Getting Into the Flow | Business Travel News

It was three years ago when artificial intelligence-powered
corporate travel startup Skylink co-founder and CEO Atyab Bhatti began knocking
on the doors of Big Four consulting firm McKinsey & Company. He had
something exciting to show: a large language model-based corporate travel tool
that not only built dynamic traveler histories and profiles but also engaged
travel policy and provided agnostic connections to corporate travel content
sources—global distribution systems, New Distribution Capability connections,
TMC marketplaces. Bhatti and Skylink’s other co-founder, Kush Maheshwari, who
holds the chief technology officer role, had started working on the idea in
2016 while still in college at the University of Illinois Urbana-Champaign, but
with the maturity of large language models advancing rapidly, the opportunities
for Skylink to become transformative in the corporate travel space were growing
by orders of magnitude.

Unlike previous iterations of such technologies, OpenAI and
ChatGPT had transformed the form factors possible for booking tools from flat
visuals with radio buttons and drop-downs into conversational voice- and
chat-activated interfaces that could remove clutter and keystrokes and give
back time. Big time. 

That was the narrative that had captured investment from
venture capital firm Y Combinator for the duo to pursue their vision. Plus,
Bhatti was a former McKinsey rank-and-file employee—it had been a couple years
since he departed, but could that, maybe, get him a foot in the door?

“Not so much,” Bhatti told BTN with a laugh. “I don’t think
really at all.”

Bhatti and Maheshwari needed someone who just “got it.” They
needed someone who saw the possibilities of large language model interactions
tied to the right content, traveler history and policies, and who could
envision how such interactions might forward corporate travel program
objectives.

That’s when they were introduced to McKinsey director of
travel and events technology Jamie Stewart.

Phase 1: The Meet-Up

“Atyab and Skylink were introduced to me by our air category
manager approximately three years ago,” Stewart told BTN. “We just had an
exploratory meeting to take a look at what he had built so far. It looked
interesting in the sense that it could help us meet some of our business
objectives of bringing the experience of booking travel more into the flow of
work with an asynchronous messaging platform.”

Those objectives for the McKinsey program included not only driving
bookings into existing workflows but also offering more personalized results,
achieving a quicker and more satisfying booking process and—here’s the kicker—realizing
more beneficial booking behaviors across the McKinsey travel population. Which,
as in most professional services firms, represents the vast majority of employees.

But travel decision-making is complex. There are multitudes
of choices, not all of them logical or optimized for price or efficiencies. There
are clear traveler preferences along with viable alternatives that should be
offered. For Stewart, guiding travelers to the right choices was paramount, and
he envisioned an AI-powered pathway toward that goal.

“We wanted data to drive the options offered to travelers,
and we didn’t want to overwhelm travelers with too many choices,” he said. That
said, the reduced set of options had to feel curated for the traveler rather
than restrictive.

“We needed to recognize their personal loyalties and travel
history, but also the intent they’ve shared in their current prompt,”
Stewart said. “We sort of knew from the beginning that recognizing all
those elements would be critical to our success.”

Phase 2: The Set-Up

While the objectives and broader strokes of a possible
collaboration were fairly apparent from the get-go, several other elements were
more open-ended:

What content would McKinsey initially require,
and how would it flow from agency partner American Express Global Business
Travel and integrate with McKinsey’s other tech and service providers?How much more development and functional
capability would be required within Skylink to handle the complexity of the
McKinsey program?How much modification would be required to
ensure the AI was weighing the right personalization-versus-policy notes in its
content returns to gain trust across the McKinsey traveler population?And finally, how could travelers be nudged effectively
toward more optimized behaviors and itineraries?

With a full procurement process and security vetting
completed, Stewart was ready to integrate Skylink with McKinsey’s Slack
workplace collaboration channels. The idea was to create a seamless pathway to
travel booking tools, reducing employee time and motion dedicated to arranging
travel in a system outside of its day-to-day platforms. To make a travel
booking at McKinsey, the user simply messages Skylink in the Slack environment,
which initiates a dialogue.

AGENCY CONTENT & COLLABORATION

With the delivery channel settled, “We set out what our requirements would be and then built against
those together,” said Stewart. That meant pulling in McKinsey’s agency of
record, Amex GBT, which has a transformation team attached to the McKinsey
account. That team collaborated alongside the client and the startup to push
toward the AI vision.

“This was sort of novel. To source the content and
bring it into a conversational workflow was a new thing. There were times where
we really had to work together across Skylink and the TMC to identify, ‘OK,
what is the natural path forward here,’ ” Stewart recalled. “There certainly
were times that what we envisioned from the outset wasn’t feasible with the
technology in place. But we’ve worked around those pretty well. Everyone
understood the business case for what we were trying to do and has been
supportive.”

While Amex GBT has been testing a number of
technologies for internal AI development—its Egencia product recently announced
AI enhancements to its virtual agent, for example—the company said it’s also
pursuing innovations with clients who are invested in their own vision.

“Customer-led innovation is absolutely part of the
equation at Amex GBT,” said chief product and strategy officer Evan Konwiser. “Certainly
with a customer like (McKinsey), we can see Jamie’s vision, (and) it’s great to
be able to… bring the resources to the table to support it.”

McKinsey came to the table with the right innovation
at the right time, according to Konwiser.

“The idea of booking a trip
in a Slack window is something that I think even a year or two ago people would
have felt a little bit, if not uncomfortable, (that) it wouldn’t have been a
natural type of adoption mechanism,” he said. “Now that they’re personally doing
similar types of search and transactions in a ChatGPT window, for example,
suddenly (booking travel) in that kind of conversational channel makes sense. You
can’t underestimate how important that is for the adoption curve. … One of the
best things we can do is adapt quickly to user behaviors that people already are
adapting in their personal and work lives, and that’s what we see happening
here.”

Skylink and GBT currently pull
in only Sabre content for the McKinsey project. Though it has the ability to
accept other sources, like NDC or specialized hotel marketplaces, those have
not been a critical need in the North American market where McKinsey initially
piloted and has rolled out the AI platform. “We will need to think about that
more as we look to roll out more geographies and with other GDSs,” Stewart
said.

McKinsey also is an established Concur client, and
while the Skylink tools do not integrate with the Concur booking pathways, they
are configured to off-ramp users to the more traditional booking platform if the
AI cannot complete requested travel bookings in the Slack pathway.

“Even two years ago, our research suggests a semantic
search in Slack was important, but it really needed a quick agency escalation,
otherwise it would be a fail,” said Konwiser. “Now, we’re learning and
experimenting that, OK, you need to get to an agent and you may need Concur, but
it doesn’t feel like it needs to be fully embedded.”

Konwiser added that he felt the industry was at the
precipice of major change.

“I don’t want to say that online booking tools’ days
are numbered, but I would say AI capabilities are enabling user experiences
that are the next leap forward.” He described the historic shift from phone
bookings to self-service OBT adoption in the early 2000s that 25 years later
have garnered 80 percent to 90 percent online booking adoption across GBT’s
different brands. “We think the AI shift will be equally significant, but it’s
really early days. Hopefully it doesn’t take 25 years.”

The McKinsey travel program is forging that path.

“We expect Skylink will eventually support the vast
majority of bookings,” said Stewart, “but we never want to leave McKinsey employees
without travel booking support.” That includes the consultative support GBT
agents provide, he said. “We are looking to effectively utilize the support
channels we have.”

Indeed, even the Skylink team is open about the fact
that there are a number of complex functions the LLM channel can’t accomplish.
“Agents are the backbone of this industry,” Bhatti told BTN, citing a web of
underlying technologies in the travel industry that will continue to require
interventions for some time.

PERSONALIZATION & ITERATION

In terms of content and booking, Stewart told BTN
that Skylink worked well pretty much out of the box. The startup was subject to
all McKinsey’s standard security reviews, and as with most companies of
McKinsey’s size and sector, data privacy was critical.

That can be tricky when it comes to personalizing content
but both Skylink and Amex GBT were impressed with the ingenuity and
accountability Stewart exhibited in making privacy and personalization come
together.

Skylink for McKinsey works from two sets of profiles.
On one hand it accesses the GDS profile, which houses the traveler’s loyalty
information along with the personal data the system requires to book a travel
itinerary. On the other hand, the Skylink system holds onto a traveler’s
historical bookings and, after assessing intent as expressed in the prompt,
will use the appropriate history to inform what travel options are returned to
the user.

“We built the AI off of
best-in-class models like you see on Facebook or other social platforms,” Bhatti
told BTN. “What we started to realize is that you should (personalize) off
historicals, but you should update that because behaviors change over time as
well.”

But Stewart wasn’t only interested in personalizing
those trip options. He also wanted policy-compliant bookings, and a core
objective was to shift noncompliant and not-well-optimized travel behaviors to
be more in line with McKinsey’s travel policies.

It was Stewart’s
thoughtfulness about the possibilities inherent in a new kind of channel
interaction that impressed Konwiser.

“This project hasn’t really
been about shifting a booking from Concur to a Slack window or from an agent
offline to a Slack window—not entirely,” said Konwiser. “Jamie is talking about
how behavior changes and how does a booking path fundamentally
give new opportunities to change how travelers are purchasing in context of
what the company is interested in and (what the individual) is interested in.”

That wasn’t necessarily something Skylink’s
fundamental tool set was set up to do, at least not in the way that McKinsey wanted
it.

“We iterated a lot to find the right balance between
traveler preferences and policy-driven options,” Stewart said. “There were
things we learned in the course of working together that were user-interface
and user-experience enhancements. Language became clearer, and notices that we
wanted to give to people in the process needed to become more effective.”

That balance, Bhatti added, is supported by multidimensional
AI “thinking” behind the scenes—because you can’t influence a smart business
traveler by showing them choices that they clearly will not take. “The tool has
to show that it understands the traveler’s intent,” said Bhatti, and based on
that intent will present viable options.

“We have a pretty complex algorithm that takes the
response and reorganizes it into an algorithm that we call Trip Rank. Trip Rank
then updates (the returns) to provide the most relevant results to the traveler
that are also smart to take.”

Phase 3: Feedback & Results

In the pilot stage in October 2024 and since its North
America launch in April, every McKinsey travel search initiated in the Slack-Skylink
workflow returns three air itinerary options. The options clearly note which is
considered the lowest logical fare selection and prompt those booking
international itineraries to choose roundtrip fares instead of one-way options.
The Slack conversation always asks the user to book a hotel within the same
workflow.

“People approaching the tool know it’s an AI-enabled
workflow and that they’ll have a reduced option set. We are totally transparent
about that,” said Stewart. “Getting that option set right establishes trust and
is really important to sustained adoption.”

While bookers have the leeway to decline Skylink
recommendations, Stewart shared that 85 percent of users from the outset were
choosing from the first three options. “There’s probably some selection bias
there (among users) in terms of being excited to use the tool, but we’ve seen
that sustained through the rollout” in North America.

Moreover, 96 percent of users have given the Slack-based
workflow a thumbs-up rating. Net Promoter Scores, Stewart said, “are on par
with some of the best consumer tech out there.”

McKinsey has seen migration toward Skylink from users who
previously preferred live agent bookings and from the more traditional online
booking tool environment. All the collaborators found it was important to keep
the asynchronous messaging environment for Skylink as primary.

Efficiency is at the heart of that feedback. The
asynchronous strategy allows people to stay in the zone of work, rather than
break away for a side convo. To that end, additionally, time and motion data of
McKinsey Skylink users shows an average booking time of 99 seconds, compared to
nearly 15 minutes in more traditional channels.

Embedded in that calculation are cool features like simultaneously
making bookings for repeat trips. A booker simply needs to request through
Skylink to repeat the booking for a given duration of weeks. For example, six itineraries
that include hotel can be booked via the same request. Likewise, travelers with
multiple trips can book them all in the same thread with the large language
model able to parse the trip stops and starts as well as different city pairs
within a given prompt.

Adoption and attention to behavioral nudges continue to
climb. McKinsey has tracked a 35 percent shift in users toward hotel attachment
and 40 percent shift in the company’s Skylink users toward booking
international roundtrips. The multi-week repeat trips as well as the ability to
book multiple trips in a single prompt has driven up advance purchase stats for
airfares.  

“That’s evidence of the advantage of having someone like
Jamie. He not only understood the problem statement, he was able to translate …
that to stakeholders and put it in the language and key performance indicators of
finance and procurement. And he also brought it all into the travel world,
which has so much complexity and nuance,” said Bhatti.

On top of that, he said, “Jamie is very kind. Not everyone
at big companies treats startups with as much respect as Jamie treated us.
Sometimes as a startup you are treated like a nobody, but Jamie was always
collaborative and always looking for ways to drive success.”

That success wasn’t just about driving a conceptual win—that
an AI-powered LLM could be applied to a travel program, for example, said
Konwiser. “He actually took a scenario like AI around which there has been a
lot of hype and delivered real value. That’s pretty awesome, and it doesn’t
happen every day.”