These slides are for the following session presented at the UX STRAT Online 2021 Conference:
"Design Tools to Get the Most from AI"
Adilakshmi Veerubhotla
IBM: UX Architect
3. Note
Designing for AI is fundamentally based on
the same principles we use to design quality
user experiences (UX) and product strategies.
However, a special lens is required towards
certain principles.
This talk will focus on bringing that lens to
âš
the forefront.
3
4. Agenda
The case for
âš
design influence on AI
âš
Cultivating a
âš
perspective to shape AI
âš
Planning towards
âš
tangible design impact
4
1
2
3
6. Why designers should influence AI
AI is just a
means,âš
not the
end goal.
6
Designers can help
define user outcomes
to achieve through AI
solutions
7. Why designers should influence AI
AI is not
âš
a magic
cure.
7
Designers can help
identify the right
problems to tackle
with AI
8. Why designers should influence AI
Businesses
must have
responsible
future plays.
8
Designers can help
bridge business value
with responsible
applications of
emerging technologies.
11. Outcome focus
11
Agree on
the user
problem
to be
solved.
We need to get
ahead in the market.
Letâs build some AI.
Okay. Whatâs our
objective with it?
We have to use our
data to drive insights.
Can you describe the
user value? Letâs try
it together âŠ
12. 12
Outline a
solution &
identify
where AI can
fit into it.
What will the user do before
and after this AI part?
Then,
what?
User
contacts
support
User objective: Get quick resolution to issue
User
sends
details
of
problem
* Uses AI models to
extract issue and
classify type
System tags
issue type,
engages
right team
Context awareness
13. 13
Within that
scope,
define how
your AI
models
should
behave.
Start
Stop
Option 1 Option 2
User action
System action System action
System action System action
Complete?
Yes
No
Detailed plan
14. Pro-tips
14
Explicitly define expected
user value in journey.
Business outcomes are not
equivalent to user value.
Understand what defines
âsuccessâ to users in this journey.
Derive value from there.
âš
â
E.g: full resolution time (vs just
first contact time); a synthesized
answer (vs a list of data points)
Rephrase the promise of AI
without buzzwords.
Re-phrasing value proposition of
AI models will help you notice
implicit data dependencies and
potential biases to work through.
âšâšâš
â
E.g: replicating an existing, well-
defined manual process;
modeling a potential scenario
based on historical data
Balance effort + value of AI.
âš
Building high quality AI models
takes effort. Sometimes, a
creative workflow can achieve a
similar gain as an AI model might.
Ensure that AI provides justified
value to your problem.
â
âš
âš
E.g: an AI model that can infer
intent of user inquiry (vs asking
user to choose from a defined list
of intents)
16. Example 1
Imagine an AI model to detect an issue
mentioned in a free form text field called
âissue descriptionâ.
Assume that a subject matter expert (SME)
has to label a few hundred examples to
effectively train this model.
Balancing user
expectations with
model needs
16
17. 17
Our example workflow here assumes that
the model can get all those examples.
That is an exhausting situation for a user.
What if the user does not label all
examples?
User begins
labeling examples
Labels all
examples
Labels some,
ignores some
??
Model learns
correctly
What if ⊠the user quits
without finishing?
Example 1: Balancing UX with model needs
18. 18
What if there are negative
consequences associated with partial
labeling?
That makes the task heavier for the
user. The probability and impact of
quitting are high.
Oh, can you fix the
AI to avoid this
confusion for users?
Uh, Iâm not sure. Itâs
not that simple. This
model learns with
positive & negative
context.
⊠and, thereâs heavy
penalty for quitting?
Example 1: Balancing UX with model needs
19. 19
What if we asked for
examples in multiple
rounds?
Hmm..interesting. I
wonder if we can build
the labeled set over
multiple rounds.
Explore
different
workflows
to reduce
load on
users.
Example 1: Balancing UX with model needs
20. Example 2
Imagine an AI model that has to learn to compare
text from two files, and raise a flag if they are not
similar to each other.
Assume this is part of automating a business process.
20
Balancing user
expectations with
technology
constraints
21. 21
Compare text in
template and file
âYou should wait
to do Xâ
âYou should NOT
wait to do Xâ
Does text
look more
than 90%
similar?
Similar.
Good to go.
Yes
Not
similar.
Flag it.
No
Simple situations like these are complex edge
cases for AI technology.
While a human can easily tell that those two
sentences mean opposite things, AI models
might struggle, given how similar they look.
Example 2: Balancing UX with tech limitations
What if ⊠thereâs a
curveball for AI technology?
22. 22
Compare text in
template and file
âYou should wait
to do thisâ
âYou should NOT
wait to do thisâ
Does text
look more
than 90%
similar?
Similar.
Good to go.
Yes
Not
similar.
Flag it.
Review
differences
No
Example 2: Balancing UX with tech limitations
Make
human
checks &
re-balances
easy to do.
23. Example 3
Imagine an AI solution which automatically alerts a
crew to quickly respond to a situation.
Assume this solution combines multiple AI models
together to achieve this objective.
23
Balancing user
expectations with
model needs &
tech constraints
24. 24
When a composite system has a failure
point, there can be a domino effect.
What if the AI model to detect text from
image fails? Or reads incorrect text?
How will it impact the final outcome of this
composite system?
And how will that impact be felt by users?
Read text in image
Extract
values
Classify based
on values
Invoke process X
based on
classification
Example 3: Balancing UX with model needs & tech limitations
What if ⊠thereâs a weak
link in the chain ?
Get an image as
input
25. Example 3
25
Unexpected
outcome.
âš
Review ?
Read text in image
Extract
values
Classify based
on values
Invoke process X
based on
classification
Get an image as
input
Uncommon word.
Please review?
Decide
when to
call the
userâs
attention.
26. Pro-tips
26
Systems thinking is crucial.
âš
You must learn to envision a
whole as more than the sum of its
parts. Youâll have to first make the
parts fit together and then fuse
together for composite value.
Adapt UI patterns for
complexity.
Typical web UI patterns are
optimized for simplicity and
predictability, with least amount
of friction.
You must learn to balance user
expectations with modified UI
patterns for complexity, friction &
unpredictability.
Document expected
behavior of AI models.
Variable outcomes are a defining
characteristic of AI models.
But your team must document
expected patterns of input &
output leading to model behavior.
This is crucial to make the model
less of a âblack boxâ.
28. Evolving tech for users
AI models
need lots of
training data.
28
How might we reduce
load on humans to
train AI models?
29. Evolving tech for users
AI model
building is
being
democratized.
29
How might we help
users make the right
decisions about their
AI models?
30. Evolving tech for users
High value use
cases have
data with
nuanced
interpretation.
30
How might we scale
AI tech to tackle
complex, high value
problems
31. Pro-tips
31
Learn to identify the right
trends to care about.
Adoption of new technology can
effect trends over time.
The perceived value of use cases
shifts over time. A userâs mental
model, behavior and skills evolve
over time.
Choose which problems are worth
caring about now vs later.
Break down complex
problems.
There are no easy solutions to
many problems.
Are you suggesting a change
doable in a few months, or is it a
5-year research project?
To chip away at complex
problems, break them down into
smaller chunks.
Be experimental.
To learn quickly while still making
progress in the right direction, you
must learn to craft good
experiments - both for good user
experience and for good AI
models.
A good experiment has a well
defined hypothesis and clear
indicators of success or failure.
33. âExecution planâ cheatsheet
Be clear on purpose
âą Outline user problem
âą Define whatâs âvalueâ
âą Identify the specific role
of AI in that value
Plot the full journey
âą Identify user goal
âą Lay out end-to-end user
journey, with steps
before & after the AI
parts.
Plan the details
âą Define AI model behavior
through system flows.
âą Identify failure points.
âą Refine user workflow
appropriately.
Prepare for surprises
âą Outline potential
curveballs on multiple
factors
âą Craft solid experiments to
test and learn
âą Refine user workflow
33
The âWhyâ
Enables you to make
decisions for optimal
value. Steps to take if
things donât go per plan.
The âWhatâ
Tells you what effort you
are taking on, and what
success are you working
towards.
The âHowâ
Aligns a cross-
disciplinary team on a
sequence of actions that
must be executed.
The âWhat-ifâ
Enables you to prepare
for unpredictable
outcomes of variables.
35. 3 timelines of work
35
Immediate
Goal:
âš
Tactical design of
AI-driven solution
Near term
Goal:
âš
Planning a useful AI-driven solution
Long term
Goal:
âš
Shaping the future of AI technology, user behavior
and business opportunities
Focus:
âą Design system & UX
behavior
âą Build high value
moments
Focus:
âą Tell a valuable story of this
solution, from a user perspective.
Focus:
âą Explore potential futures that are in line
with evolving system & user needs
Example artifacts:
âš
Story maps, system
flows, user flows,UI
patterns, micro-
interactions etc
Example artifacts:
âš
âConcept carsâ, end to end user
journeys, usage scenarios, MVP
slices, high level scenario maps etc
Example artifacts:
âš
Vision concepts, speculative stories, trend
spotting, horizon maps, experience vision
statements, strategic UX roadmaps etc