine and more like a person. In a company seIIing, HRM evaluaIes employees on Iwo major areas: Iheir individual abiliIies and how Ihey work wiIh oIhers. The agenI should be evaluaIed on boIh iIs individual abiliIies and how iI inIeracIs wiIh Ihe user. An agenI will noI be as eííecIive as possible unless iI is compeIenI in boIh oí Ihese areas. The individ- ual eííecIiveness oí Ihe agenI can be measured wiIh guanIiIaIive indicaIors such as speed (sec- onds), coverage (number oí Iasks Ihe agenI can compleIe), íalse alarm raIe, and accuracy oí resulIs. However, Ihe gualiIy oí iIs user inIerac- Iions will probably need Io be evaluaIed wiIh gualiIaIive indicaIors such as user saIisíacIion, ease oí inIeracIion, IrusIworIhiness, íriendli- ness, manners, and value provided Io Ihe user. A mixed-iniIiaIive sysIem can be evaluaIed on iIs speed, cosI, and correcIness in períorm- ing iIs Iasks. DiííerenI sysIems will place diííer- enI imporIance on each oí Ihese issues. In a mixed-iniIiaIive sysIem Ihere is a guesIion oí who will do each Iask, Ihe user or Ihe agenI. The simple answer Io Ihis is Io have Ihe one wiIh Ihe lowesI cosI, highesI speed, and high- esI correcIness períorm Ihe Iask. The agenI almosI always has Ihe highesI speed, so iI would be good Io íocus on how Io have Ihe agenI also have a low cosI and high correcI- ness. The cosI oí Ihe agenI períorming a Iask íor Ihe user can be IhoughI oí as Ihe sum oí Ihe cosIs oí Ihe íollowing íive acIiviIies: (1) creaI- ing Ihe agenI, (2) Ihe user Ielling Ihe agenI Io do Ihe Iask, (3) Ihe agenI doing Ihe Iask, (4) Ihe agenI communicaIing Ihe resulI Io Ihe user, and (S) Ihe user receiving and validaIing Ihe resulI oí Ihe agenI. All oí Ihese cosIs should be reduced Io make Ihe agenI as useíul as possible. A simple way Io have Ihe agenI be as correcI as possible is Io have iI only do Iasks (or speciíic insIances oí a given Iask) where iI is likely Io be correcI. Eigure S shows a generalized ílow oí conIrol Ihrough a mixed-iniIiaIive sysIem. The process can sIarI and go Io eiIher a user or an agenI. Then users and agenIs will work Ihe Iask unIil eiIher Ihe user or Ihe agenI ends Ihe Iask. Some mixed-iniIiaIive sysIems may reguire eiIher Ihe user or agenI Io sIarI and end Ihe Iask. One exIreme oí Ihis ílow oí conIrol is Io have Ihe user sIarI Ihe Iask, work on Ihe Iask wiIhouI help írom Ihe agenI, and Ihen end Ihe Iask. This is Ihe way many Iasks are done beíore Ihe mixed-iniIiaIive sysIem is creaIed. The oIher exIreme is Io have an agenI sIarI Ihe Iask, work on Ihe Iask wiIhouI help írom Ihe user, and Ihen end Ihe Iask. This is how many IradiIion- al AI sysIems operaIe. This would usually be Ihe lowesI cosI and highesI speed approach, buI Ihere are many domains where Ihe correcI- ness oí such a sysIem would noI be accepIable. So, Ihe proper Irade-oíí íor a mixed-iniIiaIive sysIem is Io have Ihe agenI do all oí Ihe Iasks where iI can have an accepIable correcIness aI a cosI IhaI is lower Ihan Ihe cosI oí Ihe user períorming Ihe Iask. This would move as much oí Ihe Iask as possible away írom Ihe user and Io Ihe agenI. One way Io evaluaIe job períormance írom people is wiIh a Ihree-level scale: (1) Foor÷ does noI íinish Iasks correcIly or on Iime. (2) Good÷íinishes Iasks correcIly and on Iime. (3) ExcellenI÷íinishes Iasks correcIly and on Iime plus shows iniIiaIive Io suggesI or do oIher Iasks. AgenIs could be evaluaIed on a similar scale. This would reguire iniIiaIive írom Ihe agenI in order íor iI Io be an excellenI agenI. WiIh STC, we íound IhaI call Iakers guickly losI IrusI in Ihe sysIem ií iI gave poor suggesI- ed guesIions or iniIiaIive. ThaI is why we wenI Ihough Ihe addiIional eííorI oí creaIing a deci- sion Iree on Iop oí Ihe case base. K-commerce per se does noI reguire a decision Iree, iI can auIomaIically suggesI guesIions írom Ihe mosI likely cases. However, aI Ihe sIarI oí Ihe diag- nosis process, iI is oíIen unclear which cases are Ihe mosI likely ones. We creaIed diagnosIic Irees íor each possible sympIom oí each Iype oí appliance Io improve Ihe correcIness oí Ihe sys- Iem and reIain Ihe users´ IrusI. Eigure 4 shows Ihe Iree íor a side-by-side reírigeraIor where a comparImenI is Ioo cold. The cusIomers´ saIisíacIion is primarily based on Ihe correcIness, speed, and proíes- sionalism oí Ihe help provided by Ihe call Iak- Artìcles SUMMER 2007 º7 er. We have already described how Ihe sysIem helps Ihe correcIness and speed oí Ihe process. The sysIem also helps Ihe proíessionalism oí Ihe call Iaker by providing Ihe exacI wording IhaI should be used Io ask a guesIion oí Ihe cusIomer. Call Iakers are given Iwo weeks oí Iraining beíore Ihey are able Io Iake calls írom a cusIomer. This Iraining used Io involve edu- caIion abouI Ihe appliances and how Io pro- íessionally inIeracI wiIh a cusIomer. Now, Iraining íocuses mainly on proíessional inIer- acIion, and Ihe sysIem acIs as "on Ihe job" Iraining abouI Ihe appliances. ArchiIecIure Issue The archiIecIure issue addresses Ihe design principles, meIhodologies, and Iechnologies íor diííerenI Iypes oí mixed-iniIiaIive roles and behaviors. HorviIz (1ººº) described 12 criIical íacIors íor mixed-iniIiaIive sysIems. These íac- Iors include considering uncerIainIy abouI a user´s goals, employing dialogue Io resolve key uncerIainIies, employing socially appropriaIe behaviors íor agenI-user inIeracIions, main- Iaining working memory oí recenI inIerac- Iions, and conIinuing Io learn by observing. These are excellenI design principles íor a mixed-iniIiaIive sysIem. However, even wiIh greaI archiIecIure principles iI is sIill diííiculI Io wriIe a good mixed-iniIiaIive agenI. Eor exam- ple, many people do noI like Ihe MicrosoíI Oííice assisIanI, someIimes called Clippy. Clippy did have many good properIies, buI iI was removed írom MicrosoíI Oííice because Ihe value oí iIs help was noI greaIer Ihan Ihe cosI oí Ihe incorrecI inIerrupIions. Many users con- sidered a small number oí incorrecI inIerrup- Iions or even a single incorrecI inIerrupIion Io be Ioo much. II was very diííiculI íor Clippy Io give correcI and Iimely advice aI all Iimes. However, Ihere is anoIher íorm oí mixed-ini- IiaIive inIeracIion Iaking place in MicrosoíI Oííice. The spelling and grammar checkers acI as a simple mixed-iniIiaIive sysIem. MicrosoíI Word puIs a wavy line under words IhaI may noI be correcIly spelled or may have inappro- priaIe grammar. The goal oí Ihis is Io help Ihe user wriIe wiIh proper spelling and grammar. The user should usually wanI Io have proper spelling and grammar, so Ihere is liIIle uncer- IainIy abouI Ihe user´s goals. There is a dialogue available ií Ihe user clicks on Ihe underlined word. The dialogue is a pop up IhaI includes a guess aI Ihe correcI word, Ihe abiliIy Io "ignore" Ihe suggesIion, which removes Ihe line, Ihe abiliIy Io "add" Ihe word Io a person- al dicIionary, and a descripIion oí Ihe gram- maIical rule IhaI caused Ihis Io be underlined ií iI was a grammar issue. Underlining a word is socially appropriaIe inIeracIion since iI does noI inIerrupI a user. Memory oí Ihe usage oí Ihe "ignore" and "add" opIions can allow Ihe grammar and spelling agenIs Io conIinue Io learn Ihe users´ preíerences. The STC archiIecIure puIs exIra emphasis on Ihe mainIainabiliIy oí Ihe sysIem. Every Iime Ihe case base was used Io answer a call, a descripIion oí Ihe call was wriIIen Io a reporI- ing daIabase. The descripIion included sIarI Iime, end Iime, caller Ielephone number, call- Iaker ID, Iype oí appliance, a shorI IexI descrip- Iion oí Ihe issue, all guesIions asked by Ihe sys- Iem, all answers given Io Ihese guesIions, Ihe íinal diagnosis suggesIed, and wheIher Ihis answer was accepIed by Ihe caller. Each week a case auIhor would analyze Ihis iníormaIion in Ihe reporIing daIabase íor every call IhaI was Iaken íor Ihe week. The call Iakers would help wiIh Ihis analysis and were a valuable parI oí Ihe developmenI Ieam. FersonalizaIion Issue The personalizaIion issue addresses Ihe adapIa- Iion oí Ihe agenI´s behavior Io Ihe preíerences oí iIs user. AgenIs can be more useíul ií Ihey are personalized íor Ihe user. All users are noI Ihe same, so iI can be diííiculI íor one agenI Io make all users happy all oí Ihe Iime. One com- mon way Io personalize Ihe acIions oí an agenI íor a speciíic user is Io sIore iníormaIion abouI Ihe user and use Ihis iníormaIion Io deIermine Ihe Iype and Iiming íor help Ihe agenI pro- vides. One oIher way íor Ihe agenI and user Io work IogeIher more eííecIively is Io have com- mon shared experiences. This means Ihe agenI needs Io remember pasI experiences wiIh Ihe user and use Ihese Io beneíiI Ihe user. One sim- ple example oí Ihis is when Ihe Iool MaIlab (or oIher soíIware developmenI environmenIs) Artìcles º8 AI MAGAZINE sLarL end user i agenL k lìçure S. User-Açent lnteractìon. remembers pasI user commands and allows Ihe user Io guickly selecI and reissue Ihese com- mands. This has Ihe eííecI oí allowing Ihe user Io guickly say, "Do iI again." The pasI experi- ences can be boIh shorI-Ierm and long-Ierm experiences. There can be problems wiIh personalizaIion. An agenI should be careíul abouI making assumpIions abouI a user and assuming Ihese preíerences do noI change. Freíerences can change over Iime or be misundersIood by Ihe agenI, so Ihe agenI should allow users Io ediI Iheir personal preíerences. Eor example, Iwo years ago one oí Ihe auIhors purchased Iwo DVDs on poIIy Iraining íor his daughIer írom amazon.com. Amazon was sIill annoyingly rec- ommending poIIy-Iraining books and videos Iwo years laIer. EorIunaIely, Amazon now does allow users Io improve Iheir recommendaIions by ediIing Ihe lisI oí iIems Ihey have purchased or previously raIed. AnoIher example is írom Ihe TiVo digiIal video recorder, which suggesIs and Iakes iniIiaIive Io record shows iI Ihinks iIs user would enjoy. Many people have com- menIed IhaI írom recording shows like Wìll & Grace and Queer Lye jor the Straìçht Guy Iheir TiVo now Ihinks Ihey are gay (Zaslow 2002). One person reporIedly Iried Io Ieach his TiVo IhaI he was noI gay by recording many Ihings a sIraighI man would waIch like World War II movies and documenIaries. Now Ihe TiVo Ihinks he is a Nazi. STC has some limiIed personalizaIion capa- biliIy. There is no personalizaIion íor Ihe call Iaker and some personalizaIion íor Ihe cus- Iomer. The ideal personalizaIion scenario is when cusIomers iniIialize Iheir personal iníor- maIion by regisIering Ihe producIs Ihey pur- chase. AlIernaIively, call Iakers will updaIe Ihe iníormaIion íor Ihe cusIomer when Ihe cus- Iomer calls ASI. IníormaIion abouI cusIomers and Ihe appliances Ihey own is sIored in a daIa- base. This iníormaIion is used by STC when iI suggesIs guesIions and soluIions. In addiIion, pasI experiences oí Ielephone conversaIions (guesIions asked and answers given) are also sIored. This is beneíicial ií, íor example, a cus- Iomer is íorwarded írom a íirsI-level call Iaker Io a supervisor. In IhaI case, all answered gues- Iions are visible Io Ihe supervisor. Having Ihis personalized iníormaIion readily available reduces cusIomers´ írusIraIion IhaI comes írom repeaIing Ihe same iníormaIion. The same principle allows personalized iníormaIion Io be shared beIween diííerenI call Iakers. Ií a cus- Iomer calls mulIiple Iimes, Ihen previous ses- sions can be reviewed Io geI Ihe background on Ihe currenI call. All guesIions and answers are sIored íor 30 days in one daIabase (shorI-Ierm memory) and a brieí summary oí Ihe call is sIored permanenIly in anoIher daIabase (long- Ierm memory). ResulIs STC has been in consIanI use since Ihe end oí 1ººº. More Ihan 300 call Iakers aI mulIiple locaIions in Ihe UniIed SIaIes use Ihe sysIem. The percenIage oí calls IhaI are correcIly answered over Ihe Ielephone has increased each year. Eigure ô shows Ihe success raIe íor each year. EreguenI cusIomer surveys show cusIomer saIisíacIion is higher wiIh cases IhaI are answered using STC Ihan wiIh calls IhaI are noI. This sysIem has also been a íinancial success íor GE. The iniIial developmenI in 1ººº cosI $1 million íor Ihe soíIware Iool, Iníerence Corpo- raIion proíessional services, Iwo person years oí eííorI by GE personnel, and Ihe cosI oí hard- Artìcles SUMMER 2007 ºº 1ººº 0 5 10 15 20 25 2000 2001 2002 ¥ear P e r c e n L a g e
H e l p e d 2003 2004 2005 lìçure ó. STC Success Rate. 0 0.ó 0.4 0.2 0.8 1 1.2 1.4 1.ó v o l u m e
( M i l l i o n s ) 1ººº 2000 2001 2002 ¥ear 2003 2004 2005 lìçure 7. Call Volume. WIllIam Cheetham is a senior researcher in Ihe arIiíicial inIelligence lab- oraIory oí Ihe General ElecIric Global Research CenIer in upsIaIe New York, where he has worked since 1º8S. His job is Io invenI or improve knowledge-based producIs and services. This oíIen involves Ihe applicaIion oí arIiíicial inIelligence and relaIed Iech- nigues. He has led Ihe developmenI oí more Ihan a dozen inIelligenI sysIems IhaI are in use IhroughouI Ihe General ElecIric Company. Since 1ºº8, he has been an adjuncI proíessor aI Rensselaer FolyIechnic InsIiIuIe, where he now Ieaches Ihe class Applied InIelligenI Reasoning SysIems. KaI Goebel is a senior scienIisI aI RIACS, where he is coordinaIor oí Ihe FrognosIics CenIer oí Excellence íor NASA Ames Research CenIer. Frior Io IhaI, he worked aI General ElecIric´s Global Research CenIer in Niskayuna, New York, írom 1ºº7 Io 200ô as a senior research scienIisI. He has carried ouI applied research in Ihe areas oí arIiíicial inIelligence, soíI compuIing, and iníorma- Iion íusion. His research inIeresI lies in advancing Ihese Iechnigues íor real-Iime moniIoring, diagnosIics, and prognosIics. He has íielded numerous applicaIions íor aircraíI engines, IransporIaIion sysIems, medical sysIems, and manuíacIuring sys- Iems. He holds halí a dozen paIenIs and has published more Ihan 7S papers. Since 1ºº8, Goebel has been an adjuncI proíessor oí Ihe CompuIer Science DeparImenI aI Rensselaer FolyIechnic InsIiIuIe (RFI), Troy, New York, where he has been coIeaching classes in soíI compuIing and applied inIel- ligenI reasoning sysIems wiIh William CheeIham. ware Io deploy Ihe sysIem. The main- Ienance in each oí Ihe nexI six years has averaged $0.S million íor ongoing mainIenance oí Ihe cases and a major upgrade Io a new version oí Ihe eGain soíIware in 2004. The beneíiI oí noI sending a íield service Iechnician Io a cusIomer´s home when Ihe producI is in warranIy is $S0 íor GE. The dollar savings íor GE can be calculaIed by Ihe íollowing íormula Savings = increase in success raIe * call volume * $S0 Eigure 7 shows Ihe call volume per year. The sum oí Ihe savings íor each year írom 2000 Io 200S is $44.S mil- lion. This is a projecI IhaI can boIh provide beIIer service íor cusIomers and reduce Ihe cosI oí Ihis service. OIher beneíiIs include Ihe íollow- ing: (1) Higher íirsI-call success raIe. This increases cusIomer saIisíacIion and decreases Ihe number oí calls GE needs Io handle. (2) Early idenIiíica- Iion oí new Iypes oí cusIomer prob- lems. This íeedback can be senI Io design Ieams, who can íix Ihe problem in íuIure releases and reduce íuIure cusIomer problems. (3) The abiliIy Io eníorce policies such as "Ií Ihe model is sold by a parIicular reIailer S, Ihen ask ií iI was purchased aI S." S pays íor Ihis service, buI call Iakers oíIen íorgoI Io ask, and GE ended up shouldering Ihe cosI. (4) The increased abiliIy Io mail parIs ouI wiIh insIrucIions and avoid a service call íor iIems IhaI can- noI be íixed over Ihe Ielephone buI are an easy íix and Ihe cusIomers wanI Io do iI Ihemselves. (S) STC idenIiíies parIs needed when a íield service Iech- nician is senI Io Ihe home so more íix- es can be made on Ihe íirsI Irip, saving Iime íor Ihe cusIomer and íield Iech- nician (Ihe parI needed is reíerred Io as "inIercepI" in íigure 3). (ô) The increased consisIency oí call Iakers has reduced repeaI calls írom cusIomers "íishing" (a pracIice in which cus- Iomers call repeaIedly in hopes Io geI diííerenI answers). Our implemenIaIion oí Cisco´s e- mail Iool wenI inIo use in June oí 2000. The Iool reduced Ihe average Iime a person needs Io answer an e- mail message by 40 percenI. Since our yearly projecIed cosI Io answer e-mails was $1.S million, Ihe savings oí Ihis Iool is $ô00,000 per year. Conclusion GE Consumer FroducIs has been suc- cessíul using Ihe STC mixed-iniIiaIive sysIem Io provide cusIomer supporI. The sysIem conIains exhausIive knowledge abouI appliance problems and Ihe guesIions needed Io diagnose Ihem. The human call Iaker provides some knowledge abouI appliances and a proíessional inIeríace beIween our cusIomers and Ihe diagnosIic knowl- edge sIored in Ihe Iool. The Ieam oí compuIer sysIem and human call Iak- er períorms beIIer Ihan Ihe call Iakers had beíore Ihe sysIem was creaIed. Reíerences AamodI, A., and Flaza, E. 1ºº4. Case-Based Reasoning: EoundaIional Issues, MeIhod- ological VariaIions, and SysIem Approach- es, AlCOM 7(1). Acorn, T., Walden, S. 1ºº2. SMART: SupporI ManagemenI CulIivaIed Reasoning Tech- nology íor Compag CusIomer Service. In lnnovatìve Applìcatìons oj Artìjìcìal lntellì- çence 4, ed. C. ScoII and F. Klahr. Cam- bridge, MA: AAAI Fress}The MIT Fress. Allen, J. 1ººº. Mixed-IniIiaIive InIeracIion. lLLL lntellìçent Systems 14(ô): 14÷1ô. CheeIham, W., 2003. Lessons Learned Using CBR íor CusIomer SupporI. In lro- ceeJìnçs oj the Sìxteenth lnternatìonal llorìJa Artìjìcìal lntellìçence Research Socìety Conjer- ence. Menlo Fark, CA: AAAI Fress. CheeIham, W., Frice, J. 2004. Measures oí SoluIion Accuracy in Case-Based Reasoning SysIems. Faper presenIed aI Ihe SevenIh European Coníerence on Case-Based Rea- soning, Madrid, AugusI 30÷SepIember 2. HorviIz, E. 1ººº. Frinciples oí Mixed-IniIia- Iive User InIeríaces. In lroceeJìnçs oj the Conjerence on Human lactors ìn Computìnç Systems, 1Sº÷1ôô. New York: AssociaIion íor CompuIing Machinery. Simoudis, E. 1ºº2. Using Case-Based Rea- soning íor CusIomer Technical SupporI. lLLL Lxpert 7(S): 7÷13. Tecuci, G., Boicu M., and Cox, M. 2007. Seven AspecIs oí Mixed IniIiaIive Reason- ing: An InIroducIion Io Ihis Special Issue on Mixed IniIiaIive AssisIanIs. Al Maçazìne 28(2). WaIson, I. 1ºº7. Applyìnç Case-BaseJ Rea- sonìnç. Technìques jor Lnterprìse Systems. San Erancisco: Morgan Kauímann Fublishers. Zaslow, J. 2002. WhaI Io Do When Your TiVo Thinks You´re Gay. The Wall Street jour- nal, November 2ô. Artìcles 100 AI MAGAZINE Stay tnformed with AAAt Member Announcements lf you are a member of AAAl and would like Lo receive periodic announcemenLs from AAAl, please go Lo aaai.org}cgi- dada}mail.cgi, choose AAAl- Members, enLer your email address, and selecL "subscribe.¨ ¥ou can also regisLer for Lhe semimonLhly Al AlerL.