Latest / Today’s AI News / Why 80% of AI Projects Fail, What the Cloud Era Teaches Us, and Where AI Won't Replace You

Transcript

Why 80% of AI Projects Fail, What the Cloud Era Teaches Us, and Where AI Won't Replace You · Today’s AI News

• 0:00
  Welcome back to today's AI News, brought to you by 9X Roductions.
• 0:04
  Having the world's smartest AI models in your business is like
• 0:08
  putting the greatest musicians in a room together without a
• 0:12
  conductor and sheet music to coordinate them.
• 0:14
  You don't get a Symphony, you just get really expensive noise.
• 0:18
  Yeah, just a lot of chaotic noise.
• 0:19
  Exactly. And today we are talking about
• 0:21
  how you can become that conductor.
• 0:24
  Our mission for this deep dive is to explore an exclusive
• 0:27
  interview with UI Path's Chief Marketing Officer, Michael
• 0:30
  Atalla. We're going to figure out why
• 0:32
  like 70 to 80% of AI projects just completely fail to get off
• 0:36
  the ground, which is. Just a staggering number when
• 0:39
  you think about. It, it really is.
• 0:40
  So we'll look at this concept he calls agentic business
• 0:43
  orchestration, which seems to be the missing puzzle piece here.
• 0:46
  And, you know, most importantly, we'll break down what this
• 0:49
  massive shift actually means for your daily work and your future
• 0:52
  career. To really understand the scale
• 0:54
  of what we're going through, we kind of have to look back at
• 0:56
  where enterprise AI started. Think about maybe five years
• 1:01
  ago, right around the time UI Path had their IPO.
• 1:04
  They were the undisputed leader in robotic process automation,
• 1:08
  or RPA. And the pitch back then was
• 1:10
  incredibly straightforward. It was just automate the task,
• 1:13
  free the person. OK, let's unpack this for a
• 1:16
  second. Because robotic Process
• 1:17
  Automation is essentially the digital version of a factory
• 1:21
  assembly line, right? That's the perfect way to
• 1:23
  describe it. You take these highly
• 1:25
  repetitive, soul crushing tasks like opening an e-mail, copying
• 1:30
  a client's address, pasting it somewhere else, and you just
• 1:33
  hand that specific isolated sequence over to a rigid
• 1:37
  software bot. Exactly.
• 1:38
  It was purely deterministic. You know, if X happens, do Y.
• 1:41
  And it was highly successful for what it was.
• 1:44
  But the landscape has shifted so dramatically since then.
• 1:47
  The pitch today isn't just about parallel task automation.
• 1:50
  Now the whole goal is agentic business orchestration.
• 1:54
  OK, agentic business orchestration.
• 1:55
  I mean, that sounds like the ultimate corporate jargon Bingo
• 1:58
  winner. Oh, totally, It really does.
• 2:00
  But if you break it down, it's actually pretty intuitive.
• 2:03
  It's like the difference between buying a bunch of smart home
• 2:06
  gadgets that don't talk to each other where you have, you know,
• 2:09
  one app for your Philips Hue lights, another app for your
• 2:12
  smart locks, and 1/3 for your thermostat.
• 2:14
  Right, a total mess. Yeah.
• 2:16
  Versus having a centralized system where just unlocking the
• 2:19
  front door automatically turns on the hallway lights and
• 2:22
  adjusts the thermostat based on who just walked in.
• 2:25
  That's a great analogy, yeah. And if we connect this to the
• 2:28
  bigger picture, you can see how Atala's own background shaped
• 2:32
  this philosophy. He sent 15 years at Microsoft
• 2:35
  and a really defining era for him was leading marketing for
• 2:38
  Office around 2011. Oh wow, 20. 11 yeah during its
• 2:41
  massive transition from on premise software to the cloud
• 2:45
  based Office 365 that. Feels like a completely
• 2:48
  different era in tech. I mean, he was literally sitting
• 2:51
  in boardrooms demoing things like conversation view in e-mail
• 2:55
  or the Do Not Reply All button to these skeptical executives.
• 2:59
  Exactly. And those executives were
• 3:00
  looking at him, you know, being asked to move their most
• 3:03
  sensitive information off a physical server they could
• 3:06
  actually touch in their basement and put it in this invisible
• 3:09
  thing called the cloud. Right.
• 3:11
  Terrifying for them. Completely and Atala learned a
• 3:14
  foundational lesson there. You can have the absolute best
• 3:17
  tech product in the world, but you will still lose the customer
• 3:21
  if you don't help them Rethink how the underlying work gets
• 3:24
  done. The companies that failed the
• 3:26
  cloud transition were the ones that just tried to lift and
• 3:28
  shift. So they just took their terrible
• 3:30
  bottlenecked legacy processes and hosted them on a faster
• 3:35
  server. They basically just digitized
• 3:37
  their bad habits. Spot on.
• 3:39
  And he points out that this exact same pattern of failure is
• 3:42
  playing out with AI right now. Most AI leaders are completely
• 3:46
  fixated on the model itself. They sit around debating
• 3:49
  parameters, benchmarking, reasoning, focusing on what the
• 3:52
  model can do in a vacuum. But evaluating a model in a
• 3:55
  vacuum completely ignores how a business actually functions.
• 3:58
  So if your team is looking at AI tools right now, you really have
• 4:01
  to shift the conversation. Stop asking what can this model
• 4:04
  do? And start asking what does our
• 4:07
  workflow need to look like for this to actually work?
• 4:09
  Yes, failing to redesign those workflows is the main reason why
• 4:13
  the vast majority of AI pilots just hit a brick wall.
• 4:17
  Like you said earlier, 70 to 80% of agentic AI initiatives never
• 4:22
  make it out of the pilot stage and nearly half of organizations
• 4:25
  are calling AIA massive disappointment.
• 4:28
  I have to play devil's advocate here though.
• 4:29
  Is this just a lack of ambition or I don't know, a lack of
• 4:33
  funding from conservative leadership?
• 4:35
  Because the hype around this is deafening.
• 4:36
  Why are well funded companies failing post deployment?
• 4:40
  It's fascinating because Atala is adamant that it is not a
• 4:42
  motivation problem. The ambition is definitely there
• 4:45
  from the CEO down to the employee who just wants an
• 4:48
  easier work day. The issue is a severe
• 4:50
  coordination problem. Meaning what exactly?
• 4:53
  Well, let's go back to your smart home analogy.
• 4:56
  These pilots almost always run in total isolation.
• 4:59
  You deploy one super smart AI agent in customer service to
• 5:03
  handle returns and meanwhile you have a rigid traditional bot
• 5:07
  running in accounts payable. Right, so they're succeeding in
• 5:11
  their tiny silos. The customer service bot talks
• 5:14
  to an angry customer and successfully authorizes a
• 5:17
  return. Perfectly.
• 5:19
  But that isolated success actually creates a nightmare
• 5:22
  downstream. The bot authorizes the return,
• 5:25
  but because there's zero orchestration, it doesn't notify
• 5:28
  inventory management to expect the item back.
• 5:30
  Oh geez. Right.
• 5:31
  It doesn't trigger accounting to prep the refund.
• 5:34
  It doesn't generate the shipping label and logistics.
• 5:36
  The pilot worked, but the business process completely
• 5:39
  broke and then humans have to rush in and manually bridge the
• 5:42
  gap between those disconnected system.
• 5:43
  Exactly which burns time and money.
• 5:46
  The ROI just disappears into the gaps between the tools.
• 5:49
  Leadership sees the mounting hidden costs, they realize it
• 5:52
  won't scale, and the project gets killed.
• 5:55
  Pilot purgatory. So the companies that are
• 5:57
  winning that 20%? They treat AI agents just as
• 6:01
  components of a larger governed workflow.
• 6:04
  You crack the code when you map out the territory first.
• 6:07
  Where does work originate? Where do handoffs happen and
• 6:10
  where are decisions made? Wait, but isn't the whole
• 6:12
  promise of modern AI that it's smart enough to figure out messy
• 6:16
  environments on the fly? Why are we manually mapping out
• 6:19
  workflows like it's a 2005 corporate retreat?
• 6:22
  Because without boundaries, an AI agent will just optimize for
• 6:26
  its specific task without any understanding of the downstream
• 6:29
  consequences. Governance is what prevents an
• 6:33
  overly helpful AI from promising $1000 credit to a user just to
• 6:38
  close a ticket quickly. Right?
• 6:39
  You have to map the handoffs to establish where the agents
• 6:42
  authority ends and human oversight begins.
• 6:44
  But if a company successfully maps out all those handoffs,
• 6:46
  they immediately run into a terrifying realization, right?
• 6:49
  Almost all of those handoffs are currently done by human beings.
• 6:52
  What happens to those people when agents take over?
• 6:54
  Yeah, that is the defining anxiety of this era.
• 6:57
  And the data shows a massive chasm in perspective. 75% of AI
• 7:03
  experts are highly optimistic about a is impact on jobs.
• 7:06
  But when you pull the public, only 23% share that optimism.
• 7:10
  That's a staggering disconnect. And honestly they have a right
• 7:13
  to be worried. We are seeing entry level dev
• 7:16
  jobs drop nearly 20 since 2024. Both sides are looking at
• 7:20
  reality but focusing on different horizons.
• 7:23
  Experts see the historical arc creating new industries.
• 7:27
  The end user sees their specific tasks being absorbed by a
• 7:30
  machine and wonders how they'll pay rent next year.
• 7:33
  Here's where it gets really interesting.
• 7:35
  You'd think Attala, leading marketing for a massive
• 7:38
  automation company, would lean into the humans are obsolete
• 7:41
  narrative to sell more software, but he pushes back fiercely
• 7:45
  against the idea that humans are becoming optional.
• 7:47
  His reasoning really cuts to what AI actually is under the
• 7:50
  hood, and LLM is at its core a staggering probability engine.
• 7:55
  It just mathematically predicts the next logical sequence of
• 7:58
  words based on training data. Because it's math, it has no
• 8:01
  intrinsic motivation. No personal taste.
• 8:03
  Right. No taste.
• 8:05
  And crucially, absolutely zero instinct for risk.
• 8:09
  It fundamentally cannot ask the single most important business
• 8:12
  question, should we? Exactly.
• 8:15
  An AI agent can tell you how to execute a massive marketing
• 8:17
  campaign efficiently. It can't tell you if that
• 8:20
  campaign aligns with your core values, or if it might alienate
• 8:23
  your loyal demographic, or if it's just culturally tone deaf.
• 8:27
  It doesn't lose sleep if the company goes bankrupt.
• 8:29
  Exactly. Humans have a biological and
• 8:32
  financial stake in the outcome. Humans alone possess an instinct
• 8:35
  for risk. So the work doesn't evaporate,
• 8:38
  but it profoundly changes shape. AI absorbs the routine,
• 8:42
  predictable parts. It's like spreadsheets.
• 8:45
  When Excel came out, it didn't eliminate accounting, it
• 8:48
  eliminated the human calculator and shifted workers into higher
• 8:52
  level analysis. The required skill set shifts
• 8:55
  entirely. Organizations are suddenly
• 8:57
  starved for people who understand workflow design, AI,
• 9:00
  governance, process ownership. Italia gave a great personal
• 9:04
  example about his 13 year old daughter.
• 9:05
  Oh. Yeah, I remember this part.
• 9:07
  He pointed out that when she applies for College in five
• 9:10
  years, the job she'll be competing for probably haven't
• 9:13
  even been named yet. Which is an incredible testament
• 9:16
  to the velocity of this shift, but it also presents a harsh
• 9:20
  reality. You know, telling a 24 year old
• 9:23
  that cool unnamed jobs will exist in a decade is cold
• 9:27
  comfort if they're coding job is vanishing today.
• 9:29
  Yeah, the anxiety transition is very real and leadership has to
• 9:33
  manage it, but we have to separate transition anxiety from
• 9:36
  the fear of total permanent replacement.
• 9:38
  Well, that distinction gets really clear when we look at how
• 9:41
  tasks are practically divided in the real world.
• 9:43
  Right now, the labs pushing full autonomy are running way ahead
• 9:47
  of commercial reality. Let's look at the actual UI path
• 9:49
  playbook. How do they divide work?
• 9:51
  They divide it based on 2 criteria, ambiguity and
• 9:54
  accountability. Let's imagine you're on a busy
• 9:56
  finance team. OK, your day is pulling raw
• 9:59
  data, reconciling it across clunky systems, chasing
• 10:02
  approvals. In an orchestrated system, that
• 10:05
  workload is broken into 3 distinct layers.
• 10:07
  Layer one is traditional automation.
• 10:10
  So that handles the deterministic black and white
• 10:13
  work. An invoice comes in, a bot
• 10:15
  scrapes the numbers and routes it to the database. 0 Cognitive
• 10:18
  effort. Right, the bot just follows
• 10:20
  rigid if then rules. But business is messy and
• 10:24
  eventually you hit Gray areas. That triggers layer 2 AI agents.
• 10:29
  They step in for unstructured data and ambiguity.
• 10:32
  Like, say a new vendor submits an invoice in a totally bizarre
• 10:36
  format, or a transaction is mathematically flawless, but it
• 10:40
  looks anomalous for that client. Or a furious customer e-mail
• 10:43
  comes in using sarcasm instead of standard complaint keywords.
• 10:46
  Right. A traditional bot would crash
• 10:48
  there, but an AI agent uses probabilistic reasoning to parse
• 10:52
  the sarcasm or extract data from the weird invoice.
• 10:54
  Exactly, but even the agent has trick limits.
• 10:57
  It prepares the context, but it doesn't pull the final trigger.
• 11:00
  That handoff leads to layer 3. The humans us.
• 11:03
  The people with actual skin in the game.
• 11:05
  Humans handle the ultimate exceptions.
• 11:07
  You review the context, apply your lived experience, your
• 11:10
  understanding of company politics, and make the judgment
• 11:12
  call. Humans handle the tasks that
• 11:14
  carry real world consequences. Because if an autonomous agent
• 11:18
  approves A disastrous $1,000,000 transfer, you can't fire an
• 11:21
  algorithm. You can't put an LLM in federal
• 11:24
  prison. True accountability requires a
• 11:26
  human name on the dotted line. Exactly.
• 11:29
  And when you look at the day-to-day reality of that
• 11:31
  finance worker after orchestration, they aren't
• 11:34
  manually copying numbers for six hours.
• 11:37
  Their day shifts to reviewing exceptions, managing risk,
• 11:40
  making strategy decisions. It takes the robotic work out of
• 11:43
  the human. So for you listening right now,
• 11:46
  the immediate take away is to categorize your own tasks.
• 11:49
  Ask yourself what parts of my day are rigidly rules based,
• 11:53
  what parts are ambiguous but low risk, and what decisions do I
• 11:56
  make that carry systemic accountability?
• 11:59
  That is how your job will be restructured very soon.
• 12:01
  It's a very grounded approach, less sensational than artificial
• 12:04
  super intelligence, but it actually holds up under the
• 12:07
  pressure of running a profitable business.
• 12:09
  OK, let's distill all this down to the essentials.
• 12:11
  To cap off this deep dive, let's run through Atala's most urgent
• 12:14
  quick hits for organizations adopting AI right this second.
• 12:18
  OK, the biggest mistake enterprises make is expecting AI
• 12:22
  to magically fix a broken process.
• 12:25
  Yes, people treat AI like a Band-Aid for bad management, but
• 12:29
  Atala says AI just makes a good workflow faster and a bad
• 12:33
  workflow significantly more expensive.
• 12:36
  Every time an agent executes a task you pay for compute, you
• 12:39
  burn tokens. If your approval process is a
• 12:42
  bureaucratic nightmare of endless loops, dropping an AI
• 12:45
  agent into it just executes that nightmare at lightning speed.
• 12:48
  You're just scaling your dysfunction and burning compute
• 12:51
  costs every second. You have to fix the human
• 12:53
  process first. OK, the second quick hit is
• 12:55
  about how you introduce this to your workforce.
• 12:57
  Framing is everything. It ties back to transition
• 13:00
  anxiety. If leadership introduces AI as
• 13:03
  this mysterious cost cutting force happening to the
• 13:06
  employees, the natural response is sheer terror.
• 13:08
  But if you slip the script and frame as a capability they will
• 13:11
  build with giving them agency to deploy agents to eliminate the
• 13:14
  worst parts of their own jobs, you bypass that resistance.
• 13:17
  It becomes a copilot, not a replacement.
• 13:19
  Which raises an important question about where our focus
• 13:22
  should be, leading to Atala's final point.
• 13:25
  He's fascinated by what he calls the clarity problem.
• 13:28
  Yes, the gap between what an organization believes AI will do
• 13:34
  and what they actually set up the infrastructure to
• 13:36
  accomplish. We spend all our energy
• 13:38
  agonizing over whether models are smart enough or fast enough,
• 13:42
  but the technology is almost never the bottleneck anymore.
• 13:45
  The models are incredible. They are.
• 13:47
  The bottleneck is human clarity, our process design
• 13:50
  documentation, our understanding of our own business mechanics.
• 13:54
  It's lagging decades behind the tech.
• 13:56
  The instruments are perfectly tuned and ready to play, but the
• 13:58
  conductor doesn't know what sheet music they're handing out.
• 14:01
  You can't just wave a baton at 100 brilliant musicians and hope
• 14:04
  they accidentally play Mozart. You need orchestration.
• 14:07
  That leaves us with a somewhat uncomfortable thought for you to
• 14:09
  Mull over. Today.
• 14:11
  We established that applying AI to a bad workflow just makes
• 14:14
  failure more expensive. Look critically at your own work
• 14:17
  day. What hidden, deeply broken
• 14:20
  processes are currently running in your daily routine?
• 14:24
  What chaotic tasks are you brute forcing manually that you
• 14:27
  desperately need to untangle before you even consider handing
• 14:31
  them to an AI agent? Because if you don't fix the
• 14:33
  workflow first, you are just automating the chaos.
• 14:36
  Don't automate the chaos. That's a perfect place to leave
• 14:39
  it. That's all the time we have for
• 14:41
  today's deep dive. Make sure you hit subscribe so
• 14:43
  you never miss out and stay updated with today's AI news.
• 14:46
  If you had an aha moment today, please take a second to rate the
• 14:49
  show 5 stars. It really helps us out.
• 14:51
  Thank you so much for joining us, and we will see you right
• 14:53
  back here for tomorrow's Deep dive.