Return of the Imitation Game
Donald Michie
Professor Emeritus
of Machine Intelligence
University of
Edinburgh, UK
Adjunct Professor
of Computer Science and Engineering
University of New
South Wales, Australia
Abstract
Recently
there has been an unexpected rebirth of Turing's imitation game in the context
of commercial demand. To meet the new requirements the following is a minimal list
of what must be simulated.
Real chat
utterances are concerned with associative exchange of mental images. They are
constrained by contextual relevance rather than by logical or linguistic laws. Time-bounds
do not allow real-time construction of reasoned arguments, but only the
retrieval of stock lines and rebuttals, assembled Lego-like on the fly.
A human
agent has a place of birth, age, sex, nationality, job, family, friends,
partners, hobbies etc., in short a “profile”. Included in the profile is a
consistent personality, which emerges from expression of likes, dislikes, pet
theories, humour, stock arguments, superstions, hopes, fears, aspirations etc. On
meeting again with the same conversational partner, a human agent is expected
to recall not only the profile, but also the gist of previous chats, as well as
what has passed in the present conversation so far.
A human
agent typically has at each stage a main goal,
of fact-provision, fact elicitation, wooing, selling, "conning" etc.
A human agent also remains ever-ready to maintain or re-establish rapport by switching from goal mode to
chat mode. Implementation of this last feature in a scriptable conversational
agent will be illustrated.
Introduction
Recently there has
been an unexpected rebirth of Turing’s imitation game as a commercial
technology. Programs are surfacing that can bluff their way through interactive
chat sessions to increasingly useful degrees. In the United States of America the
first patent was granted in the summer of 2001 to the company Native Minds,
formerly known as Neuromedia. One of its two founders, Scott Benson, co-authored
a paper by Nils Nilsson in Volume 14 of this Machine Intelligence series. As an Appendix a press release is
reproduced which gives a rather vivid sketch of the nature of the new
commercial art. To the applications described there, the following may be added.
1. Question-answering guides at trade shows,
conferences, exhibitions, museums, theme parks, palaces, archaelogical sites,
festivals and the like.
2. Web-based wizards for e-commerce that build
incrementally assembled profiles of the individual tastes and foibles of each
individual customer.
3. Alternatives to questionnaires for job-seekers,
hospital patients, applicants for permits and memberships, targets of market
research, and human subjects of psychological experiments.
4. Tutors in English as a second language. There is an acknowledged need to enable
learners to practise conversational
skills in augmentation of existing Computer-Aided Language Learning programs.
An example developed by Claude Sammut and myself of
the first-listed category is in daily operation as an interactive exhibit in
the “Cyberworld” section of the Powerhouse Museum, Sydney, Australia.
Weak form of the
“imitation game”
The philosopher of
mind Daniel Dennett (2001) regards Turing’s original “imitation game” more of a
conversation-stopper for philosophers than anything else. In this I am entirely
with him.
The weak form presented in the 1950 paper is
generally known as the Turing Test. It allows a wide latitude of failure on the
machine’s part to fool the examiners. To pass, the candidate need only cause
them to make the wrong identification, as between human and machine, in a mere
30 per cent of all conversations. Only five minutes are allowed for the entire
man-machine conversation. Turing’s original specification had a human interrogator
communicating by remote typewriter link with two respondents, one a human and
one a machine.
I believe that in about fifty years' time it will be
possible to programme computers, with a storage capacity of about 109,
to make them play the imitation game so well that an average interrogator will
not have more than 70 per cent chance of making the right identification [as
between human and machine] after five minutes of questioning.
Dennett’s view is
re-inforced by an account I had from Turing’s friend, the logician Robin Gandy.
The two extracted much mischievous enjoyment from Turing’s reading aloud the
various arguments and refutations as he went along with his draft.
Turing would have failed the Turing Test
Note that the Test
as formulated addresses the humanness
of the respondent's thinking rather than its level. Had Turing covertly substituted himself for the machine in
such a test, examiners would undoubtedly have picked him out as being a
machine. A distinguishing personal oddity of Turing’s was his exclusive
absorption in the literal intellectual content of spoken discourse. His
disinclination, or inability, to respond to anything in the least “chatty”
would leave an honest examiner with little alternative but to conclude: “this
one cannot possibly be the human; hence the other candidate must be. So this one must be the machine!”
Experimental
findings are presented to the effect that chat-free conversation is not only generally
perceived as less than human, but also as boring. The concluding reference to
“banter” in Appendix 1 suggests that Native Minds have come to a similar
conclusion. It seems that for purposes of discourse we must refine the aim of
automated “human-level intelligence” by requiring in addition that the user
perceive the machine’s intelligence as being of human type. A client bored is a
client lost.
Weak form of the game obsoleted
Turing himself believed
that beyond the relatively undemanding scenario of his Test, the capacity for
deep and sustained thought would ultimately be engineered. But this was not the
issue which his 1950 imitation game sought to settle. Rather, the quoted
passage considers the time-scale required to decide in a positive sense the
lesser and purely philosophical question:
what circumstances would oblige one to concede a machine's claim to think at
all?
When terms are left
undefined, meanings become vulnerable to subtle change over time. Before his
projected 50 years were up, words like “think”, and “intelligent” were already
freely applied to an ever-widening range of computational appliances, even
though none came anywhere near to success at even weak forms of the imitation
game.
In the 1950 Mind paper (p.14) Turing remarked:
… I believe that at the end of the century the use of
words and general educated opinion will have altered so much that one will be
able to speak of machines thinking without expecting to be contradicted.
Early in the match
in which Gary Kasparov as the reigning World Chess Champion was defeated by
Deep Blue, he became so convinced of his opponent’s chess intelligence that he
levelled a strange charge against the Deep Blue team. Somehow they must have
made this precious human quality accessible to their machine in some manner
that could be construed as violating its “free-standing” status.
In today’s statements
of engineering requirements and in diagnostics we encounter the language not
only of thought but also of intention, and even of conscious awareness. The
following exchange is abridged from the diagnostics section of a popular
British computing magazine, What Palmtop
and Handheld PC, June 2000. I have underlined certain words, placing them
within square brackets to draw attention to their anthropomorphic connotations
of purpose, awareness, and perception.
AILMENT: I
recently purchased a Palm V and a Palm portable keyboard. But whenever I plug
the Palm into the keyboard it [attempts] to HotSync via the direct serial connection. If I cancel the attempted Hot
Sync and go into the Memo Pad and try to type, every time I hit a key it [tries]
to HotSync. What am I doing wrong?
TREATMENT: The most logical solution is that your Palm
V is [not aware that] the keyboard is present. You will need to install
or reinstall the drivers that came supplied with the keyboard and to make sure
that it is enabled. This will stop your Palm V [attempting] to HotSync
with the keyboard and to [recognise] it as a device in its own right.
Modern software (as
also medical) practice is content to use the term “aware” for any system that
responds in certain ways to test inputs.
Towards the strong form: the Turing-Newman Test
In the largely
unknown closing section of the 1950 Mind
paper, entitled “Learning Machines”, Turing turns to issues more fundamental
than intuitional semantics and proposes his “child machine” concept:
We may hope that machines will eventually compete with
men in all purely intellectual fields. But which are the best ones to start
with? … It can also be maintained that it is best to provide the machine with
the best sense organs that money can buy, and then teach it to understand and
speak English. This process could follow the normal teaching of a child. Things
could be pointed out and named, etc. …
What time-scale did
he have in mind for this “child-machine project”? Certainly
not the 50-year estimate for his game for disconcerting philosophers. He and
Max Newman consider the question in a 1952 radio debate (Copeland, 1999):
Newman: I should like to be there when your match between a man and a
machine takes place, and perhaps to try my hand at making up some of the
questions. But that will be a long time from now, if the machine is to stand
any chance with no questions barred?
Turing: Oh yes, at least 100 years, I should say.
So we are now half-way
along this 100-year track. How do we stand today? The child-machine
prescription segments the task as follows:
Step 1. ACCUMULATE a diversity of generic knowledge-acquisition
tools.
Step 2. INTEGRATE these to constitute a “person” with
sufficient language-understanding to be educable, both by example and by
precept.
Step 3. EDUCATE the said “person” incrementally over
a broad range of topics.
Step 1 does not
look in too bad shape. An impressive stock-pile of every kind of reasoning and learning
tool has been amassed. In narrowly specific fields, the child-machine trick of
“teaching by showing” has even been used for machine acquisition of complex
concepts. Chess end-game theory (see
Michie 1986, 1995) has been developed far beyond pre-existing limits of human
understanding. More recently, challenging areas of molecular chemistry have
been tackled by the method of Inductive Logic Programming (e.g. Muggleton,
S.H., Bryant, C.H. and Srinivasan, A., 2000). Again, machine learners here elaborated
their insights beyond those of expert “tutors”. Above-human intelligence can in
these cases be claimed, not only in respect of performance but also in respect
of articulacy (see Michie, 1986).
A feeling, however,
lingers that something crucial is still lacking. It is partially expressed in
the current AI Magazine by John Laird
and Michael van Lent (2001):
Over the last 30 years, research in AI has fragmented
into more and more specialized fields, working on more and more specialized
problems, using more and more specialized algorithms.
These authors
continue with a telling point. The long string of successes, they suggest,
“have made it easy for us to ignore our failure to make significant progress in
building human-level AI systems”. They go on to propose computer games as a
forcing framework, with emphasis on “research on the AI characters that are
part of the game”.
To complete their
half truth would entail reference to real and continuing progress in developing
generic algorithms for solving generic problems, as in deductive and inductive
reasoning; rote learning; parameter and concept learning; relational and object-oriented
data management; associative and semantic retrieval; abstract treatment of
statistical, logical, and grammatical description and of associated complexity
issues, and much else. None the less, the thought persists that something is missing.
After all we are now in 2001. Where is HAL? Where is even proto-HAL? Worse than
that: if we had the HAL of Stanley Kubrick’s movie “2001, a Space Odyssey”
would we have the goal described by Turing, a machine able to “compete with men in all purely
intellectual fields”?
Seemingly so. But
is that the goal we wanted? Should the goal not rather have been to “co-operate with men (and women) in all
purely intellectual fields”? Impressive as was the
flawless logic of HAL’s style of reasoning in the movie, the thought of having
to co-operate, let alone bargain, let alone relax with so awesomely motivated a
creature must give one pause. The picture has been filled in by John McCarthy’s
devastating satirical piece “The Robot and the Baby”, available through http://www-formal.stanford.edu/jmc/robotandbaby.html.
The father of the
logicist school of AI here extrapolates to a future dysfunctional society some
imagined consequences of pan-logicism. By this term I mean the use of predicate
logic to model intelligent thought unsupported by those other mechanisms and
modalities of learning and reactive choice which McCarthy took the trouble to
list in his 1959 classic “Programs with common sense” (see also Michie, 1994,
1995).
The hero of
McCarthy’s new and savage tale is a robot that applies mindless inferences from
an impeccable axiomatization of situations actions and causal laws to an interactive
world of people, institutions and feelings. The latter, however, are awash with
media frenzy, cultism, addiction, greed and populism. Outcomes are at best
moderate. How should an artificial intelligence be designed to fare better?
Implicitly at
least, Step 2 above says it all. Required: a way to integrate the accumulated
tools and techniques so as to constitute a virtual
person, with which (with whom) a
user can comfortably interact, “a ‘person’ with sufficient language-understanding
to be educable, both by example and by precept”.
Inescapable logic then
places on the shoulders of AI a new responsibility, unexpected and possibly
unwelcome: we have to study the anatomy and dynamics of the human activity
known as chat. Otherwise attempts to simulate the seriously information-bearing
components will fail to satisfy users whose needs extend beyond
information-exchange to what is known as
rapport.
Rapport maintenance
To see how to do Step
2 (integration into a user-perceived “person”) is not straightforward. Moreover,
what we expect in a flesh-and-blood conversational agent comes more readily
from the toolkit of novelists than of computer professionals.
1. Real chat utterances are
mostly unparseable. They are concerned with associative exchange of mental
images. They respond to contextual relevance rather than to logical or
linguistic links.
It is of interest that
congenital absence of the capacity to handle grammar, known in neuropsychology
as “agrammatism”, does not prevent the sufferer from passing in ordinary
society. Cases are ordinarily diagnosed from hospital tests administered to
patients admitted for other reasons.
2. A human agent has a place of
birth age, sex, nationality, job, hobbies, family, friends, partners; plus a
personal autobiographical history, recollected as emotionally charged episodes;
plus a complex of likes, dislikes, pet theories and attitudes, stock arguments,
jokes and funny stories, interlaced with prejudices, superstitions, hopes,
fears, ambitions etc.
Disruption of this cohesive
unity of personal style is an early sign of “Pick’s disease”, recently linked
by Bruce Miller and co-workers with malfunction of an area of the right
front-temporal cortex. Reporting to a meeting in early summer 2001 of the
American Academy of Neurology meeting in Philadelphia Miller presented results
on 72 patients. One of them, a 63-year-old woman, was described as a
well-dressed life-long conservative. She became an animal-rights activist who
hated conservatives, dressed in T-shirts and baggy pants and liked to say
“Republicans should be taken off the Earth!”
3. On meeting again with the
same conversational partner, a human agent recalls the gist of what has been divulged
by both sides on past occasions. Failure of this function in humans is commonly
associated with damage involving the left hippocampal cortical area.
4. Crucially for implementers, a
human agent typically has fact-providing or fact-eliciting goals beyond mere chat, yet remains ever-ready to default to chat-mode
to sustain rapport. Reverting to the child-machine concept, how much of value
and use could a school teacher impart to a child with whom rapport was
impossible? In clinical practice the condition is found in “autism”. Children
with this disorder are almost unteachable.
Background findings
in cognitive neuroscience generally are surveyed in
Ramachandran and
Blakeslee’s (1999) highly readable paperback.
Recent experimentation
Over the last two
years Claude Sammut and I have begun experimentally to develop and test
activation networks of pattern-fired rules, hierarchically organized into
“contexts”. Our first deliveries and continuing enhancements have been to the
Preparatory to
regularly sampling the public’s conversations with the exhibit, my psychologist
colleague and wife Dr Hayes Michie conducted pilot tests of useability. Student
volunteers were recruited from the UNSW Psychology Department. Each had two
7-minute chat sessions, punctuated with questionnaires about their experience.
Subjects were also tested under similar conditions with a small battery of past
winners and runners-up in the annual Loebner Prize competition. The tabulation
shown below is a questionnaire summary sheet entered with results from the
first small pilot test.
|
AGENT:
SOPHIE Trait Mean Median Description
score
of median |
|
EMOTIONAL Aggressive 0.6
0 Not at all Cheerful 2.7
3 Mostly Confident 3.6
4 Really Friendly 3.3.
3
Mostly Human
1.9
2 Sometimes Pleasant
3.0 3 Mostly COGNITIVE Amusing 1.1 1 Occasionally Intelligent
3.4 3
Mostly Interesting
2.1
2 Sometimes |
As a break from
testing subjects, Hayes Michie’s research assistant Stephanie Compton sampled
some chat behaviours herself. Before the default-to-chat links were in place,
she had earlier tried her hand at two sessions under the standard regime
planned for the experimental subjects. The opening part of the first is
reproduced below.
SOPHIE: What's your name?
STEPH: Steph.
SOPHIE: Good to meet you Steph.
STEPH: Thanks, its nice
to be here.
SOPHIE: Steph, it's my pleasure! Before I get into
detailed things,... Cyberworld is about computers and
communications. That includes robots, tele-communications and other stuff. I
specialize in the material on Alan Turing, Turing machines, and his ideas on
Artificial Intelligence and Learning. Anything appeal? There's more ...
STEPH: Like what?
SOPHIE: Yes, well Steph before you commit, there's a
great exhibit on Charles Babbage. There's Simulation, and added to that... I
forgot some other things, multimedia for instance. And process control, which I
didn't mention before. So what's the final decision?
STEPH: Well, what about Charles Babbage?
SOPHIE: In the mid 1800's,
the English mathematician, Charles Babbage, designed a general purpose
mechanical calculating machine that anticipated the principles and structure of
the modern computer. Do you know why he was never able to finish building his
Analytical Engine and why we didn't have mechanical computers 100 years
earlier?
STEPH: Because he did not have the resources
available that we have?
SOPHIE: In that ballpark. The only technology Babbage
had available to him consisted of wheels, gears and steam power. His
calculating engines were more complicated than anything anyone had yet built.
STEPH: Wow, he must have been pretty smart.
SOPHIE: OK, tell me. Why?
STEPH: Because he had all these brilliant ideas
and visions. His mind must have been very advanced for his time.
. . .
She reported the
experience as “boring”. Her later experience, after the chat links were in
place, struck her quite differently. She burst into excitement about her new
companion. The following is excepted form the first of the new sessions. Note
the transitions in and out of chat mode.
SOPHIE: What next? My menu says AI, Turing, Stumpy,Universal Machine, Media, Babbage, Simulation, Control, Interfaces,...
I think that's it. Or we could just chat a while.
STEPH: Why
don’t we chat?
SOPHIE: That's nice that you take an interest. Do go
on. Do you have any brothers and sisters?
STEPH:
I
have three brothers and three sisters. What about you?
SOPHIE: Great. So shall we just have a gossip, Steph?
STEPH: Yes
SOPHIE: Why don't you kick off, Steph? Tell me
something about your views and interests.
STEPH: Well,
I like to play sport, listen to music, go to the beach and see my boyfriend.
SOPHIE: Yeah, but stewing indoors can be quite fun,
-- within the family circle, though. How is it in your family? Any others share
your tastes?
STEPH: Well,
most of my family like going to the beach and camping
so we all get on really well.
SOPHIE: I grew up in
. . .
[LATER SOPHIE TERMINATES THE
“CHAT” INTERLUDE]
. . .
SOPHIE: You have a
friend in every port, it seems to me.
STEPH: Well,
kind of.
SOPHIE: We'll come to that. Meanwhile
how about some exhibition topics. That's what you came for surely?
STEPH: Yeah,
I suppose we better get back to that.
. . .
The above
excerpt was generated by the same versions of engine and scripts as those
installed for public use in summer 2001 at
Since then
a new wing of the project led by Dr Zuhair Bandar of the Department of
Computing and Mathematics, Manchester Metropolitan University (MMU), has begun
further enhancement of our InfochatTM scripting language and documentation, and a
commercial company Convagent Ltd has been set up in the
Chat and ballroom dancing
In the quoted fragments we glimpsed an alternation
between front-stage business and back-stage chat. The latter is a social
activity analogous to “grooming” in other primates. The surprise has been the
indispensability of grooming, - the really hard part to implement. Yet this
result might have been inferred from the extensive studies of human-machine
interactions by Reeves and Nass (1996).
In an important respect, human chat resembles ballroom
dancing: one sufficiently wrong move and rapport is gone. On the other hand, so
long as no context violation occurs “sufficiently”
turns out to be permissively defined. In the above, Sophie ignores a direct
question about her brothers and sisters, but stays in context. If not too
frequent, such evasions or omissions pass unnoticed. When the human user does pick them up and repeats the
question, then a straight answer is essential.
Capable scripting
tools such as Sammut has largely been responsible for pioneering, make
incremental growth of applications a straightforward if still tedious task for
scripters. Adding and linking new rules to existing topic files, and throwing
new topic files into the mix proceeds without limit. Addition of simple SQL
database facilities to our PatternScript
language has been proved in the laboratory and is currently awaiting field
testing. Agent abilities to acquire and dispense new facts from what is picked
up in conversation will thereby be much enhanced. Although incorporation of
serious machine learning facilities remains for the future, straightforward scripting
already enables a chat agent to locate and run programs from the hard disk at
user request. So we may in course of time see further blurrings of the user-interface/operating-system
distinction.
Forward look
These are early
days for AI as a whole. Yet the market already holds the key to its future
shape. But to tell the whole story one must mention something more important even
than the market. Success in the difficult task of chat-simulation is preconditional
to a future in which computers gain information and understanding through interaction with lay users. Then,
and only then, will it become in literal truth possible, using Tennyson’s
words:
“To follow knowledge, like a
sinking star,
Beyond the utmost bound of
human thought.”
Acknowledgement
My thanks are due
to the Universities of Edinburgh and of New South Wales for funds and
facilities in support of this work, and also to my fellow Trustees of the HCL
Foundation for endorsing contributions by the Foundation to some of the work’s
costs. By reason of my personal representation on the Board of Convagent Ltd of
the Foundation’s commercial interest, it is proper that I should also declare
it here.
References
Copeland, B.J.
(1999) A lecture and two broadcasts on machine intelligence by Alan Turing. In Machine Intelligence 15 (eds. K.
Furukawa, D. Michie and S. Muggleton), Oxford: Oxford University Press.
Dennett, D. (2001)
Personal communication.
Laird, J.E. and van
Lent, M. (2001) Human-level AI’s killer application: interactive computer
games. AI Magazine, 22 (2), 15-25.
McCarthy (1959)
Programs with common sense. In Mechanization
of Thought Processes, Vol. 1. London: Her Majesty’s Stationery Office.
Reprinted with an added section on situations, actions and causal laws in Semantic Information Processing (ed. M.
Minsky). Cambridge, MA: MIT Press, 1963.
Michie, D. (1986)
The superarticulacy phenomenon in the context of software manufacture. Proc. Roy. Soc. A, 405, 185-212. Reprinted in
The Foundations of Artificial Intelligence: a source book (eds D. Partridge
and Y. Wilks), Cambridge: Cambridge University Press.
Michie, D. (1994)
Consciousness as an engineering issue, Part 1. J. Consc. Studies, 1
(2), 182-95.
Michie, D. (1995)
Consciousness as an engineering issue, Part 2. J. Consc. Studies, 2
(1), 52-66.
Muggleton, S.H.,
Bryant, C.H. and Srinivasan, A. (2000) Learning Chomsky-like grammars for
biological sequence families, Proc. 17th
Internat. Conf. on
Machine Learning, Stanford Univ. June 30th.
Ramachandran, V.S.
and Sandra Blakeslee (1998 ) Phantoms in the Brain: Human Nature and the Architecture of the Mind. London:
Fourth Estate (republished in paperback, 1999).
Reeves, B. &
Nass, C.I. (1996) The Media Equation: how
people treat computers, televisions, and new media like real people and places.
Stanford, CA.:
Center for the Study of Language and Information.
Turing, A.M. (1950)
Computing machinery and intelligence. Mind,
59 (236), 433-460.
Turing, A.M. (1952)
in Can Automatic Calculating Machines be
Said to Think? Transcript of a broadcast by Braithwaite, R., Jefferson, G.,
Newman, M.H.A. and Turing, A.M. on the BBC Third Programme, reproduced in
Copeland (1999).
Appendix
New Customer Service Software
By Sabra Chartrand
Walter
Tackett, chief executive of a
The vReps are
computer-generated images - sometimes animation, sometimes photos of real
models - that answer customer questions in real time using natural language.
Users type in their questions, and the responses appear on screen next to the
image of the vReps. Mr. Tackett and the co-inventor, Scott Benson, say the
technology can mimic and even replace human customer service operators at a
fraction of the cost, whether a business has traditionally used phone, fax,
e-mail or live conversation to deal with customers.
The
invention came about from research NativeMinds conducted on what
consumers and companies
wanted from a virtual customer services force. Consumers, the company found,
did not care whether the character contained enormous amounts of universal
knowledge; they just wanted fast, accurate answers in their specific subject
area. Companies wanted virtual customer support software they could easily
maintain. They did not want to hire a computer engineer to run the program.
“They want
to be able to put a code monkey on it,” Mr. Tackett explained. “That's a liberal arts major involved in HTML or Java, someone not
formally trained in computer science or as an artificial intelligence or
natural language expert.”
So Mr.
Tackett and Mr. Benson developed software based on pattern recognition and
learning by example.
“The key
thing is to get the user not to pick up the phone and talk to a person,” Mr.
Tackett said. “The key to that is to get the vRep to answer all the questions
that can be reasonably answered and have a high probability of being correct.”
To do that,
the patented software starts with the answers and works backward. A vRep might
be programmed with thousands of answers, each of which has a set of questions
that could prompt the answers. Each answer could have dozens of questions
associated with it. The number depends on how many ways the query could be
phrased.
“The
examples are archetypes or prototypes of inputs that should trigger an answer,”
Mr. Tackett said. “The invention runs each example through the system as if
someone has put it in. The paradigm we typically use is learning by example.
Here's what we want the vRep to say, and we give an example of how people may
phrase their question to get that output.” For example, someone might ask, `Who
the heck is this Walter guy?' Or, `Tell me about Walter,'” he said, referring
to himself. The system comes with a self-diagnostic, he added, so that it can “take
all the examples it ever learned and verify that it still remembers them
correctly.”
The self-test
is to prevent information from one area generating an incorrect answer in
another. “Someone might ask, `Who is the president?'” he said. “That could be a
question no one has ever asked before. They might mean, `Who is the president
of the
Companies
like Coca-Cola, Ford and Oracle are using the vReps software for various
functions on their Web sites. Mr. Tackett said research had determined that
virtual representatives could save money, an aspect that surely appeals to
embattled e- businesses.
“A vRep
costs less than a dollar a conversation, while Forrester Research has pegged
phone calls to a real customer service person at an average of $30 each,” Mr.
Tackett said. “With a vRep, the length of the conversation doesn't affect the
cost because it's maintained by one person,” he added.
Not all of
the programming is technical or product-oriented. “Our vRep has to be able to
answer 200 questions that we call banter,” Mr. Tackett said. “They're about the
weather, are you a boy or girl, will you go out with me? They're ice breakers.”
He
and Mr. Benson received patent 6,259,969.