you2idea@video:~$ watch p3vg3TZ-XE0 [16:55]
// transcript — 218 segments
0:03 Let's welcome up our panel, shall we? >> it. All right, put your hands together.
0:06 >> welcome panel. Go ahead, Stephen. All right, we have Joe Dormani, partner at
0:11 Thomson Reuters Ventures. Please give Harshita Mirchandani, principal at Avesta Fund, and
0:22 we have Ryan Taylor, principal at Fusion Fund. Put your hands together. Thank
0:25 you. >> All right, so thanks for joining us, and thanks for being here. I'm excited to
0:33 talk about this. We uh we're talking about autonomous advanced AI systems,
0:37 and so a lot of what this is translated is from what can be done to what
0:42 actually is reliable to is this a real business, right? And so uh Ryan, I want
0:46 to start with you. What how do you separate sort of uh the the signals from
0:51 all the noise right now out there? Yeah, that's Yeah, that's a really good
0:55 question. Um I think what we're looking So, just to give some context here, we
0:59 invest at the earliest stage, uh pre-seed to series A. Often, a lot of
1:03 the companies that we're speaking with um don't have traction yet. Um so, what
1:09 we're really looking for is really tier one research, um really strong pedigree
1:14 both from an experience perspective, because fundamentally we're betting on
1:17 the team. And then and at the very earliest side, um strong POC or
1:23 development partners um that actually really understand the pain point that
1:27 this company's trying to solve, and can really help advocate on behalf of the
1:31 company and say, "Hey, no, this is real. Um this problem's scalable. It exists
1:36 throughout our entire industry, and we see real line of sight to this not only
1:39 scaling throughout the broader industry within our org to ideally millions in
1:44 contract value." So, that's really what we're looking for to kind of help
1:48 identify who's doing something real, and who's doing something that might be a
1:54 not as tangible. Let's just put it that way. Yeah, that's That's uh that's helpful. Thank you for
2:00 that. And uh Joe from uh workflow, from a work uh flow system standpoint, what
2:04 are you seeing that's that's working right now that's being adopted?
2:08 Uh yeah, it's it's a You're on. You're on. Okay. Yeah, so it's a great question. Um so, TR
2:15 Ventures, we invest in uh Series A companies in professional services. So,
2:19 a lot in legal, tax and accounting, risk, fraud, and compliance. Uh so, we
2:23 get a pretty broad spectrum of different types of AI companies that are being
2:27 built across those different areas. And so, you know, what we've seen work really
2:32 well is uh you know, a vertical specialization. Um so, areas within um
2:39 areas of law or areas of tax and accounting or from a risk standpoint.
2:44 So, you know, a focus on personal injury, for example, or focus on
2:49 intellectual property, or focus on uh transactions. And and that focus and
2:56 specialization enables companies um generate AI companies to generate a
3:00 ton of value because they're able to focus on very specific problems. They're
3:06 able to create very specific uh outcomes or outputs or documents that
3:12 are, you know, critical for moving those types of workflows forward. And so, we
3:17 see a lot of traction in this kind of vertical or subcategory specificity. Uh
3:22 and and just basically what that enables is a a really fast time to value for
3:26 these companies so they can prove their ROI to their customers very quickly.
3:31 That makes a lot of sense, Harshit uh Harshit, I'd love for you to to chime in
3:34 as well. What what are you seeing that helps you separate signal from noise?
3:39 And and to tag on to that, there's a lot of great demos that are out there,
3:42 right? And how can you tell something's real versus just another great demo?
3:49 Great question. Uh that's something that we're still learning as we go. So, for
3:53 some context, we're a Colorado-based early-stage VC, and we invest in
3:58 physical AI, especially as it relates to really high-energy industries. So,
4:04 things like manufacturing, energy, um the grid. And for me, there were several of the points
4:10 that my panelists said that I I fully concur with. But in addition to that,
4:16 I'd say number one is we see a lot of great demos that say, "Hey, we're 99%
4:24 accurate." My question beyond that is, you know, do you understand the edge
4:28 cases? Do you understand what that long tail of problems is? Because
4:33 um you know, in a steel factory, you need to be explainable, even when you're not
4:39 perfectly accurate, and we understand that's normal for early-stage companies.
4:43 Second, we hear a lot about data flywheels, and it's kind of a a hand-wave gesture to
4:49 say, "Hey, we're going to build it at scale." The problem with industrial settings and
4:56 with just manufacturing as a whole is that a lot of deployments on prem, uh
5:01 it's really hard to generalize. So, that question on how are you collecting
5:06 proprietary data? I mean, not just do you have it, but how are you collecting
5:10 it? Why is it hard for somebody else to collect the same thing? Is it just, you
5:16 know, and that needs to be a living organism. Um proprietary data becomes redundant if
5:22 you don't constantly keep updating it. So, having a way to really build that
5:26 data set, to collect on-site data that sometimes your customers may not even
5:30 know that they have, is extremely valuable. So, those are the Those are a
5:34 few of the points that I'd want to touch on. >> Joyce, uh um I see you're over there,
5:39 Ryan, ready to go. Yeah, and to add And to add on to that, I just want to
5:43 highlight that yes, we need to be able to, you know, understand these edge
5:45 cases. But a lot of companies, they really emphasize on like a full
5:49 end-to-end autonomy stack. And unless you really had the foresight, like
5:53 Tesla, to acquire this data over 10-year periods, your data set's very robust and
5:57 very rich, the chances that you're going to get 100% autonomy in a lot of these
6:02 vertical use cases is very low without a lot of upfront capital, a lot of upfront
6:07 data collection. It's hard. And so, some of the companies that we're seeing get
6:11 to value the quickest right off the bat is they they automate as much as they
6:15 can. And for these edge cases, they have human in the loop, whether that's like
6:18 remote control in some way, case, or form. And then they're actually very
6:22 quickly able to identify, "Hey, where do we not have good data? Where is it
6:26 failing? Where do we need to actually, you know, focus on?" And so, when the
6:30 humans operate and they loop in, then that's really where they focus.
6:33 They really focus on that data specifically and then retrain on that.
6:37 And that's where you can actually deploy at the earlier stage. We even see
6:40 companies generating revenue in the seed series A with this framework. Um yes,
6:45 it's not full autonomy. Yes, it's a little bit harder to scale, um
6:49 but it works. No, I love that answer. Thank you for that. And and Joe, tell me what when we
6:54 think about these pilots, cuz there's a lot of you know, obviously there's a lot
6:56 of pilot programs that are out there. What are you seeing? Tell the founders
6:59 out here. Why why are things Why do they fail? Um most common reasons.
7:04 Yeah, I guess I guess, you know, there are a couple failure modes that come to
7:08 mind. I guess, you know, when when you don't have the right level of engagement
7:13 from a customer executive sponsor point of view, you know, that's one kind of
7:17 failure mode. Um another is when you don't have a very clear
7:22 um ability to deliver uh real and tangible ROI. Um and, you know, another one is just when,
7:31 you know, there's a lot of experimentation going on with with with
7:35 companies and customers. And so, maybe if you're one of many
7:39 um doing pilots at the same time, you know, I think this is an acceptable
7:43 situation where you have to prove value, but it's definitely more of a risk
7:49 when you know their companies has a budget that may not be in place for the long term. It's more so
7:58 to do experimentation and learn about what options exist in the market.
8:02 So I mean that's kind of the reality that you'll have to live with, but I
8:05 think you know, you you really want to make sure that there are dollars
8:09 allocated, that those dollars aren't going to go away, that the use case that
8:14 you're focused on is a is a real pain point for the enterprise, and that you
8:19 have the right executive or share stakeholder engagement to convert that POC to a longer term
8:26 relationship. That's helpful and Harshit, you had mentioned something pre-event I thought
8:32 was great. You had the term I think pilot purgatory. So so tell us a little
8:36 bit about what pilot purgatory is, how you get stuck in it, and how the heck
8:39 you get out. So pilot purgatory is when right after you get an initial pilot and complete
8:48 it, there's you're in a period where you don't really know what the next step is.
8:51 You're almost in a purgatory. You don't really know whether the enterprise that
8:57 you piloted with is willing to actually you know, scale your technology, do you
9:01 go to phase two, and there's there's a lot of just in between that you're stuck
9:04 in before it turns into a full company-wide deployment with your
9:13 customer. And it's a really tough spot. Plenty of companies get stuck in it.
9:17 Think it's driven by a couple of things. Number one, you touched on this, but
9:21 just not connecting with the right buyer. And by right buyer, I mean the
9:26 person actually signing the check. It needs you know, whatever they're
9:30 buying needs to align with the customer's KPIs. You are not going to
9:35 make a large enterprise change their KPIs, and the person who's buying from
9:40 you needs to be incentivized to push it forward, whether that's in the form of
9:44 hey, I'll get a promotion if I do this, or um this is very closely tied to our
9:48 priorities. That's number one. Number two, the folks actually using your technology
9:55 need to find deep value in it. What as an example, we see a lot of sort of industrial tech
10:04 where the technicians in question um are unwilling to use it for a variety
10:08 of reasons. It could be because they find it too hard to learn, the
10:12 onboarding process is cumbersome, um they think it'll replace them as
10:16 opposed to augment them, and there's there's a lot of just, you know, end
10:20 user understanding that needs to be built in. Um three, I think KPIs and scaling pathways need
10:28 to be talked about from day one. I think sometimes companies in an attempt to get
10:33 in with an enterprise uh do everything they can to get an unpaid pilot or just
10:37 to just get something going for that logo, but the odds of it really converting aren't
10:43 that high if there isn't a willingness to pay for it. Um I think it's a great
10:48 signal when a company's willing to pay for for a technology, and they're also
10:53 just willing to also try it with live data. So, you know, there's a lot of
10:57 companies that will put you in a sandbox and try with historical data. If they
11:02 can get you on a production line, even just one, I think that's that's a that's
11:05 a great signal. So, to kind of summarize a bit, you know, executive sponsorship for sure
11:09 super important. We're talking to the right people, you have the right people
11:13 involved, they're invested. The KPIs, right? So, understanding from the start
11:16 what the expectations are and making sure there's an incentive there too to
11:20 to drive that. I love that. But, it feels like we're kind of having a moment
11:23 right now in this specific category with advanced and autonomous systems. And and
11:27 Ryan, I know you have some some thoughts on this, sort of the the economics of
11:30 what's happening right now, what's driving the growth. We'd love to hear
11:32 from you. Well, in addition to the economics, it's also technology. So, I think there's two
11:39 core variables, right? So, we we saw that the scaling law actually applies to
11:43 physical AI, which is really exciting. You know, you take a lot of compute uh
11:47 you take a transformer architecture, you throw a lot of data at it, and like you
11:51 get generalizable, like pretty high performance. The main issue with
11:54 physical AI or robotics right now is just we just have a huge data scarcity.
11:58 And so, a lot of people are spending a lot of time trying to collect real-world
12:01 data. You have a lot of people, you know, building great simulation models
12:04 to be able to generate scalable synthetic data that's as accurate as
12:08 possible. The other thing that's really driving adoption, though, outside of
12:12 these technological revolution is just cost of labor in the United States is
12:15 getting more expensive. Specifically, cost of of physical labor. Um not only
12:19 do you have a trend of like people not wanting to do it, but you also have, you
12:23 know, the southern border being closed as an example, right? That's driving up the
12:27 cost of labor. You see a lot of people I mean, you talk to even restaurant owners
12:31 in this area, they can't get anyone for less than 30, 35 dollars an hour with
12:36 tax and benefits included. That's That's very pricey. Um then you also have
12:40 tariffs, you know, that's getting rid of international labor arbitrages. All of
12:44 this is driving up cost of physical labor, which then increases the
12:48 incentive for physical AI. And what we're seeing, um at least at the early
12:53 stages with our portfolio companies in this space, is that the adoption as a
12:57 result is a lot faster. So, you have both the technology trend and then this
13:00 macroeconomic trend, which is really pushing adoption, which, you know,
13:07 That makes a lot of sense, and thanks for sharing that. Um so, we have just a
13:10 few minutes left. I want to make sure that we address the founders in the
13:14 room. And for each of you to tell them when you're meeting somebody early
13:18 stage, uh Joe, we'll start with you, what what uh signals to you to lean in a
13:22 bit more than than perhaps losing interest? What are you What are you
13:25 listening for? Yeah, look, I I mean, there's a number of different types of parameters that we
13:29 look for, but you know, first is, you know, team quality, founder market fit.
13:34 So, is the team well suited to solve the problem that they're going after? Is
13:37 there anything about their prior experience, whether that be academic or
13:42 professional, that gives them an edge? Um and then, you know, we're big on like
13:46 do what you say you're going to do, tell us that you did it, and then remind us
13:50 that you did it. So, keep on telling us, here's the plan for the next 6 months,
13:54 and this can be you know, traction type of met you know, metrics, product
13:57 development metrics, but you know, here's what our plan is, and then we're
14:00 executing on that, and then this is what we've executed on. Um obviously, we we
14:05 care a lot about unit economics. We want to see scalability in in what it costs
14:10 to acquire a customer, the value create and over the lifetime of that customer,
14:15 does it justify that spend? Um and so, you know, that type of um
14:20 scalability is important for us. Um and then, you know, we we we look through a
14:23 number of different lenses on the product side as well to make sure that
14:27 there's defensibility or or that you're able to create distance between you and
14:31 other competitive products in the market. So, th- those are a couple of
14:35 lenses. We can we can go to detail about any of those, but you know, that's kind
14:38 of high-level. Appreciate that. That's helpful. And Harshita, over to you.
14:45 Oh. In addition to all of those things, I'd say uh team productivity,
14:50 understandable that, you know, you may not be generating revenue, but how
14:54 productive has your team been since it's launched? Um go-to-market, have you been able to
15:01 shorten sales cycles beyond the status quo? Have you been able to think about,
15:05 you know, how are you going to actually scale? Uh direct-to-customers really hard.
15:10 Um and then, in terms of unit economics, I sort of the converse of that is also how are you
15:19 thinking about your cash burn and cash management? Very often, companies
15:23 uh run out of money and we're not necessarily in a period where
15:29 venture funding is unlimited. So just being mindful of cash management
15:34 something that's to me a very positive signal. Ryan you finish us off with this one.
15:42 Just one final item on that and it's really just the market that you're going
15:45 after. I think there's a lot of focus on automating a lot of low
15:53 low cost labor not a lot of focus on automating like high value labor. Like
15:56 one of the companies that are actually both looking at quite a bit right now is
16:00 actually trying to automate as much labor in clean room environments as
16:04 possible. Those are PhD level people doing very manual repetitive tasks that
16:11 they don't want to do. So the ROI on that is massive and so if you can find
16:15 unique niches that haven't been competed in yet that is really producing a ROI and you
16:24 can make a a lot of noise with that. Thank you guys for that. I mean we heard
16:27 communication and talking you know not only calling your shot but sort of along
16:31 the line informing and communicating how you're performing and team performance
16:35 as a whole right the team itself making sure you've got enough cash obviously
16:38 important cuz not a lot happens if you don't right and then focusing on those
16:42 areas that others might not be the niche the small the the higher end labor I
16:46 think all really great points but thank you all for your your your time today
16:49 and for for adjusting the audience big round of applause for our panel.
$

AI, Autonomous + Advanced Systems VC Panel

[fundraising and investment][AI agents and automation]
// description

Ryan Taylor (Fusion Fund), Harshita Mira Venkatesh (Avesta Fund), Joe Dormani (Thomson Reuters Ventures) + Stephen Katrivanos (Lenovo) This was filmed at our annual flagship Conference in Silicon Valley. Check out upcoming events near you and get involved with your local chapter at https://www.startupgrind.com. ---- Startup Grind Conference - returning April 2027. More at https://www.startupgrind.com/conference/ LINKEDIN: / startup-grind X: https://x.com/StartupGrind FACEBOOK: / start

now: 0:00
// tags
[fundraising and investment][AI agents and automation]