Latest / Elon Musk Podcast / AI UPDATE: Infosys Replaces Human Labor With Anthropic

Transcript

AI UPDATE: Infosys Replaces Human Labor With Anthropic · Elon Musk Podcast

• 0:00
  Welcome back to the AI update. It is really good to be here.
• 0:02
  Indian IT giant Infosys has officially partnered with
• 0:06
  Anthropic to deploy AI agents across their enterprise network.
• 0:10
  Which is honestly a massive signal hiding in plain sight.
• 0:14
  You see these corporate handshakes all the time, and you
• 0:18
  might just brush it off as another chatbot integration.
• 0:20
  But that completely misses the point of what is actually
• 0:23
  happening. We were looking at a fundamental
• 0:25
  rewiring of a $283 billion IT industry.
• 0:30
  So what happens when the business of outsourcing human
• 0:35
  labor gets replaced by orchestrating software?
• 0:37
  That is the exact tension we are seeing in the market right now.
• 0:40
  We are going to look at the new Anthropic Center of Excellence
• 0:43
  and exactly how these legacy systems are being transformed,
• 0:46
  and we will get right into that after this very short break.
• 0:48
  To really understand this shift, you have to look at how this
• 0:51
  industry has operated for decades.
• 0:53
  Right, the traditional model is incredibly straightforward.
• 0:56
  Yes, for a long time, that $283 billion industry has run on a
• 1:01
  very specific fuel, and that is human labor.
• 1:04
  Bodies in seats. Exactly.
• 1:06
  You hire smart people, often in lower cost geographies like
• 1:09
  India, to write the code or manage the servers or answer
• 1:14
  support tickets. It is the classic outsourcing
• 1:16
  model. But what Infosys is announcing
• 1:18
  here with Anthropic is a hard pivot away from that.
• 1:21
  We are seeing a move from outsourcing humans to
• 1:25
  orchestrating agents. Orchestrating agents, which
• 1:28
  sounds very futuristic but also a little ominous if you are one
• 1:33
  of the humans used to being outsourced.
• 1:35
  Well, it is the existential tension of the decade for the
• 1:37
  services sector. So here is the key point.
• 1:39
  We are pulling this discussion from the transcripts of Emphasis
• 1:42
  Investor Day 2026, which just wrapped up right along with
• 1:46
  their official partnership announcements from February
• 1:48
  18th, 2026. And the press release is very
• 1:52
  specific about the wording they use.
• 1:54
  Very specific. They are not just deploying
• 1:56
  generative AI, they are deploying AI agents.
• 1:58
  Why did they choose that specific word?
• 2:00
  It is probably the most important word in the entire
• 2:03
  document. A chatbot like the early
• 2:05
  versions of ChatGPT or Clod is passive.
• 2:08
  It is conversational. You ask a question and it gives
• 2:11
  an answer. Exactly.
• 2:12
  It just sits there waiting for you.
• 2:14
  But an agent represents action. An agent does not just tell you
• 2:17
  how to reset a password. It actually does it.
• 2:19
  Right, it logs into the system, it navigates the user interface,
• 2:23
  it resets the password and then emails the user.
• 2:26
  So a chatbot is basically a consultant offering advice, but
• 2:30
  an agent is an actual employee doing the job.
• 2:33
  That is the perfect way to look at it, yeah.
• 2:35
  And Infosys is betting that their enterprise clients, the
• 2:38
  massive banks and telecom giants, they do not need more
• 2:42
  consultants giving advice. They need digital employees to
• 2:45
  do the grunt. Work, yes, and they have a very
• 2:47
  clear road map for this. They are not just releasing wild
• 2:50
  agents into the corporate network to do whatever they
• 2:52
  want, Which? Would be a total disaster.
• 2:54
  Absolute chaos. So they are rolling this out
• 2:58
  sector by sector, starting with telecom and once they prove the
• 3:01
  agents can handle that piece, they move to financial services
• 3:05
  than manufacturing and finally software development itself.
• 3:09
  Starting with telecom is an interesting choice.
• 3:12
  Why not finance, which seems to have the deepest pockets?
• 3:15
  Telecom is high volume and high complexity, but the data is
• 3:19
  extremely structured. It is basically the perfect
• 3:22
  testing ground. Because you have millions of
• 3:24
  logs and signal data and routing tickets.
• 3:26
  Right. And if an agent messes up a
• 3:28
  network configuration, it is bad, but you can usually fix it
• 3:31
  in the software layer. Whereas if an agent messes up a
• 3:34
  SWIFT wire transfer in financial services.
• 3:36
  That is a resume generating event for everyone involved.
• 3:39
  You do not want to accidentally wire a billion dollars to the
• 3:42
  wrong account. So they start where it is safer.
• 3:45
  Which brings us to the partner choice.
• 3:47
  Why Anthropic? Yeah, they could have gone with
• 3:50
  Open AI or Google or even open source Llama models.
• 3:53
  But they went all in on Claude. I assume this links back to the
• 3:57
  constitutional AI and safety branding Anthropic is known for.
• 4:00
  It is the primary driver here. Infosys brings the domain depth.
• 4:05
  They know exactly how a bank's messy legacy back end works.
• 4:08
  They know where the skeletons are in the closet.
• 4:10
  Exactly, but Anthropic brings Frontier models with a safety
• 4:13
  first design. In a highly regulated industry,
• 4:16
  you cannot have an AI that hallucinates.
• 4:19
  You cannot have a model that accidentally leaks customer data
• 4:23
  into its training set. And Anthropics architecture is
• 4:26
  built specifically to prevent that.
• 4:28
  This follows their push into the enterprise sector last month.
• 4:32
  Right. In January 2026, they launched
• 4:34
  Claude Cowork. Which was their direct play for
• 4:37
  the white collar desktop, helping individual workers with
• 4:40
  office tasks. So if Claude Co work is the
• 4:43
  individual tool for a worker, this Infosys partnership scales
• 4:48
  that up to the enterprise level. Exactly.
• 4:50
  Anthropic builds the engine and Infosys builds the chassis.
• 4:53
  The transmission and the steering wheel, and they are
• 4:56
  building it inside something they are calling the Anthropic
• 4:59
  Center of Excellence. The Coee.
• 5:01
  Vivek Sinha from Infosys called this a milestone moment.
• 5:04
  But every massive company has a center of excellence for
• 5:07
  something. Usually it is just a conference
• 5:09
  room with a nice plaque. And slightly better coffee.
• 5:11
  Right. Is this one actually doing
• 5:13
  anything different? I was deeply skeptical too,
• 5:15
  until I looked at the engineering capabilities they
• 5:18
  listed. They're not just doing basic
• 5:20
  prompt engineering. They're doing real technical
• 5:22
  work. Deep technical work.
• 5:23
  The first one that stood out was knowledge distillation.
• 5:26
  OK, so let us look closely at that knowledge distillation.
• 5:30
  To me that sounds like making moonshine in the server room.
• 5:33
  Not quite, though the concept of boiling something down to its
• 5:36
  essence is very similar. Think of the massive frontier
• 5:40
  models like Claude 3.5. They are incredibly smart.
• 5:44
  They know French poetry and quantum physics.
• 5:46
  Right, they are like a brilliant PhD student who has read every
• 5:50
  book in the entire library. But they are massive and very
• 5:53
  expensive to run. Extremely expensive and
• 5:56
  computationally heavy. Now, if you are a Bank Lauder,
• 5:59
  you do not need a PhD student to process mortgage applications.
• 6:03
  You need a smart clerk who knows mortgage rules perfectly and
• 6:06
  nothing else. Exactly.
• 6:08
  Knowledge distillation is the process where the massive
• 6:11
  teacher model teaches a smaller student model only what it needs
• 6:14
  to know for a specific task. So you strip away the poetry and
• 6:18
  the quantum physics. And you were left with a model
• 6:20
  that is incredibly good at checking credit scores, but
• 6:22
  literally nothing else. And crucially, that smaller
• 6:25
  model is 100 times cheaper to run.
• 6:27
  And 10 times faster margins are everything in enterprise tech.
• 6:32
  Because if every single query costs a dollar, you go broke
• 6:36
  very fast. But if it costs a fraction of a
• 6:39
  cent, you have a sustainable business if Infosys can sell a
• 6:42
  service that costs them pennies to run because they distilled
• 6:45
  the model. But they charge the client based
• 6:48
  on the value of the outcome. That is where the profit lives,
• 6:51
  That is how they survive this industry transition.
• 6:53
  The next capability on their list was large scale agent
• 6:57
  engineering and orchestration. We touched on agents earlier,
• 7:00
  but orchestration sounds like a massive headache.
• 7:02
  It is arguably the biggest engineering challenge in AI
• 7:05
  today. Building 1 agent to check your
• 7:07
  e-mail is easy. But what happens when you have
• 7:09
  5000 agents running simultaneous continuously inside a global
• 7:12
  supply chain? Absolute chaos.
• 7:14
  Imagine 5000 interns running around a corporate office
• 7:18
  without a manager. You need a manager AI.
• 7:20
  Yes, an orchestrator. An AI that assigns tasks and
• 7:23
  resolves conflicts. So if two agents try to update
• 7:26
  the exact same database record at the same time, who wins?
• 7:29
  The orchestrator decides Infosys is claiming their Center of
• 7:33
  Excellence has actually solved this orchestration layer.
• 7:36
  So it is like being the conductor of a very robotic
• 7:39
  orchestra. That is a very apartment
• 7:41
  analogy. Without the conductor it is just
• 7:43
  noise. Which leads right into the Third
• 7:45
  Point, system intelligence for complex, legacy and hybrid
• 7:49
  environments. The polite term is legacy.
• 7:52
  The real term is spaghetti code from 1995.
• 7:55
  We are talking about mainframes and green screens and cobalt
• 7:59
  code that no one has touched since the Y2K scare.
• 8:02
  Exactly, and you cannot just plug a modern API into a
• 8:05
  mainframe. It does not have one.
• 8:06
  It was built before the Internet even existed.
• 8:08
  So that is the billion dollar problem.
• 8:11
  How does a shiny new AI agent interact with that?
• 8:14
  Do they just rewrite the old? Code never.
• 8:16
  That is way too risky. Usually involves building
• 8:18
  complex wrappers or connectors. Emphasis says they're using AI
• 8:22
  to build the bridge. So an AI that can read the old
• 8:25
  green screen and understand what is happening.
• 8:27
  And then translate that into modern data for the anthropic
• 8:30
  agent to use. It is literally like teaching a
• 8:32
  new robot to use an old metal filing cabinet.
• 8:34
  That makes sense, because banks are certainly not going to RIP
• 8:37
  out their core mainframes just to use a new chatbot.
• 8:40
  The risk is too high. You have to meet the legacy tech
• 8:43
  exactly where it is. If Infosys can solve that
• 8:45
  translation layer seamlessly, that alone justifies a
• 8:49
  partnership. There's one more technical
• 8:51
  capability here, synthetic enterprise data for safe
• 8:55
  training. Yes, this one feels slightly
• 8:57
  counterintuitive. We always hear that data is the
• 9:00
  new oil. Why would you want fake oil?
• 9:02
  Because the real oil is toxic, or rather, it is legally
• 9:07
  radioactive. You mean privacy laws like GDPR?
• 9:10
  Exactly. Say you want to train an AI to
• 9:12
  detect fraud in credit card transactions.
• 9:14
  You have petabytes of real transaction data.
• 9:17
  But you cannot feed that into a cloud based model.
• 9:21
  No, because it contains real names and real credit card
• 9:23
  numbers and real spending habits.
• 9:25
  That is a massive data breach waiting to happen.
• 9:28
  Right, so you use the Center of Excellence to generate synthetic
• 9:30
  data. The AI analyzes the statistical
• 9:33
  properties of the real data. The patterns and the variance
• 9:36
  and the frequency of the fraud. And it creates a completely new
• 9:39
  data set there. Looks mathematically identical,
• 9:42
  but contains zero wheel people. So it creates a John Doe who
• 9:47
  lives on 123 Fake Street but spends money exactly like a real
• 9:51
  banking customer. Precisely.
• 9:53
  You can train the model on that safely and crash test it without
• 9:56
  any risk. And then deploy it on the real
• 9:58
  data later. It bridges the deployment gap.
• 10:00
  It allows innovation to happen without the compliance officer
• 10:02
  having a heart attack. And that leads us to the Garn
• 10:05
  and stuff they listed cost optimization and reliability and
• 10:09
  observability. Which sounds boring to us, but
• 10:12
  for a CIO, observability is how they sleep at night.
• 10:15
  They need to know why the AI made a specific decision.
• 10:18
  If an agent denies a business loan, you cannot just tell the
• 10:22
  regulators that the black box said no.
• 10:24
  We need a trace. And Infosys is building that
• 10:27
  governance layer. So here's what happened with the
• 10:29
  financials. Infosys released some stats
• 10:31
  during this Investor Day. The numbers are very revealing,
• 10:34
  but you do have to read between the lines.
• 10:36
  They announced they are currently working on 4600 AI
• 10:39
  projects. 4600 sounds like a massive volume.
• 10:42
  It is a high volume, but project is a very vague term in
• 10:46
  consulting. Right A2 week proof of concept
• 10:48
  is a project. And a five year global
• 10:50
  transformation is also a project.
• 10:52
  We do not know the exact mix. However, they did reveal that AI
• 10:56
  services accounted for 5.5% of their total revenue in the
• 10:59
  December quarter. And CEO Salil Parekh called that
• 11:02
  robust. So what does that mean in plain
• 11:05
  terms? Is 5.5 percent actually robust?
• 11:08
  It is decent. You have to remember that
• 11:10
  Infosys was the very last of the big three Indian IT firms to
• 11:14
  actually break out this number. They were playing it very close
• 11:16
  to the vest. Exactly.
• 11:18
  So how does that compare to the industry leader TCS, Tata
• 11:22
  Consultancy Services? TCS recently released their Q3
• 11:25
  results. They reported $1.8 billion in
• 11:28
  annualized AI revenue. Which if you do the math, comes
• 11:31
  out to about 5.8% of their fiscal year revenue.
• 11:33
  So Infosys is at 5.5% and TCS is at 5.8%.
• 11:37
  They're basically neck and neck. That is surprising because the
• 11:40
  narrative in the market has been that TCS is way ahead and
• 11:43
  emphasis is scrambling to catch up.
• 11:44
  That narrative was entirely based on silence.
• 11:48
  Because Infosys was not reporting the numbers, analysts
• 11:51
  just assumed they were weak. But this data shows they are
• 11:54
  right in the mix. The main difference is strategy.
• 11:58
  TCS has been very quiet and building internally, focusing on
• 12:02
  their own wisdom Next platform. Well, Infosys is taking a much
• 12:05
  louder partnership heavy approach.
• 12:07
  Partnering with NVIDIA earlier and now this major push with
• 12:11
  Anthropic, they are making noise to show they are on the cutting
• 12:14
  edge. It feels like Infosys is trying
• 12:16
  to win the mindshare battle, even if the actual revenue
• 12:19
  battle is currently a tie. Absolutely have to.
• 12:21
  And this brings us to the existential pivot.
• 12:24
  Right. We have talked about the tech
• 12:25
  and the horse race between these massive companies.
• 12:27
  But there is a much bigger threat looming over this entire
• 12:30
  sector. We have that Raiders report
• 12:32
  regarding the broader Indian IT landscape.
• 12:35
  It highlighted rising investor concerns, yes.
• 12:37
  And the core concern is simple. The traditional business model
• 12:40
  is about to break. Broken how, exactly?
• 12:42
  These companies are incredibly profitable giants.
• 12:45
  For now they are, but their model is based on labor
• 12:49
  arbitrage. The old time and materials
• 12:51
  trick. Exactly.
• 12:53
  I hire a developer in Bangalore for $30.00 an hour and I bill
• 12:57
  them to a client in New York for $100.00 an hour.
• 13:00
  The margin is entirely on the human hour.
• 13:02
  The more hours a project takes, the more money Infosys makes.
• 13:06
  But what happens when an Anthropic agent running on
• 13:09
  Infosys Topaz can write that exact same code in three seconds
• 13:12
  for a cost of $0.05? You cannot bill 3 seconds of
• 13:15
  work to a client. Right if efficiency as he goes
• 13:17
  up by 1000%, your billable hours go down by 90%.
• 13:21
  If Infosys and TCS just stick to the old model, AI is going to
• 13:25
  destroy their revenue stream. They are effectively designing
• 13:28
  the technology that eats their own lunch.
• 13:30
  That is the ultimate innovator's dilemma.
• 13:32
  So this pivot to agentic AI is not just about offering a cool
• 13:36
  new service, it is about fundamentally changing how they
• 13:39
  get paid. It has to be.
• 13:41
  They're trying to move from selling effort to selling
• 13:45
  outcomes. Selling the hours worked versus
• 13:47
  selling the result. Exactly.
• 13:49
  Instead of charging you for the 50 people managing your call
• 13:51
  center, they charge you for 10,000 resolved customer
• 13:54
  tickets. Regardless of whether a human or
• 13:56
  a robotic agent solve them. Right, if they could
• 13:59
  successfully make that shift, AI suddenly becomes a massive
• 14:03
  margin booster. But if they cannot make that
• 14:05
  shift and clients still insist on paying for hours, their
• 14:09
  revenue simply collapses. So Infosys, Topaz and this new
• 14:12
  Center of excellence, these are essentially the vehicles for
• 14:16
  this transformation. They are the lifeboats to get
• 14:18
  them from the service economy to the outcome economy.
• 14:22
  The lifeboat implies a rescue. I would call it a rocket ship.
• 14:26
  They want to be the ones managing the robots.
• 14:28
  They are betting that the sheer complexity of these agents is so
• 14:31
  high that clients will still need a massive partner to manage
• 14:34
  it all. The orchestration and the safety
• 14:36
  and that legacy integration we talked about.
• 14:38
  So the work does not actually go away, it just changes from
• 14:41
  writing code to managing the thing that writes code, correct.
• 14:46
  But that requires a completely different skill set and a
• 14:48
  completely different sales pitch.
• 14:50
  It is a really bold bet, but looking at the speed of
• 14:52
  anthropics development it seems like the only bet they can make.
• 14:56
  It is disrupt yourself or be disrupted.
• 14:59
  There is no third option here. So here's what I want you to
• 15:02
  hold onto from all of this. The technology has officially
• 15:06
  moved from chat to act. Agents are here to do actual
• 15:09
  work, not just talk. This is operational AI. 2nd, the
• 15:14
  barrier to entry is deep engineering.
• 15:16
  You cannot just plug these models in.
• 15:18
  You need knowledge distillation and synthetic data and deep
• 15:21
  legacy integration. That is the Moat Infosys is
• 15:24
  trying to dig. And finally, the scorecard is
• 15:26
  much tighter than we thought. Infosys and TCS are in a dead
• 15:30
  heat. But the finish line keeps
• 15:32
  moving. The winner will not be the
• 15:33
  company with the most engineers, it will be the one with the best
• 15:36
  agents. It really changes your
• 15:38
  perspective on what an IT company actually is.
• 15:40
  We have spent 50 years building an industry around outsourcing
• 15:44
  tasks to people. And the next 50 years will be
• 15:46
  about outsourcing tasks to intelligence.
• 15:48
  So that leaves one question for you to consider.
• 15:50
  If these IT giants are successful and they actually
• 15:52
  build autonomous agents that can handle complex tasks end to end,
• 15:57
  does the outsourcing industry eventually just become the
• 15:59
  software management industry? At what point does the human
• 16:02
  element of IT services disappear entirely?
• 16:06
  The answer to that might come a lot sooner than we think.
• 16:09
  Thanks for joining us for this update.
• 16:10
  Catch you next time.