Good to see you.
Thanks for being here.
This is a ways away
from campus for
me, so I hope you're
ready to nerd out
with me for a little
bit. I told somebody
I think this will
be the nerdiest
room in the whole
session, so welcome.
So let me tell
you a little
I lead the Human First
AI group at MIT. It's
part of MIT's initiative
on the digital
economy, which is the
biggest research center
at MIT. And my research
comes through a
behavioral science
lens. So I try to figure
out how people make
decisions and the
circumstances under
which they make
better decisions
or worse decisions.
And then in the overlap
of this and artificial
intelligence,
which is a catch-all
phrase for anything from
agentic AI to machine
learning to algorithms
in general, to gen AI,
and how that interface
of the way in which
people think and
make decisions and when
they outsource these
decisions to AI affects
things like trust.
And then that
last concentric
circle is market
strategy, which
is probably what
you care about.
So what does this
mean for trust in your
particular business
and your competitive
advantage? And so I
look at the behavior of
the humans at the
nexus of these three
concentric circles now
when salesforce asked
me to come today and
talk about what makes
people trust ai agents
i wanted to remix
the question a little
bit and say what
makes agents trustworthy
right we want to
think about what
makes them trustworthy
because that'll give us
predictive power about
whether or not they'll
trust those agents.
Now, everyone here
in this convention
center is getting a
lot of encouragement
around AI agents, and
that's because they
can improve customer
experiences. And in
fact, as we see here,
81% of professionals
trust agentic AI with
their customer data.
The thing about
it is customers
In fact, only 36%
of customers agree
with that sentiment.
And we know that
trust in AI has fallen
in the past year.
So how do we resolve
this tension?
How do we bring to
the customers the
value we want while
bearing in mind that
they may have some
reluctance around
this? Well, my research
and my team and
my lab at MIT have
led me to share with
you three ways for
us to think about.
the first way I want
to sort of motivate
with an example. So
imagine you're trying
to get here. This is
MIT. This is where I
work and you have a
goal and you need to
choose a route. You're
using a GPS app.
Now, as you can see,
this is Waze. This app
is giving you three
options. Option one,
option two, and
option three. In fact,
this is my route to
get from my home to
campus. So you can see
that Boston traffic
is is pretty brutal,
takes over an hour.
But these three
are virtually the
same, right?
They're all an hour
and 11 minutes,
give or take.
So there's no wrong
answer to this question,
and I want you to
show me by applause
which one you
would choose.
So if you would
choose route
number one,
please applaud.
If you would choose route
route three,
please applaud.
I'm hearing
crickets chirp.
If you were to look
around and see how
people voted, which
route they would
take, who would
be more persuasive
now? The people who
applauded with you
for the route that
you chose or people
who chose a different
route from you?
And our research
question then for agentic
AI is if this could
happen between humans,
could this have an
effect between humans and
their agents? That is,
could an agent activate
what we know is
similar attraction that
is you think just like
me I like you would
that lead people to
trust the agents more
so we ran an experiment
to find out if
essentially we presented
people with information
that told them that the
agent was essentially
their logic twin When
would they be more
trusting with this
agent and be more likely
to delegate tasks to
this agent for them?
And we know, for
instance, in research,
people can be
psychologically
similar or
psychologically distant.
Can the same thing
happen with AI agents?
Well, it turns
out it can.
And what we found is
that similar reasoning,
when you show the
agent's logic flow,
and it mirrors the
user's same logic
flow, people are
far more trusting.
So we see in the
orange bar, this is the
control condition. We
see in the pink bar,
this is when the user
has been shown the
logic flow of the
agent that is different
to theirs. and in
the green bar this is
when they've been shown
an agent that has a
similar logic flow
to theirs. We asked
people to do tasks to
figure out problems. We
measured their logic
flow that is the
steps they took to arrive
at their answer and
we randomly assigned
them to either see
no logic flow to see
a path that was the
same path as theirs
or to see a path that
was different. Mind
you, they all got to
the same destination,
so controlling for
accuracy, the only
thing that explains this
difference is the
presentation of a logic
flow that is either
similar or different.
We then asked people
to self-report in
order to see what mediates
this effect, because
this first chart is
simply behavioral.
It's not what people tell
And we ask them, how
reasonable, then, do
you think the agent
you used was in its
logic flow? And as you
can see, when people
see a logic flow
that is like theirs,
they think the agent
is more reasonable.
when people get a
logic flow user agent
that is like them,
they are less effortful
in their engagement
mentally. So we see,
for instance, how much
mental and perceptual
activity was required
for this task.
In the control condition,
Similarly, the mental
demand was lowest.
What this means then
is that if we can
show people the logic
flow of the agent,
Just like you saw
people who voted like
you, if we see that
the agent thinks like
us, we are more likely
to trust it. And
we can do this now
because agents aren't
just trained on
foundational data sets,
they're trained on
us. And therefore,
they can become like
our digital twins.
Now, you might
be asking, well,
what does this mean
for calibrating
trust, meaning
not over-trusting?
Certainly this could
help you not under-trust,
but would that lead
us to over-trust? And
what does this mean
for critical thinking?
If you're asking
that question, you're
asking the right
question, and a question
that we've asked as
well. So let me show you
a B2B application.
This is an experiment
I did with a B2B
company, a financial
services firm, and
they want to use agents
so that the auditors
in their firm could
be more efficient in
the auditing process.
But the concern is
that since auditors
play a really important
role in reducing
risk in the firm,
they want to make sure
that they don't
overtrust on the AI and
kind of turn off their
critical thinking.
So, their current
system has an AI
agent make suggestions
for which risk
profiles should be
part of the audit.
And then the human
agent gets to
decide based on the
agent's recommendation,
the human employee,
excuse me,
gets to decide whether
or not to adopt
what the AI agent has
suggested or remove
some or add some and
modify in some way.
In our experiments,
we ask, well, could we
do something to improve
trust calibration?
So one thing we do
is we tinker with
when the agent's
suggestion happens.
What if the agent
makes a suggestion
first, as is
the case here?
What do we see? but
what if we flipped
that and had the
human make an
opinion first and
then have the agent
come in? Would it
make a difference
in timing the
agent's involvement?
The second thing we
looked at with our
experiment is what if
you got the agent's
suggestion and then
you got a playback
for the rationale
of that agent's
suggestion. So just
a very brief two to
three sentence rationale
playback of this
is why I made the
suggestion I made.
And moreover, what if
we had an experimental
condition where
not only did people
get the rationale
playback from the AI,
but what if we
asked the human, do
you know the rationale
of the AI? and
we ask them to
play back the AI's
rationale or to
play back their own.
So we randomly assign
people to different
conditions in this
three-by-two experiment,
and here's what we found.
There's an interaction
between the timing
of when the agent
makes a suggestion
and the playback
that the human gets.
If the AI agent makes
a suggestion first,
we find that
pendulum swings where
people trust it,
perhaps overtrust even.
plays back the rationale,
when we ask the
human, what did you
just hear? What
do you think that
agent just
suggested? And what's
the rationale that
you are seeing?
pendulum back
to the center.
Now, when the human
goes first, we
find the pendulum
swings in the other
way. That is, they
have a significantly
lower reliance
on the agent.
But when they
get the AI's
playback of
its rationale,
What this does then
is it calibrates
trust. And so for you,
if your agent goes
first, you're going
to want a human
playback to calibrate
that trust. But if the
human goes first,
you're going to want
an AI playback to
calibrate that trust.
Mind you, when
we ask people in
self-report, did
this help you
calibrate your trust
better? However,
nope, had no
effect on me.
This is why I run
experiments. Because
what people say
and what people
do are not always
highly correlated.
Last but not least,
when it comes to AI,
understand that trust
in agents is highly
correlated and
exponentially so with
the value you create
for those customers.
And in my research
with thousands
of companies over
15 years at MIT,
I've come up with
a framework which
I call the EXIT
framework, E-C-S-I-T.
And that's five
forms of value that
you can deliver
to your customer.
Each one of these
forms of value is
represented in the
acronym. At MIT, we love
an acronym. It's in
our name. And so, this
was actually inspired
by the work of
Pierre Bourdieu, the
French sociologist.
So I want you to go
back to undergrad.
And Pierre
Bourdieu talked
about cultural
capital, social
capital, and
economic capital.
It turns out when
we think about the
role that agents play
in our customers'
lives, we can think
about this in terms
of capital, but in
a different way.
There are five forms
of capital or value
that you can deliver
to your customers.
The first is
economic capital.
People love
free money. They
love to save.
They love a deal.
When you provide
your customers with
deals, with
incentive through
compensation, they're
going to have more
value because you're
giving it to them.
The problem is that this
is easily replicated,
right? Because if you
give me 10% off and
somebody gives me 11
% off, then I'm gone.
So there are other
forms of capital
that increase
trust and loyalty.
The second is
cultural capital.
That's the mission
of the brand, the
values, the ritual
that I am part of. And
to the extent that
you provide customers
this, well then you
can increase the
value that they get
in the experience.
The S stands for
social capital. That is
connections between
customers, where
the connections
between the customers
increase the experience
and the value that
I get out of the
product. Anybody that's
seen dozens of Harley
-Davidson riders
go on the highway
together know that
they connect with
one another and that
improves their experience
with the brand.
The fourth kind of
value is information
capital. That's making
your customers smarter
but that's also using
their data in a way that
is transparent and
respects their privacy.
we have temporal
capital. That's saving
them time, that's
anticipating their
needs, and therefore
providing them value.
ideally five, what you
see is your customers
don't feel like you're
extracting value
from them. you're
exchanging value with them.
And it turns out that
agents are pretty
good vectors for
economic capital.
That is, they can
personalize promotions
based on the data
that they know about
customers and offer
them great deals.
They're really
good when it comes
to temporal capital.
That is, they
can save time and
be more efficient
in the experience.
and they're also
pretty good at
information capital.
That is, they can take
the data they have on
the customer and then
turn it around and
enable discovery in the
customer experience.
But you see, humans
are still needed here
because when it comes
to cultural capital,
social capital, and even
information capital,
humans are important
vectors. So what
we see here is human
and AI interaction.
We see mutually
beneficial forms
of value that increase
the probability
that the customer
feels agency
and as a result,
that they trust.
So when we look at
these forms of value,
you want to see,
Am I making my
customers feel like
you're helping them to
make in a B2B context
or save money?
That if this is
a way of life,
that they feel
connected and in the
know and being saved
effort and time.
We can see agents
work with humans
to check all
of these boxes.
So, how can we
think about customer
trust when it
comes to AI agents?
Number one,
logic similarity.
We trust agents
that think like us.
So if you are able
to show a similar
logic flow, your
customer thinks that the
suggestions from
the agent are more
reasonable, which means
that they feel like
it's less cognitively
effortful to do
whatever the task is,
and therefore, this
shows up as greater
trust in the agent.
A bit of beneficial
friction can
beneficially impact
the trustworthiness
of AI. So if you
take the time to both
think about how
you can calibrate
the similarity as
well as the timing,
that's effort well spent.
And taking that
step can improve
not only trust
in the agent, but
the trustworthiness
of the agent.
Last but not least, I
talk to thousands of
executives, I teach
executive education at
MIT, I do all of these
talks, and I hear folks
talking about AI in
this kind of way. Do
you see anything that
these have in common?
It's about
being AI first.
I'm sure you've heard
that mentioned by
colleagues or in
the convention here,
but I ask you, if
you're AI first,
Is it the human? Is
it your customer?
in order to be
customer first,
you need to
exchange value. And
the more value
you exchange,
the more your
customers trust.
Thank you all for your
attention. I really
appreciate your being
here. Thank you.
Thank you. Thank
you to Salesforce.
And I've been asked
to put up a QR code.
So please feel free
to use this. And I'll
be hanging out if
any of you have any
questions. Enjoy the
rest of your time.