Beyond Chatbots: Deployable AI Intelligence
Key Points
- The hype around chat‑based interfaces overstates AI’s true potential; we should view large language models (LLMs) as general intelligence that can be embedded throughout applications, not just as a chat window.
- LLMs represent “deployable intelligence,” meaning they can be assigned tasks much like a high‑performing employee, with future versions gaining more autonomous, agent‑like abilities.
- Even as LLMs become more autonomous, ultimate accountability remains with the human delegator, especially because AI still lacks true business judgment, which relies on nuanced, context‑dependent decision‑making.
- Current LLMs excel at processing explicit information but struggle with the implicit, political, and cultural cues that underpin effective business decisions, limiting their ability to replace human judgment.
- Designers need to broaden their mindset to integrate LLM capabilities across diverse workflows and interfaces, moving beyond the narrow chatbot paradigm to unlock richer, more functional AI deployments.
Full Transcript
# Beyond Chatbots: Deployable AI Intelligence **Source:** [https://www.youtube.com/watch?v=ODLyBQd4VHU](https://www.youtube.com/watch?v=ODLyBQd4VHU) **Duration:** 00:17:29 ## Summary - The hype around chat‑based interfaces overstates AI’s true potential; we should view large language models (LLMs) as general intelligence that can be embedded throughout applications, not just as a chat window. - LLMs represent “deployable intelligence,” meaning they can be assigned tasks much like a high‑performing employee, with future versions gaining more autonomous, agent‑like abilities. - Even as LLMs become more autonomous, ultimate accountability remains with the human delegator, especially because AI still lacks true business judgment, which relies on nuanced, context‑dependent decision‑making. - Current LLMs excel at processing explicit information but struggle with the implicit, political, and cultural cues that underpin effective business decisions, limiting their ability to replace human judgment. - Designers need to broaden their mindset to integrate LLM capabilities across diverse workflows and interfaces, moving beyond the narrow chatbot paradigm to unlock richer, more functional AI deployments. ## Sections - [00:00:00](https://www.youtube.com/watch?v=ODLyBQd4VHU&t=0s) **Beyond Chatbots: Deployable AI** - The speaker argues that AI should be seen as deployable intelligence rather than merely a chat interface, urging designers to envision LLMs as agents integrated across applications. ## Full Transcript
AI is not a chat box it's just not and I
want to talk about that because I think
that we have overstated the value of the
chatbot because we have seen how
powerful the release of chat GPT 3.5 was
in the world landscape and that specific
release changed our perception of what
generative AI is capable of that's great
but one product release does not make a
user interface and so I actually want to
talk through where I think the
capacities of llms are going and then I
want to revisit and say how does this
not fit a chatbot interface because I
want you to walk away from this with
some more open Horizons around where we
could be designing uh llms into our
applications we need more designers
thinking about how llms could be more
effectively deployed outside that little
chat window that you see in the corner
okay
so the first thing I want to call out
there will be five of these I want to
call out on the capacities of AI by the
way so number one AI should be thought
of as Deployable intelligence in other
words we've never really had
intelligence that we can deploy
management Theory would say that you can
deploy intelligence among your employees
but if you're working with fairly
high-powered knowledge workers they are
deploying their own intelligence
autonomously and they should be that's
great AI is directly Deployable by the
person managing it and that means on the
one hand that you have to put time into
managing it and I think that we're going
to start to see llms develop more
agentic properties so they'll be able to
do things autonomously more over the
next couple of years which means
eventually you're going to get to a
point where you can delegate in a way
that's sort of similar to the way you
would delegate to a high-powered
employee and they will just go off and
and go after a particular task or
particular objective
independently that's not where we are
today but I I think even if we get there
Deployable intelligence is still the
right frame because at the end of the
day they are going to be accountable to
the person who is driving that
delegation in a way that an employee is
not an employee is expected to exercise
business judgment and an
llm may give you business perspective
but I think their ability to exercise
business judgment is going to be one of
the last things that AI is able to
conquer so to speak and the reason for
that is that business judgment is a very
context dependent art and it's very very
difficult to get right and I actually
haven't seen a lot of evidence that llms
are getting better at it despite their
progress in a lot of other areas they
are certainly getting better at talking
business but business judgment is
actually an act of decision against a
particular set of data and I think part
of why llm struggle with that is they
can't really take in a lot of the
implicit data that business judgment is
grounded in they take in explicit data
and so if you're looking at the
complexity of the politics in a
particular organization as part of your
business judgment llms can get that if
you exhaustively describe it to them but
that's still a low Fidelity picture of
what's really going on because they
can't experience it that's just one
example so wrapping that back to the top
fundamentally if we think of AI as
Deployable
intelligence we will have a much more
flexible view of where it can be
deployed it's not just in a chatbot all
right number two AI is fast and
therefore cheap and I think that is one
of the fundamental attributes that we
have forgotten chatbots are designed to
be callon response right like you're
supposed to have an interaction what we
miss there is this idea that llms are
fundamentally much much faster at doing
a bunch of tasks once they're trained
than humans are and so even if they do
them at lower Fidelity if we are
tolerant of a lower Fidelity task done
with additional
guidance we you're going to go way
faster on a whole range of knowledge
working tasks and that's in anecdotally
what I see AI being used for in
organizations people are using it to do
a bunch of tasks that would previously
take a long time and they're just
getting it done
faster all right uh number
three AI is opening up new business
models because of that cheapness and
speed and I want to call that out
separately because we are just at the
beginning of uncovering what that looks
like
I think that what we see with the uh
layoffs at McKenzie is a good example of
this McKenzie basically functioned on
this idea that human intelligence is
something that is hard to replicate
anywhere else and so if you have human
intelligence as a
consultant you sort of have for lack of
a better term a monopoly or a corner on
the ability to provide consultant
services and I think what McKenzie and
others are discovering is that at the
end of the day a lot of people are using
chat GPT as their cheap MBA
consultant and even if it's not quite as
good as McKenzie uh depending on your
opinion of
McKenzie it's still good enough and
that's really the fundamental issue a
substitute doesn't have to be 100%
equivalent it can be good enough and
McKenzie is finding out that a lot of
companies in an era of belt tightening
when the cost of cash is
higher they're going to be looking for
cheaper options and there's this new
technology right at the CEO's fingertips
that basically functions as an MBA on
top
so the reason I'm calling that out is
that we have gotten to a point where
we're starting to see the old business
models disrupted but we have not yet
gotten to a point where we're seeing the
new business models come into play and I
think one of the interesting hints at
new business models is the insistence
companies have on training their own
models they want to see their own models
ins inside the house trained on their
data that are effectively private models
that they can use the way they see fit
and I think that that may be a hint that
we are seeing a move toward for lack of
a better term vertically integrated
intelligence so there's this idea in
business that if you vertically
integrate you gain efficiencies right
you reduce costs up and down the supply
chain what we're seeing here is
potentially that businesses want to
vertically integrate the intelligence
stack and they therefore can do more
with what they have if they are building
inside the house they save costs but
they also save on risk right like they
know what that AI will produce they know
what that model is capable of and I
think the risk piece is bigger
anecdotally than the cost savings piece
right now no one wants the debacle from
Air Canada where the chatbot started to
issue fake policies and people see
vertical integration is a way to handle
that so my guess is the new business
model disruption may actually look a lot
like vertical integration and what's
interesting is that may Empower smaller
firms who are tightly vertically
integrated and who have a compelling
value proposition to compete really
really effectively because they can move
fast as a business and they're moving
fast as a business because like AI is at
the core of everything they do we will
see it's a guess uh but I think
fundamentally we're at a stage where
we're sort of at that crossover point
where we're starting to see the
disruption of the old models like
McKenzie and we're starting to just
begin to see hints of the new business
models okay number four smart deployment
is a function of Habitual access and so
this is skipping over to the UI side of
things if you are building for llms you
need to build assuming that you have to
meet a very low friction bar for any
kind of success and I see a lot of
people assuming that a low friction
bar means what it doesn't for lack of a
better term people people assume a
degree of commitment to using
AI that is
absurd I've seen people say oh they'll
go to the special page on my Wiki and
they'll use my special model they're
just going to use chat GPT because
that's the default in their heads like
that's one of the consequences of having
a highly successful external product is
that people have been wired to use this
UI and even if it's not an ideal UI
they're still using it because that's
the default habit and so if you want to
deploy an llm you have to be thinking as
a product person how do I reduce the
friction in this space so that it is so
easy to be habitual with this where are
people already in my space and where can
I put the llm right there where they're
where they're at so they have the habit
of using it in the spaces they're
already in I think that's really
important to drive effective usage okay
fifth one generative AI means
hallucination it means it because you're
gener generating data and I think that
when people talk about hallucination
they keep talking about it like a defect
and it's not it's built into the system
you're designed to generate data well
it's going to generate
data and generating data means
inherently some of the data is not going
to be factual and so we need to get past
the point where we assume Hallucination
is a defect from a technical point of
view and we look at it as an undesirable
outcome from a business point of view
and there's two comments I have there
number one I think there's a billion
dollar opportunity out there for a third
party who can validate the factual
accuracy of llm outputs maybe that's a
service on tap where you just call them
and they validate it I have no idea I'm
not building it but somebody should
somebody should think about the problem
of factual accuracy as if it's a first
class problem because it's about to be
something that people are going to pay
big bucks to be sure of and that is a
lot of the reason why companies are
putting all this effort into fine tuning
and building those internal models as I
touched on earlier they want to be in
control of the risk imagine a world
using the Air Canada example where the
llm had to call a factchecking service
before it could put that policy language
out there in front of the customer it
would be a different world we wouldn't
have that story in the news because the
llm would have given the correct
response and so businesses May build
this internally using tools like Lang
chain but fundamentally I think that
factual accuracy as a service is going
to be an interesting opportunity and now
I I grant you learning the facts inside
the business is something that a third
party service May struggle with so
there's still going to be that internal
element but we have a lot of widely
accepted facts around the world that are
publicly available that are in
newspapers that are in scientific
Publications those are all things that a
third party service could act as a Data
Bank for and fact check against Food For
Thought I will also say the second Point
here on generative Ai and hallucination
I have seen this personally get better
and I would bet that even though we may
need a thirdparty industry or service of
some sort to to validate the outputs of
generative AI we are still going to be
smart to bet on progress from the man
model manufacturers here open Ai and
others are going to get better at
producing more factual AI outputs and
they're not going to talk about how they
did it but they're going to get better
and I've seen that since starting to use
chat GPT 3.5 is that it start it's
getting better every generation at
producing outputs that are factual and
that is reducing my risk radar which is
dangerous in one sense because I
probably should be more thoughtful about
those edge cases that still emerge than
I am but as a human being I have to make
risk assessments all day and I am
starting to find that it is Dependable
enough that I am relaxing and trusting
it more and that's a very intuitive
thing and maybe that's just me but my
sense as those models are getting better
over time as open Ai and others who are
building those models start to get more
and more emphatic about driving clear
factual output so that they don't get in
trouble anyway let's recap five things
that militate against this idea of a
chatbot and I'm going to kind of get
into how chat
Bots are ineffective in the next section
but this is the first section so the
five things where AI is Deployable
intelligence which means it's much more
widely deployed able than a chatbot AI
is fast and therefore cheap which means
that it doesn't have to be limited to a
chatbot right AI opens up new business
models that's more of the business model
side of things AI is all about smart
deployment as a habitual access point
and so that's a UI thing where where I
was talking about the fact that you have
to put the llm where people are and not
just assume people will go to a special
page or whatever to use it and finally
generative AI means hallucination and we
should stop pretending it doesn't and
that also has chatbot implications
because chatbots and we're just getting
into this second section here right the
this second section of the video is all
about the three flaws that I think are
really killing the chatbot as a UI for
llms so the first one that I want to
call out uh and I just jumped into it is
that chatbots simulate a human
conversation and that means that they
cause hum humans to overstate the
veracity or the accuracy of llm outputs
the problem is that humans believe
chatbot outputs more because it looks
like the kind of conversation we're used
to having in text messages with other
humans all day
long and that's a big problem and I
think if we had a different UI we might
be tuned to believe chatbot outputs less
compellingly right like we would be less
likely to believe them second reason or
second issue with llm uh in
chatbots llm capacity continues to
evolve but the chatbot interface largely
is not I think anthropic has done a
little bit of a push on that by S of
starting to develop a canvas where the
chatbot can paint something on the side
like a chart and that's a great step but
in
general the chatbot interface is staying
static as models evolve and so people
are actually not really keeping track of
how models are evolving because there's
no indicator to think to say in the
interace face that chat chat GPT 40 is
smarter than chat GPT 3.5 if you are
like me and you study this all day then
you know that they are but if you're
just a casual user using the chatbot it
looks exactly the same and apple figured
out that you have to make the iPhone
look different car manufacturers figured
out you have to make the car look a
little bit different in order to sell
the new car and I think there's
something some insight there around
human psychology in sort of how we
handle interfaces with evolving models
and I kind of want to put a pin in that
for a future
conversation okay the third reason that
chat Bots are an issue chat Bots require
Advanced llm knowledge because it's just
a text input box it's up to you to know
enough about the llm to use it well and
I have literally seen sidebyside
examples where I've been prompting next
to someone else at the same time and I'm
getting different results and it's not
just that the generative AI is producing
hallucinations it is that the generative
AI is something I am more able to prompt
because I know more about it
and I know more about how it works than
the person sitting next to me and that
is fundamentally inequitable and I mean
that in the sense that if the Deployable
intelligence is there we should be
developing user interface principles
that allow anyone to access that
Deployable intelligence it should not be
dependent on the ability to read an
exhaustive number of Twitter posts and
read a bunch of articles about how llm
works and prompting in order to drop an
effective conversation with a chatbot
you should be able to Max the capacity
without that esoteric or hidden
knowledge you just should okay so where
did we end up with all this I wanted to
call out sort of that those five themes
that I did in the first half of the
video around how AI is changing in ways
that make it bigger than a chatbot and
then I wanted to spend the second half
talking about the three issues I see
specifically with chat Bots because I
think we don't talk about them enough so
so I talked about how chat Bots require
Advanced llm knowledge I talked about
how llm capacity is evolving in ways the
in for interface is not and I also
talked about the fact that chatbots
simulate human conversation in a way
that probably causes us to overstate
their intelligence and the veracity of
facts that they give us I think those
are all concerning so where I net out on
this is I think we're in an inflection
point where we need both new business
models and also new UI models for large
language models like I I just think we
do and I know overused the word models
here but it's an AI talk so what do you
want um so that's where I'm netting out
and I would be interested to hear
examples in the comments of where you
see Mo business models or user interface
models that you think are a direction
that the future could build on as far as
how we deploy large language models in
Tech because I'm always looking for new
examples all right this has gone on long
enough cheers