ReflexAI on Why Most Enterprise AI Projects Fail in Production | AppDevANGLE

0:07>> Our industry research shows that 80%

0:10or more of AI projects fail

0:12to reach meaningful production deployment

0:15and development delivering measurable business value.

0:18This is highlighting the operational

0:19and organizational challenges in scaling beyond prototype.

0:22My name is Paul Nashawaty, I'm the practice lead

0:24and principal analyst for the AppDev practice,

0:27and I'm joined by John.

0:28John, how are you doing today? You're from

0:29ReflexAI, correct?

0:31>> Doing well. Thank you. Thank you for having me.

0:34>> Yeah. John, tell me a little bit about yourself

0:36and about ReflexAI.

0:37>> Yeah, well, thanks for having me on. I am John Callery.

0:42I am the co-founder and chief product

0:45and technology officer at ReflexAI,

0:48and we build AI-powered simulations

0:50and quality assurance software

0:52that improves human performance in complex conversations.

0:57The company grew out of work my co-founder, Sam Dorison,

1:01and I did when we were both leaders at The Trevor Project,

1:04and there we built AI-powered conversation simulations

1:08to train crisis counselors at scale,

1:11and that work ended up getting major recognition,

1:14including Times 100 Best Inventions

1:17of 2021.

1:19One through line from

1:21that experience is AI can accelerate experimentation

1:25dramatically, but the systems you rely on in the real world

1:28still require strong engineering discipline

1:32and operational rigor,

1:33and I think that's really the lens that I bring

1:36to this conversation about vibe coding

1:38and scaling AI in enterprises.

1:41>> Well, John, it's super relevant.

1:42I mean, it makes a lot of sense.

1:44We see a lot of organizations moving from vibe coding

1:47and AI experimentation really to production

1:49to scale deployment.

1:51But this is a challenge that

1:53no longer can just model access,

1:54it really requires operational discipline.

1:56Because when we're looking at things like governance,

2:00compliance, regulations,

2:01and control, enterprises really must have an established new

2:05engineering standard that manage AI-native tech debt,

2:08ensure governance, but also prevent uncontrolled complexity

2:12really from compounding across code, data,

2:15and automation layers.

2:16This is really kind of an extension of what we were seeing

2:20with shadow IT and shadow development,

2:24but this is now becoming shadow

2:26code and application development.

2:28At the same time,

2:29we're seeing innovation can no longer be confined

2:32to these technical teams

2:34because they just don't have the amount of

2:37resources in order to do it for professional developers.

2:40So we have to let citizen developers in product

2:43and operations and finance

2:44and go-to-market functions really help

2:46build products of their own.

2:47So, John, what are your thoughts on that?

2:48I mean, it really is kind of an area

2:51that's expanding from the professional

2:52developer to the citizen developer.

2:55>> I think democratizing AI is essential

3:00because a lot

3:01of the best AI opportunities come from non-technical teams.

3:05If you think product, design, operations,

3:08and go-to-market, those are folks who are often closest

3:11to real workflow problems.

3:14The risk is exactly what you said.

3:16If people are experimenting independently,

3:20you can get kind of fragmentation

3:22and inconsistent standards in shadow AI,

3:25but the solution isn't to stop experimentation.

3:29The solution is to make experimentation safe and structured.

3:34One of the reasons I feel really strongly about this is

3:37we're living it right now.

3:38This week, we intentionally changed how we work

3:42so every role in the company builds real AI fluency as part

3:46of their daily workflow, not just as a side hobby,

3:50and the point of this is to remove bottlenecks.

3:52If only engineers can use AI effectively,

3:56then product context, customer nuance,

3:59and operational insight don't make it into the system,

4:02you end up with output that's fast

4:05but lacks user-centered soul

4:07that it actually needs to succeed.

4:09Practically speaking, the organizations

4:13that do this well provide shared platforms, clear policies

4:17and support, and they also create a culture where

4:21people can share what worked and what didn't without fear.

4:25I just want to be clear, this is not about replacing people,

4:27it's about giving people leverage

4:30and raising the bar of quality across the organization.

4:35>> Yeah, absolutely.

4:37I think that there's this misnomer of AI is going

4:41to come in and take a new job.

4:43I don't think that's really the case.

4:45I mean, it's an operational tool.

4:47Because you can do math on pen and paper,

4:50but you can use a calculator.

4:51It's kind of a similar approach here.

4:53But let's talk about something you talked about,

4:54which is bridging the vibe coding to production reality.

4:58The gap isn't about model quality,

5:00it's about operational rigor.

5:02When we look at security, observability, versioning,

5:04governance, it really comes down

5:06to prototype success often masked with missing fundamentals.

5:11You think things like data quality or latency constraints

5:14or cost controls or compliance.

5:16Those are things that organizations must

5:18understand and know.

5:20And then when you have AI system

5:21that have probabilistic kind of production environments, it

5:27really demands guardrails, right?

5:29Because if you don't have those guardrails in place,

5:31then it's the kind of a little bit of chaos.

5:33And I get it that when...

5:36You kind of want to use the words of former Andy Grove from

5:41Intel who said like, "Let the chaos...

5:43and then rein it in." I guess that's true in the context

5:46of moving production requires evaluation frameworks,

5:51it requires clear ownership,

5:52it also requires production grade data pipelines.

5:55So, John, when I'm looking at this, we're seeing a lot

5:57of vibe coding and AI experimentation,

6:00but what fundamentally breaks when organizations try

6:04to move these prototypes into production?

6:07>> Vibe coding is incredibly powerful for exploration.

6:11You can iterate with an LLM quickly,

6:15you can test ideas,

6:16and get to early signal way faster

6:20than traditional development.

6:21The problem is that prototypes hide operational complexity.

6:27A demo can look great while skipping all of the fundamentals

6:31that only surface later.

6:32So once you go to production,

6:35the requirements completely change.

6:37Now, you need reliable data flows,

6:41like you said, automated testing, observability,

6:44and security controls,

6:46and really clear ownership over

6:48how the system actually behaves.

6:51Like you mentioned, another piece is

6:53that the AI systems are probabilistic.

6:55Traditional software is mostly deterministic,

6:58but AI behavior can shift based on everything from context

7:03to model changes to just subtle differences

7:05and input, so production environments really need those

7:08guardrails and evaluation.

7:10It's more than just a working demo.

7:13I think even here at ReflexAI,

7:15we've seen this firsthand in high stakes environments.

7:18When the work also affects real human outcomes,

7:22you can't just rely on something that merely seems to work.

7:24You really need evaluation frameworks and monitoring

7:28and disciplined release practices.

7:31There's also a hype dynamic right now

7:33where people are talking about shipping features at

7:36incredible speed.

7:38Even the frontier AI labs talk about the

7:43tools that they've built and shipped in weeks.

7:44But if you've

7:49actually used a lot of these AI apps

7:51or looked at their status pages over time, it's a reminder

7:55that operating AI systems reliably is still very hard.

7:59So if the companies

8:00who are building the models themselves are constantly

8:04running into stability

8:06and performance of even their own apps, it

8:11just shows you how big the gap really is

8:13between rapid prototyping and dependable production systems.

8:17So the takeaway isn't like, "Don't move fast," it's,

8:21"Systematize velocity,

8:22and move fast with those guardrails so speed compounds,

8:27instead of creating a lot of fragility in the system."

8:30>> Yeah, there's no doubt.

8:31I think that the other thing to kind of think about here,

8:34John, is to manage the AI-native technical debt.

8:37You mentioned shadow AI. I have said shadow IT.

8:40But I think when we look at prompt sprawl

8:42and agent sprawl, that's really the new kind of sprawl

8:46that we're seeing in the shadow world of AI or IT.

8:50But model upgrades can silently break workflows,

8:52like you were talking about.

8:55It hides the complexity,

8:57and versioning discipline is now critical, right?

9:00AI really can introduce data debt, context debt,

9:04alignment debt.

9:07Emerging best practices are really kind

9:09of implementing things like AI governance layers

9:11and model registries

9:13and prompt lifecycle management as well

9:15as AI observability stacks.

9:16So the question really isn't, "Can we build it?

9:19" It's, "Can we sustain it at the end of the day?

9:22" When I look at that, AI-

9:24native technical debt is becoming a real issue.

9:27It really is because there's a lot of development out there.

9:30But how is it different from traditional software debt,

9:32and what is new about engineering standards that help...

9:35that barely emerge to manage this debt

9:38that's coming a whole lot?

9:42>> Traditional technical debt usually comes from messy code,

9:47shortcuts in architecture, or missing tests,

9:49and it's painful,

9:51but it's familiar territory for the industry.

9:54AI creates new categories of debt

9:58that are a lot less visible.

9:59Like you mentioned, you get prompt sprawl, agent sprawl,

10:03inconsistent context handling,

10:05and systems that quietly break when the model

10:09of prompt changes, and so

10:12at the core difference here is the failure mode.

10:15Because traditional software tends

10:17to fail deterministically, you can reproduce the bug.

10:20AI systems are probabilistic, so

10:23that same workflow can behave differently depending on the

10:28context, on data drift,

10:30or how the model's behavior shifts over time.

10:34What this means is that you really need

10:36to bring new standards.

10:38Versioning can't just apply to code anymore.

10:40It has to also cover prompts, the models, evaluation suites,

10:45and the data context that drives behavior,

10:49and you also need continuous evaluation.

10:51So the question isn't, "Did it work once?

10:54" It's, "Is it still working across scenarios at

10:58scale and under real conditions?

11:01" And I think the organizations

11:03that handle this well treat AI features like real

11:06products with lifecycle management.

11:07It's not just one-off experiments

11:10that are stitched together into a fragile foundation.

11:13>> Yeah, that makes a lot of sense.

11:16I also want to kind of go back to something

11:17that you said earlier, democratizing, right?

11:20You talk about democratizing the

11:22AI deployments by creating and such.

11:24I want to call it democratizing innovation,

11:26but really democratizing innovation without losing control.

11:29And looking at it in the context of non-technical teams,

11:33they're really closest to their own business problems.

11:35That's where AI opportunity often lives

11:38because they're trying to solve a problem.

11:40But the risk is decentralizing experimentation without

11:43guardrails causes...

11:44obviously, they can go off the rails real quick.

11:48So what we see in best practices,

11:49we see winning organizations really are creating safe

11:52sandboxes, internal AI platforms, curated model access,

11:56and clear usage policies.

11:57Those are the ways to put those guardrails in place.

11:59But AI literacy is becoming a competitive advantage,

12:02not just a developer skill,

12:04and the shift from citizen developer

12:06to citizen AI operator is really a model that's happening.

12:09I'm seeing that.

12:11In a lot of organizations, I'm sort of looking at it.

12:14So, John, when we look at democratizing AI innovation,

12:18it sounds powerful,

12:19but how do you really bring non-

12:21technical teams into the fold without creating

12:24chaos or shadow AI?

12:27>> Yeah. Like I mentioned, we're kind of going

12:30through this internally,

12:31and what we did was we started with practical

12:36hands-on training that is focused on prompt fundamentals

12:42and workflow patterns

12:43and then immediately moved to applied AI work

12:47where teams

12:49beyond engineering are producing real

12:52deliverables on real tasks.

12:56We also set principles to keep that safe.

13:00Those principles include individuals on the output, review

13:04of the output is non-negotiable, we share learnings openly

13:09in a shared Slack channel or in team meetings,

13:12and then iterate on

13:14that over time rather than waiting for perfection.

13:18And so that's how we've tackled that internally, especially

13:21as we've wanted to upskill the entire organization on

13:25AI and make it a useful tool in everybody's workflows

13:29beyond engineering.

13:31And I think, if we ask ourselves like,

13:35"Why do engineers seem ahead of everyone else?

13:37" I think engineers are starting ahead

13:40because code is structured

13:42and the feedback loop is immediate.

13:44You write something, you run it,

13:46and then you see what happens,

13:47but the next wave is cross-functional.

13:50The biggest unlock happens when domain experts can express

13:55workflows clearly and consistently

13:56and AI becomes

13:58that daily overlap across every role in the organization.

14:02>> Yeah. No, I really like where this conversation's going

14:04because it really kind of drives the point home of culture,

14:07bottoms-up versus a top-

14:09down AI strategy when you're looking at culture.

14:11>> Yeah. - So bottoms-up really drives

14:14that innovation speed, like you were talking about.

14:16Tops-down really ensures risk management and alignment.

14:20But when you look at it, too much of a bottom-

14:22up approach can potentially equal fragmentation,

14:25and too much of a top-down approach is bureaucracy-

14:29installed adoption, right?

14:30We kind of call it what it is.

14:31But I think the emerging model into where you're going

14:35with this is centralized AI platform strategy

14:39and decentralized use case

14:42innovation is really key.

14:44But I also think,

14:45and you kind of nailed this in my opinion,

14:48executive sponsorship is critical,

14:50but so is grassroots momentum, so you need to kind of have

14:53that balance between the two.

14:55And when we look at it, there's often tension between

15:01the bottom-up experimentation and the top-down mandates.

15:04So, John, what's the right balance

15:06to scale AI responsibility really while

15:09maintaining that velocity?

15:10And it's kind of on both sides, right?

15:13>> Yeah. I mean, this tension shows up everywhere right now.

15:17Bottoms-up experimentation creates speed and creativity,

15:21and that's where the surprising use cases are coming from.

15:25But too much bottom-up leads to fragmentation, so

15:30different tools, different standards, duplicated work,

15:35and risk exposure you potentially don't even know about.

15:40On the flip side, top-down leadership is necessary

15:43for alignment and risk management,

15:46but mandates don't really work either.

15:50A target like increased usage of AI

15:55by 20% is a metric, it's not a strategy.

15:59And so the best pattern we've found is a hybrid model,

16:03centralize the infrastructure and the guardrails,

16:05but decentralize use case discovery.

16:09That's also what we're doing internally as well.

16:12Leadership sets the direction, the principles

16:15and the safety boundaries,

16:16and then teams experiment out loud within those constraints

16:20to share what works.

16:23I think when you get that balance rate,

16:24you don't kill velocity.

16:26It's, again, kind of going for that velocity compounded.

16:31When I think about what should executives actually do,

16:34I think it's fund

16:36and prioritize shared infrastructure, define governance,

16:41and align the teams around the right outcomes,

16:45but don't expect the best use cases to emerge...

16:50actually, rather, expect the best use cases

16:53to emerge from the people who are closest

16:55to the work and not from the top.

16:58I think you could also avoid stifling innovation

17:00by making the rules about safety and quality

17:04and not about controlling ideas,

17:06and so you want people to try things freely inside

17:10of a sandbox, as you mentioned, where failure is really safe

17:13and also those learnings get shared.

17:16>> Yeah, for sure. But it also comes down

17:18to organizational readiness and orchestration.

17:20I mean, when we look at this, the high success

17:22of AI successes, it increasingly is about cross-

17:25functional orchestration, right?

17:26Whether you have engineering, security, legal, product,

17:28and finance, they're all working together,

17:31and that's a big factor.

17:32But if you have an AI change in cost model,

17:36which has variable inference costs versus fixed

17:38infrastructure, you have procurement

17:40and government processes that they have to evolve,

17:43and then there's also measurement for success

17:45because it's from that deployment that we're talking about,

17:49and this is something that is really near

17:51and dear to my heart when it comes to like the SDLC

17:54and across delivery, you have

17:56to look at things like productivity gains

17:58and cycle time reduction

18:00and revenue influence and risk reduction.

18:03Those are the areas that I think are really going

18:05to help measure the balance, to your point there.

18:11But I guess I would ask this question here, it's like,

18:14today, if you were advising a board

18:18of leading indicators that really would help drive

18:22or tell whether the organization is truly optimized

18:25or operationalizing AI, what would you say is

18:31the optimized

18:32and operationalized AI versus just experimenting with it?

18:35What would you say that to the board

18:38from those measurement perspective?

18:40Because I think that's important for the

18:41audience to understand.

18:43>> Yeah. I think

18:44that the biggest trap is mistaking

18:47pilots for progress.

18:49It's really easy to create impressive demos,

18:53and it doesn't mean that the organization is

18:55generating real value.

18:56So I think the clearest leading indicator

19:02is whether AI is actually embedded into real workflows,

19:07and are teams using it to produce real outputs every week,

19:12and are those outputs improving speed, quality

19:16or decision making.

19:19Another indicator is measurement.

19:22Mature organizations track things, like you mentioned,

19:25productivity gains, cycle time reduction,

19:30risk reduction, and it's not just the number of AI projects.

19:35And I think a third is governance maturity.

19:38Are there evaluation pipelines? Is there monitoring?

19:41Is there version discipline?

19:44Is there clear ownership in the organization?

19:46And then I think there are real business outcomes.

19:49In the training context, for example, like our customers,

19:53we've seen things like faster onboarding, higher confidence,

19:58and measurable improvements in communication skills when the

20:01AI is genuinely embedded in how teams learn and improve.

20:05I think boards should be asking themselves three questions,

20:08"Where is AI creating measurable value?

20:12How are we managing the risk?

20:14And do we have AI-native engineering standards

20:18that make it sustainable?

20:19" And so if I had to pick, I think, one leading indicator,

20:24it's workflow penetration.

20:25If AI is showing up naturally in the cadence of work

20:29and people are sharing repeatable patterns, it's real.

20:32If it's only demos and isolated pilots, it's not.

20:35>> Yeah. Yeah, no, that makes a lot of sense.

20:37I think there's a lot that we talked about

20:39during this presentation, a lot

20:41for the audience to think about.

20:43It's new to a lot of organizations,

20:45but it's also something that they need to be aware of.

20:47It's a competitive advantage,

20:49and those organizations that don't take advantage are going

20:51to be kind of left behind, honestly.

20:53But, John, thanks for being on.

20:55What would you leave with the audience on

20:57where they can learn more

20:58about what we're talking about today?

21:00>> One thing I just want to add is one thing

21:03that we've learned across industries is that

21:06as automation takes on routine work, the conversations

21:11and decisions that reach humans get more complex,

21:15more nuanced, and even more higher stakes.

21:20So the opportunity isn't only, "Where can AI replace people?

21:25" The real opportunity is,

21:27"How do we make humans dramatically better at the

21:31moments that matter the most?

21:32" And to do that responsibly, you need both sides.

21:35You need rapid experimentation

21:37and strong operational discipline

21:40so the systems you build are actually trustworthy at scale.

21:45I think one other item is if you'd like

21:48to learn more about ReflexAI, I really encourage you to go

21:51to reflexai.

21:52com and feel free to reach out. We'd love to chat with you.

21:56>> John, thanks for being on today.

21:57It's really an interesting conversation, really a lot

22:00for the audience to think about regardless

22:02where they are on their journey.

22:03It's definitely something for them to consider.

22:05And a big thank you to all of you who've tuned in.

22:07We really do appreciate you being part

22:09of the AppDevANGLE community.

22:10But for now, that wraps up this episode,

22:12but we'll be back next Wednesday

22:13with another conversation diving into the tools, trends,

22:15and talent shaping the future of application development.

22:18So when you're deploying at the edge, building with AI,

22:21or modernizing your cloud stack, we've got you covered.

22:23Be sure to follow us on social if you have any thoughts,

22:25questions, or just want to connect.

22:27Until next time, stay curious and stay building.