In this episode
Chris Willis, founding Chief Design Officer and Futurist at Domo, argues that the organizations struggling with agentic AI are not the ones moving too slowly — they are the ones moving fast on a data foundation that never earned internal trust. Willis walks through why customer-facing trust gets most of the attention while internal data trust sits on far shakier footing, why agents fail at the first exception nobody wrote down, and why “engineering for trust” is the wrong goal entirely. His alternative is concrete: build for visibility, predictability, and control, judge every candidate workflow by what it costs to verify rather than what it costs to generate, and treat adoption — not outcomes — as the first signal that any of it is working.
Key takeaways
- AI inherits whatever confidence you already have in your data. If teams keep shadow spreadsheets and argue about numbers that should be certified, an agentic layer does not resolve that — it restates the same shaky numbers in a far more confident voice.
- Internal data trust is on shakier footing than customer trust. Organizations that spent years and budgets on warehouses, lakes, and lakehouses are discovering through agentic projects that the foundation underneath is weaker than they assumed.
- The real cost of deploying before the foundation is ready is predictability. Confident-sounding machines are easy to produce; systems that behave the same way in the same situation are not, and that is the step most often skipped.
- Undocumented judgment results in undefined behavior. Willis’s example: a customer-facing agent hits a promo exception that a human manager would have carved out from experience. Because nobody codified it, the model improvises and drifts off the goal. The debt comes due at the first exception that only ever lived in someone’s head.
- Models have jagged intelligence — they fail without signaling the failure. A model can perform well on one variant of a task and badly on a slightly different one, and it cannot absorb an organization’s norms just by observing them.
- Trust is not something you can engineer for directly. Willis pushes back on the “design for trust” framing he encountered at recent design roundtables: you can only engineer the three things that produce trust — visibility (can people see what the agent actually did), predictability (same situation, same behavior), and control (can you stop it or adjust it when it goes wrong).
- Judge automation by cost to verify, not effort to generate. Plot effort-to-generate against cost-to-verify and keep work below the diagonal. Thousands of agent-generated landing page variants are cheap to make and expensive to check; that is the wrong side of the line.
- Adoption is the leading indicator; outcomes lag it. Willis has watched internal tools move from 2–6% usage to roughly 80% — and only after that did cost savings, vendor displacement, and satisfaction gains appear. Marketing tools that amount to glorified chatbots rarely clear the adoption bar at all.
- Expect roughly a year, and expect it to come from many small bets. The marketing organizations Willis cites did not run one moonshot; they ran many small experiments, and the payoff arrived as hundreds of hours saved and pipeline lift after adoption took hold.
- Governance is becoming a design problem, not a technical one. Willis expects governance to move into the interface — something people interact with directly — alongside the emergence of management layers where agents supervise other agents.
Chapters
- 0:00 — Intro
- 1:08 — The real AI risk: speed on a foundation that hasn’t earned trust
- 4:23 — Internal data trust is shakier than customer trust
- 6:40 — AI inherits whatever confidence you already have in your data
- 8:07 — Why the cost of moving fast shows up as unpredictability
- 9:05 — Jagged intelligence and the promo exception nobody codified
- 11:20 — Why you can’t engineer trust directly
- 12:03 — Engineer for visibility, predictability, and control
- 16:08 — The verification problem: effort to generate vs. cost to verify
- 19:55 — Guardrails start with an unsexy, strong data foundation
- 22:19 — Adoption is the leading indicator; outcomes lag
- 25:45 — Agents managing agents, and governance as a design problem
- 27:10 — Staying agile: tools that don’t come with a manual
Why internal data trust is the weaker link, not customer trust
Marketing conversations about trust almost always run toward the customer relationship. Willis makes the case that the more urgent gap is inside the building. Companies that invested heavily in data warehouses, data lakes, and lakehouses are now running agentic projects and finding the foundation shakier than expected. The tells are mundane and everywhere: teams maintaining their own shadow spreadsheets, people arguing over numbers that should already be certified and validated. Agentic systems do not paper over that condition. They inherit it, and then deliver it back in a register that sounds far more authoritative than the underlying data warrants.
What “undocumented judgment results in undefined behavior” means in practice
Willis’s sharpest formulation applies directly to marketing operations. Consider a customer-facing agent that encounters a promotion with an exception attached. A human customer service rep or manager would likely already carry a mental carve-out for that case and apply it confidently. The agent has no such carve-out, because nobody ever wrote it down — so it improvises, and drifts away from the intended goal. The failure isn’t in the model; it’s in the organizational knowledge that was never made explicit. This is why Willis frames context debt as coming due at the first exception, not at deployment.
Why Willis refuses to “design for trust”
Asked repeatedly at design roundtables how to engineer or design for trust, Willis declines the framing — not because trust is unimportant, but because it isn’t a thing you can build directly. What Domo engineers for instead is three separable properties. Visibility means a person can see what the agent actually did, surfaced at the interface rather than buried in a log. Predictability means the same situation produces the same behavior, with variance in identical use cases treated as something to measure. Control means the ability to intervene when something goes wrong, with the ambition that control becomes less necessary over time. Trust, in this account, is the output of those three, never the input.
The verification problem: the chart Willis draws for every team
Willis sketches two axes: effort to generate on the horizontal, cost to verify on the vertical, with a diagonal separating work that belongs to machines from work that doesn’t. The goal is to keep problems below the line. Historically, landing page copy carried natural constraints — a person wrote it, and it moved through layers of review. An agent producing thousands of near-identical variants removes the constraint on generation while leaving verification cost untouched, or raising it. Vibe coding lands in the same quadrant: cheap to produce, but what SQL query actually ran behind the revenue figure, and will the next run use the same WHERE clause? The operating rule Willis draws from it is to automate what can be verified easily, route the rest to human judgment, and where possible decompose expensive-to-verify work into verifiable pieces.
Why adoption is the metric to watch first
Most organizations track outcomes from their AI investments, which Willis considers a lagging indicator. The smart marketing teams he has observed watch adoption instead. The pattern is consistent: one person identifies a problem that was difficult to solve but not difficult to describe, builds a focused tool for it, and usage climbs from 2–6% to something near 80%. Once that signal appears, the outcomes follow — cost savings, vendor displacement, satisfaction gains, pipeline. What does not produce that signal is a moonshot innovation project, or a marketing tool that amounts to a glorified chatbot. Willis’s timeline expectation is roughly a year, arrived at through many small bets rather than one large one.
FAQ
What does Chris Willis mean when he says AI inherits your data confidence? An agentic system does not improve the reliability of the data underneath it. If an organization’s people don’t trust their own numbers — evidenced by shadow spreadsheets and disputes over figures that should be certified — the AI layer will reproduce that unreliability while sounding considerably more confident than the data justifies.
Can you design or engineer for trust in AI systems? Willis says no, not directly. Trust is a downstream result. What can be engineered are visibility (seeing what the agent actually did, at the interface level), predictability (the same behavior in the same situation), and control (the ability to stop or adjust the system). Build those three and trust follows.
What is “jagged intelligence”? Willis uses the term for the way models perform strongly on some tasks and fail badly on tasks that differ only slightly. The failure mode matters more than the failure rate: a model cannot absorb an organization’s norms by observation, so a small exception outside its training produces improvisation rather than a flag.
How should a team decide what to automate with agents? Compare effort to generate against cost to verify. Work that is cheap to generate and cheap to verify is a good automation candidate. Work that is cheap to generate but expensive to verify — thousands of landing page variants, or generated code whose underlying queries are opaque — is where organizations accumulate risk. Human judgment belongs on the expensive-to-verify side.
What is the first sign that an AI investment is working? Adoption, not outcomes. Willis has seen internally built tools go from 2–6% usage to around 80%, and only after that inflection did measurable outcomes appear. Outcome metrics are a lagging indicator; adoption is the leading one.
What does Chris Willis expect a year from now? Two shifts: management layers in which agents supervise other agents, with the agentic portion of the organization appearing on org charts alongside the human judgment layer; and governance being treated as a design problem rather than a technical one, embedded in the interface people actually interact with.omo.
About Chris Willis
Chris Willis is Chief Design Officer at Domo, where he brings more than 25 years of design and product leadership to the company’s data, analytics, and AI platform. Since joining Domo early in its history, he has played a key role in shaping the platform’s design strategy, helping make complex data more accessible and useful for customers across industries. Before Domo, Chris co-founded HOUR Detroit magazine and Footnote.com (now Fold3.com), which was acquired by Ancestry.com, and worked as an award-winning illustrator, journalist, and author. His experience blends design thinking, technology, and emerging trends to drive innovation and build tools that solve real business problems.,Yes,This will be completed shortly
Chris Willis on LinkedIn: https://www.linkedin.com/in/cwillis
———- Resources ———-
The Agile Brand podcast is brought to you by TEKsystems. Learn more here: https://aglbrnd.co/r/2868abd8085a9703
Reach your customers with Reddit. Spend $500 in ad spend, get $500 back in ad credit! Learn more: https://advertalize.com/r/491818c79fb1873f
Enjoyed the show? Tell us more at and give us a rating so others can find the show at: https://aglbrnd.co/r/faaed112fc9887f3
Connect with Greg on LinkedIn: https://www.linkedin.com/in/gregkihlstrom
Don’t miss a thing: get the latest episodes, sign up for our newsletter and more: https://aglbrnd.co/r/35ded3ccfb6716ba
Check out The Agile Brand Guide website with articles, insights, and Martechipedia, the wiki for marketing technology: https://www.agilebrandguide.com
The Agile Brand is produced by Missing Link—a Latina-owned strategy-driven, creatively fueled production co-op. From ideation to creation, they craft human connections through intelligent, engaging and informative content. https://www.missinglink.company
Transcript
[00:01:08] Greg Kihlström: Hi, I’m Greg Kihlström, host of The Agile Brand, and here’s a question for you. What if the real risk with AI isn’t moving too slowly, but instead moving fast on a foundation that hasn’t quite earned its trust? Because agility was never really about speed. It’s about being able to change again and again without the whole thing coming apart, and that depends on groundwork most organizations are tempted to skip. Today we’re going to dig into that groundwork, or the unglamorous part that decides whether AI actually pays off. Specifically, we’re gonna talk about why the same AI investment can compound trust in one company and quietly erode it in another, and what makes that difference. The tension between personalizing an interface for every individual user and keeping one coherent, recognizable product, and why building a real data culture is the step nearly everyone skips and no one can skip for long. To help me discuss this topic, I’d like to welcome back to the show Chris Willis, chief design officer and futurist at Domo. Chris, welcome to the show.
[00:02:52] Chris Willis: Oh, thanks. Thanks for having me back, Greg.
[00:02:54] Greg Kihlström: Yeah, yeah. Looking forward to it. Yeah, love, love talking with you before. Thought it would, um… Things change so quickly, I thought it would be great to, to have you back on. Uh, for those that didn’t catch you before, though, why don’t you give a little background on you and your role at, at Domo?
[00:03:06] Chris Willis: Sure. Uh, I’m the, uh, the founding chief design officer and futurist at Domo. Uh, so it’s kind of a unique title, which I’m, I’m very happy with. But my day-to-day job is I’m responsible for, um, the sort of strategic product vision and the experience layer of, uh, the Domo governed agentic platform. And, uh, so I’m looking at both very tactical things and things that are happening two or three years out, which you sort of alluded to is, is getting more challenging every day.
[00:03:35] Greg Kihlström: Yeah, yeah, exactly. And so for those, for those listening that may not be as familiar with Domo, um, can you tell us a little about, you know, what’s, what’s some of the core problems that, that you solve and, and who are the, who are the customers?
[00:03:49] Chris Willis: Sure. As, as a governed agentic data platform, the, the main problem we solve for a lot of customers is helping them gather all of their data that’s scattered everywhere else and create a more solid data foundation that they can create more, uh, trusted agentic systems, it, you know, dashboards, analytics, data products, et cetera. That’s, that’s still the everyday hard work [laughs] that companies-
[00:04:14] Greg Kihlström: Right
[00:04:14] Chris Willis: … um, really need and, uh, I think it turns out that a platform is a, is a great way to do that nowadays, especially with, um, you know, the s- the rise of the agents.
[00:04:23] Greg Kihlström: Yeah, yeah, definitely. Yeah. So let’s, um, let, let’s dive in and I wanna talk about… Well, we’re gonna touch on, on a few things here, and I think it’s a good follow-up to our, to our last conversation just, uh, a- as well. But I wanna start with this idea of, of trust and, and, and data, and I think not only on this show, but, uh, a lot of the conversations about trust involve brands and their customers and, and customer trust which, you know, definitely critical importance, but there’s a little less focus sometimes on how much people actually trust their own organization’s data that work within that, that organization. And so, you know, where does that internal trust
[00:05:08] Greg Kihlström: stand right now and, you know, how, how does this then lead to that, the customer, you know, the customer lack of trust as well?
[00:05:18] Chris Willis: So I’m, I’m glad you brought up that distinction, which is, you know, definitely if you’re looking at everything from a marketing perspective, trust is something that’s, you know, very core to what you do and what you believe, right? It’s about, it’s about forming a bond between your customers and, and being able to enable, you know, things to happen, which is great. I would say honestly, internally, if you’re talking about organizational trust or data trust, that seems to be on much shakier footing still, and I think that’s somewhat surprising for many organizations, [laughs] you know, who’ve spent lots of money and lots of effort, you know, creating data warehouses and data lakes and, [laughs] and lake houses, right? I mean, it just goes on and on and on.
[00:05:57] Greg Kihlström: Yeah.
[00:05:57] Chris Willis: And they’re like, “Wait a minute, aren’t, aren’t… Can’t we trust this data?” And, and that’s, I think, one of the, the big sort of weaknesses a lot of these agentic, um, [lip smacks]… projects are starting to reveal, which is the data fa- foundation might be a little shakier than you expected. And, and I think you’re starting to see that, you know, where are, are people keeping, like, kind of sh- their own shadow spreadsheets, [laughs] you know, their own numbers? Are they ar- arguing about numbers that should be, you know, um, sort of certified, val- validated and verified? You know, like that’s, that’s something we see day in and day out, um, because, and, and this is the key part of it, AI inherits whatever confidence you already have in your data.
[00:06:40] Greg Kihlström: Hmm.
[00:06:40] Chris Willis: Uh, I think the challenge, of course, is, um, that data may not be super confident, but it sounds really confident when you run it through some sort of agentic system.
[00:06:49] Greg Kihlström: Right.
[00:06:49] Chris Willis: And I think that’s also adding, uh, or amplifying some of the, the confusion out there.
[00:06:53] Greg Kihlström: Yeah, I mean, I think it’s… A- and there’s many things going on here, but you know, there’s, there’s enormous pressure to show returns on all of these invest- you know, ev- everything from the token costs to just the, the dollars and the time spent in, in all of this AI transformation. And yet, to your point, uh, these shadow spreadsheets probably always existed, but now there’s nothing like, uh, in, uh, in, in my work with, with enterprise orgs as a consultant, there, there’s nothing that shows some of these things quicker than trying to get AI, to your point, to interpret and, well, then it starts showing what’s really… either what’s really going on or where all
[00:07:39] Greg Kihlström: the disconnects are, and, and how, just how fragmented things are. So I mean, maybe it goes without saying, but, you know, what, what’s the real cost of, of deploying quick, you know, just quick, if speed is, is, is, um, you know, the, the paramount thing here. You know, what’s the real cost of deploying so quickly before you get the right foundation in place?
[00:08:01] Chris Willis: And, and maybe we’ll dive into this a little bit deeper. I’d love to dive in a little bit more on the trust issue. But I would say to-
[00:08:07] Greg Kihlström: Yeah
[00:08:07] Chris Willis: … answer your question directly, like, where is the cost? It’s in predictability.
[00:08:11] Greg Kihlström: Hmm.
[00:08:11] Greg Kihlström: Yeah.
[00:08:11] Chris Willis: You have machines that might be very confident sounding, but-
[00:08:16] Greg Kihlström: Yeah
[00:08:16] Chris Willis: … you know, it’s difficult to create something that’s predictable, and that’s often skipped, right?
[00:08:22] Greg Kihlström: Yeah.
[00:08:22] Chris Willis: I, I mean, you see this problem all the time where, um, we were just talking earlier before you started the show about, you know, sort of the, the flood of AI slop, right? That, that’s a form-
[00:08:31] Greg Kihlström: Yeah
[00:08:32] Chris Willis: … of it, where it’s just like we’re just throwing things at it and seeing what comes out, and it looks very confident. Um, but it, these machines, even though they have certain areas that are very powerful, there are certain places that they fail miserably, and we call this a jagged intelligence, right? So the models might work really well on one aspect, and that seems predictable, but something that’s just a little bit different, maybe a slight exception to what, you know, a, a model could do, it can’t correct and absorb your norms just by watching what’s going on.
[00:09:05] Greg Kihlström: Yeah.
[00:09:05] Chris Willis: And so I think, I think there’s a little bit of confusion and ambiguity around there. Plus, it’s a very different or difficult kind of problem space, ’cause you’re kind of staring into a black box. You know, so I’ll, I’ll give you, like, one quick example of, you know, I think a marketing example we’ve seen, uh, you know, for your audience would be you have a customer-facing agent, there’s some sort of promo, there’s a promo exception, but no one actually codified that.
[00:09:33] Chris Willis: And so the model will start to improvise. As opposed to, you know, maybe, maybe the, the, the manager or the customer service rep might already have some sort of carve-out idea for that particular promotion, and they can sort of confidently pr- apply that judgment to that customer. Models can’t do that. They kind of drift off from, you know, your goals. So I th- I, I think the lesson that is emerging from all of this is that undocumented judgment results in undefined behavior. [laughs]
[00:10:03] Chris Willis: And so, you know, I think a lot of it we’re, uh, starting to talk about in terms of context, right? That’s… I’ll use the air quotes because it’s such a broad, broad term.
[00:10:12] Chris Willis: But the debt is coming due at that first exception that nobody ever wrote down, that sat in somebody’s head. And I’m sure we’ll talk about this, but this is where kind of finding the, the, the right collaboration between what is repeatable and predictable with machines versus where is human judgment required, and that’s something different than, I would say, human in the loop, which I think is, is a somewhat tired term.
[00:10:41] Greg Kihlström: Yeah, and I think it’s, it’s a, I mean, I think it’s a, a bit of a paradox because we’re, you know, we’re, we’re asking humans to trust AI to do the work, and yet we’re asking them to be skeptical about the work, and we’re feeding, um, sometimes less than accurate information into the AI. You know, there, there’s a lot of, uh, may- and maybe this just goes back to judgment in, in the first place, but, you know, how do you, how do you build trust in data whe- knowing that the foundation might be shaky or it may not? [laughs] You know, like it’s… I know that’s a very open-ended question, but-
[00:11:20] Chris Willis: No, no. No, it’s, it’s a great question, and I, I am struck by how people are sort of approaching this, um, more and more nowadays. Uh, trust was something that was kind of at the periphery. People were just very excited about the tools.
[00:11:36] Chris Willis: But I was at a few design roundtables and working with some customers. I was out on the East Coast, out in New York doing some stuff, and this was one of the number one conversations. They were like, “How do we design or engineer for trust?”
[00:11:49] Greg Kihlström: Hmm.
[00:11:49] Chris Willis: And I, I don’t like to, you know, like throw cold water on ideas ’cause trust is important. But trust is not something you can engineer for easily. It doesn’t exist.
[00:12:03] Greg Kihlström: Yeah.
[00:12:03] Chris Willis: So, so my approach, and the approach we take is, at, at Domo, is- We engineer for three things. Visibility, predictability, like we just talked about, and control.
[00:12:17] Chris Willis: Bring those things together. So I’ll j- just touch on those really quickly. Visibility is can people see what the agent actually did, right? So it’s not something that’s sort of, you know, buried in the log somewhere, but you can actually see it. And I think, I think there’s more, you know, as a designer, I think people would say, “Of course you would say this,” but I, I think there’s more importance of sort of surfacing these things at the interface level, uh, because trust and visibility… Well, trust fed by visibility, predictability, and control are gonna be more important to see as agents are acting.
[00:12:50] Chris Willis: So the visibility is can I see what it actually did? Predictability is same situation, same behavior, right? [chuckles] That’s deterministic.
[00:12:58] Chris Willis: Somewhat of a challenge, right? And then you try to measure, like, how that variance happens in, in identical use cases. And then the last part, of course, is control. Can we, if we see something going wrong, can we stop it? Can we adjust it? Um, can we, over time, make it so that control is less and less of an issue? So I, I think that’s the way you start building trusted systems is by engineering for visibility, predictability, and control.
[00:15:10] Greg Kihlström: Not only does trust need to be earned, but things are moving so quickly, uh, that, I mean, trust kind of has to be proven consistently. So in other words, kind of what you’re ta- the visibility, predictability, control, that’s a, at least in my mind, that’s a way to prove that something can be trusted as opposed to… ‘Cause w- if we just blind- you know, if it works three times, and then we’re like, “Cool, um, just [chuckles] whatever, whatever you say, let’s do it.” I mean, not that any enterprise org is gonna do that, but, eh, maybe it’d be surprised-
[00:15:44] Chris Willis: I was gonna say… Sorry, I wanna… I just sort of building that. I, I’ve actually, I’ve run into that with big, big organizations where they actually did do that, where they thought, “Hey, we’ve got some great agents. They seem very powerful. Let’s apply it to a certain kind of use case.”
[00:15:58] Chris Willis: And then they go test it, and to their surprise, it works perfectly the first time.
[00:16:03] Greg Kihlström: [chuckles] Yeah. Yeah.
[00:16:04] Chris Willis: It’s the second time. Because in many ways it is kind of like rolling the dice.
[00:16:08] Chris Willis: But, um, yeah, no, you’re right, and I think, again, this kind of points back to we sometimes call this a verification problem, which is many times… And, and I apologize, I could o- [chuckles] I, I hate to have to sort of mime this, but just imagine a simple chart with two axes.
[00:16:25] Chris Willis: Um, on your X axis, think of it as effort to generate something. Could be an answer, could be a landing page, could be, you know, a promotion, whatever. On the vertical axis, think of that as your cost to verify.
[00:16:41] Chris Willis: Is that thing right? It does… Is it free of errors? And I kind of think of an imaginary diagonal line, and I try to get problems below that line. So, you know, like, one of the challenges is, like you were sort of alluding to, is let’s say you’re, you know, you’re kind of generating a bunch of stuff for landing pages, and back in the day, there was sort of a natural constraint. Um, people had to write copy. That copy got checked.
[00:17:08] Greg Kihlström: Yeah.
[00:17:08] Chris Willis: It went through a few different layers of checks and balances and, and but people were kind of in the loop on all of that. Well, now you might be using agents to create thousands of new kinds of landing page copies, or it might be the same one but slightly different. That’s very, very difficult and expensive to verify. Um, vibe coding is a great example of relatively easy to generate. You know? I mean, yeah, it costs you some tokens, but what is in that code? What queries were made, SQL queries were made to that data? So when you said revenue, okay, this time maybe it got the right revenue. Maybe the SQL’s a little different on the next one. You know, there’s a different where clause or something. So that’s, that’s a way of kind of thinking about the world of problems. And, and what you wanna do is you wanna think about,
[00:17:54] Chris Willis: all right, let’s automate things that can be verified easily, and then things that require, you know, are, are slower and more costly, those are where human judgment come in.
[00:18:04] Chris Willis: Or you have to kind of figure out how to take pieces of that and make those verifiable.
[00:18:08] Greg Kihlström: Yeah, yeah, ’cause, I mean, otherwise you’re- I, I don’t know if you’ve s- saw the cartoon about, it’s like how to build a car, you know, manually versus with AI. It’s like the, the manual one is like you start with wheels and build the frame and whatever, and, and then, you know, slowly it becomes a car. But to do it with Gen AI, it’s like it starts out as some like crazy monstrosity machine, and you like take all the stuff away that’s not a car. You know, I’m just reminded of, of things like that when I’m vibe coding or doing things, which is, it just… To me, that’s the verification process or just the, the shaping it into the thing you want. It’s like either it starts from zero, like hu- with human effort built, you know, bit by bit, or AI
[00:18:54] Greg Kihlström: creates something, and then to your point, you’ve got to whittle it down into what you actually wanted in the first place, right?
[00:19:00] Chris Willis: I, I, I think the tools we have right now, uh, you know, they might seem impressive, but they’re, you know, relatively primitive, and they’re-
[00:19:07] Chris Willis: … kind of built backwards in that regard. So yeah, to your point, um, a lot of models, they just give you a magic answer.
[00:19:16] Chris Willis: Um, as opposed to sort of providing the natural friction that we’ve always had to deal with in the creative process.
[00:19:22] Chris Willis: You know, think about it. Think about it deeply. Try different things. Make sure you can explain it. Can you get someone else to get the idea, right? These are all things that refine the idea before you actually ever even do it, and AI is, you know, here’s a magic answer, and then you spend most of your time, in a weird way, uh, just kind of making that next prompt hoping it’s all gonna figure it out. You know? [laughs] And it’s in a way it feels kind of like gambling. [laughs]: … yeah, that can be, that can be somewhat confusing and exhausting.
[00:19:55] Greg Kihlström: So let- let’s talk about then, you know, how, how do we move forward here, you know, and even starting at, at adoption and governance and, and things like that is just… You know, how do we, how do we know that we’re on the right path, but also how do we put those guardrails in place be- You know, because again, you mentioned the having the tools like visibility and predictability are, are great, but how do we, how do we put a process in place to enable teams to really do that? And then, then I feel like we move forward if, if we’re able to do that, right? But, but how do, how do you start with that?
[00:20:41] Chris Willis: And I know because we do a lot of that, y- you’d say, “Of course you would say that.” But-
[00:20:45] Greg Kihlström: [laughs]
[00:20:45] Chris Willis: … I’ve just seen it everywhere, right? And if you don’t have that, nothing else matters. So because you could be just using a- agents to come up with very fast, confident sounding, but wrong answers, or it could be w- even worse. It could be introducing new kinds of security issues or new kinds of [laughs] brand crisises or like a crisis, right?
[00:21:08] Greg Kihlström: Right. Right.
[00:21:08] Chris Willis: You know, it’s like, oh wait, it, it just sent out, you know, either, uh, uh, responded via email to a bunch of people, and it shouldn’t have said those things, or it gave people a huge discount, it shouldn’t have done those. Um, so I know it’s not very sexy, but that, that strong data foundation is really critical. I also think it’s important to kind of start in a very kind of focused and limited way.
[00:21:31] Chris Willis: Uh, because we’re all still learning, right?
[00:21:33] Chris Willis: And so try to figure out applications that don’t necessarily have to do everything all at once. I think there’s a big interest in like these sort of moonshot innovation projects.
[00:21:46] Chris Willis: And two things kind of happen there. One, those are very risky and costly. Uh, but two, uh, they’re, they’re not necessarily things that move the needle as quickly, right? So one of the things we’ve seen is people seem to skip a step. They skip a, a few steps in different ways. But in this particular regard, I think it’s important to make a distinction between, hey, we have an agent or agents that are doing things, and we are really focusing on the outcomes. That’s great, but that tends to be a slightly lagging indicator.
[00:22:19] Chris Willis: Internally what we’ve seen with like really smart marketing teams, for example, is they focus first on adoption. So it… ‘Cause everyone’s sort of being swimming around in a, in a pool of all these new sort of tech tools.
[00:22:32] Chris Willis: But it only takes really like one really special person to kind of figure out one part of the problem, create a solution, and then start seeing adoption. So for example, what have we seen that’s failed? Sometimes you’ll, you’ll see like, uh, you know, like marketing tools, which are essentially just sort of glorified chatbots. Those usually don’t get great adoption. But what we’ve seen is, um, teams that have tried to solve a problem that up to this point was kind of difficult to solve. It wasn’t difficult to describe, right? It wasn’t a super complicated problem. But they started building their own tools, and those tools went from 2%, 5%, 6%-
[00:23:11] Chris Willis: … to like 80% adoption. That became, and I’ve seen it across other organizations, the leading indicator. So once you start getting that signal, usually the outcomes start to happen, whether it’s cost savings or, you know, vendor displacement of certain tools, um, or, you know, customer satisfaction or things like that. So that’s the way that I, I’ve seen the best organizations kind of navigate this time, and I think this will change, right? So tho- those things will change. Those, um, uh, sort of approaches will change, but the foundation has to be there.
[00:23:45] Chris Willis: Right? There’s certain things that are non-negotiable.
[00:23:47] Chris Willis: And those things aren’t very sexy, but every great company is going to have those things.
[00:23:51] Greg Kihlström: Yeah. Well, and I think, you know, to, to that point, I mean, I think leaders need to understand that that takes a little bit of patience, but the, to your point, the, the payoff is, is much great. It’s, it’s actually probably much quicker in the long run, it just starts slow. Is that kind of what you’re saying?
[00:24:11] Chris Willis: Yeah. Um- Well, you know, it’s interesting. Does it start slow? I guess in this moment everything feels slow, right?
[00:24:20] Greg Kihlström: Yeah. Right, right.
[00:24:20] Chris Willis: Because it feels like you’re so far behind. So you’ve already started behind, so even before you begin you feel like it’s not fast enough.
[00:24:27] Chris Willis: Um, but yes, I think the expectation should be it’s going to take a while. Um, the, the example I gave you, um, I’ve seen in some marketing organizations, it took about a year for that to be figured out. Um, but they were also not doing just one thing. They were trying lots of little experiments.
[00:24:45] Chris Willis: They were making bets, right? That’s, that’s a very-
[00:24:48] Chris Willis: … common way of thinking about it. It’s like, “Yeah, we’re gonna take a little bit.” But it actually paid off where, you know, they, they saw hundreds of hours of time savings, they saw, um, increase in satisfaction, they saw increase in pipeline. However, what they saw first was adoption.
[00:25:07] Chris Willis: People started using the tools, and maybe they couldn’t articulate why they were using it, but something was starting to work, and that was the first leading indicator, and that’s the kind of thing you want. And then you can k- sort of verify, you know, sort of outcome based-
[00:25:20] Chris Willis: … um, uh, outcomes. Yeah.
[00:25:23] Greg Kihlström: Yeah. Love it, love it. Well, Chris, thanks so much for, for joining again. Um, I could definitely talk about this way longer here-
[00:25:29] Chris Willis: [laughs]
[00:25:29] Greg Kihlström: … but we’ve gotta, gotta, gotta wrap here. So, um, but, uh, c- couple of quick things as, as we, as we wrap up here. Uh, first thing, if we were having this interview one year from today, what is one thing that we would definitely be talking about?
[00:25:45] Chris Willis: I, I think… And this is looking out a bit. I mean, one year seems like a pretty long, a long horizon.
[00:25:50] Chris Willis: But I, I think once we start getting those other pieces in place, the data foundation, you know, the guardrails, the deployments, I think you’re gonna start to see agents starting to manage their own agents. Right now you have people just trying a bunch of things, and agents are running all over the place, and it… that’s hard to, to, um, to judge. But I think what you’re gonna start to see is kind of these management layers. Like, in a weird way, I, I think you’re gonna kind of see, like, the human part of the org that’s doing judgment-type things-
[00:26:16] Chris Willis: … [laughs] and then you’re gonna see the agentic part of the org. You’re already starting to see that show up on org charts where people are like, “Oh yeah, here’s our agents.” [laughs] You know? Like this agent reports to that agent. But yeah, I think… I, I, I could see that happening. I’m starting to see organizations create fine-tuned models that do certain things. They really understand parts of the business or processes, and I don’t see why they wouldn’t start to sort of combine those to allow people to do more higher order things. I think that’s gonna be one. Um, I also think governance is gonna be seen less as a technical thing and more, more of a design problem.
[00:26:48] Chris Willis: You know, companies realizing that governance has to kind of be, you know, the interface in their policy, right? It has to be something you, you interact with in maybe a, a more direct way than you have in the past. ‘Cause without that, it’s just gonna be chaos.
[00:27:01] Greg Kihlström: Yeah. Yeah. Love it. And last question for you. Uh, what do you do to stay agile in your role, and how do you find a way to do it consistently?
[00:27:10] Chris Willis: Uh, I read a lot, and I just try a lot of things. [laughs] I know that sounds, uh, also not super sexy, but I feel like when there’s a new technological disruption, the best way to engage with it is deeply. Um, and, and it’s, I think, especially true for agents. Um, you know, if, if you’re not deeply engaging, trying things, um, experimenting, um, questioning what’s happening, and also kind of in a way understanding what your relationship should be with that technology, um, you’re just gonna be pulled in by the hype, and you’re gonna be really confused and frustrated, and potentially, it… you might get pulled into some areas that aren’t, aren’t super helpful. So I think of these as tools, but they’re tools that don’t have a manual to them. [laughs] They’re tools that you can’t just open up and see what they do. Uh, so the, the best way to, uh, approach that, I think, is to engage deeply and, and just try lots of things. Same thing we talk to, uh, organizations about.
