Some thoughts on why Khosla’s prediction of the imminent demise of IT and BPO services won’t play out in 5 years

I read this yesterday https://www.hindustantimes.com/india-news/aisummit-2026-it-bpo-services-will-disappear-in-five-years-says-venture-capitalist-khosla-101771182147907.html

Vinod Khosla is a man I deeply admire, and hence when he says something I pay very close attention and spend some time thinking about it. He has seen computing evolve up close and is quite bold with his investments.

His underlying thesis for saying IT and BPO will go extinct is based on AI getting far superior to humans in the next few years. I agree with the rate of progress – with the kind of capital being burnt through, it is only fair to expect that the breath taking innovation in AI will progress .

That said – I don’t think his predictions will come true in the 5 year time line he is putting forward for a few reasons. I have ten things in mind that makes me believe that to be the case.

As usual, these are strictly my personal views and not of my present or past employers.

1. Not all BPO and IT deals are FTE based. FTE based model of course will get disrupted much like SaaS models based on seat based pricing is already getting existential threats. The difference is that unlike SaaS where a minority of companies have consumption and outcome based pricing, BPO and IT have quite a lot of outcome based pricing already. Khosla probably is referring to just the FTE based models which is just a subset of the industry. My thesis is that AI will just help IT and BPO providers shift even faster to a fully outcome based model which will be awesome for their clients. Business leaders largely care about guaranteed outcomes when the push comes to shove.
2. There is a lesson on agent proliferation that we should keep in mind from micro services. A decade ago or so, we fell in love with micro services and a lot of IT shops embraced their elegance. In a couple of years, we realised that it’s a real headache to deal with a lot of services and the scaffolding needed to operate them in an enterprise grade fashion took several more years. The basic infrastructure needed for a lot of agents running loose in an enterprise landscape is quite immature today and such platforms don’t happen overnight .
3. Enterprise inertia is a real thing. There is hardly a CIO I know in financial services who hasn’t told me about their goal of replacing mainframes and moving everything to public cloud. And yet – mainframes are still alive and thriving and most companies haven’t moved even half their workloads to public cloud. Change is quite hard in enterprises on all fronts – people, process and technology all usually have secondary and tertiary effects if changes happen quickly and hence corporate leaders tend to move deliberately. They won’t risk breaking things when moving fast. I am not saying that this is a good thing – trust me I have been frustrated all my career with this kind of inertia but I understand why senior leaders are careful
4. Enterprise buying models don’t change fast either. FTE models are considered favorites by most procurement teams because of the ease of managing such contracts – they are easier to negotiate, execute and monitor even if their value is less than outcome based contracts. It will take a lot to switch this behaviour to outcome based models in a mainstream way. Now imagine the challenge of moving this to a software license model !
5. Law making almost never keeps pace with innovation. Laws are written with human actors in mind. If a human accountant does tax fraud – they go to jail. There isn’t an equivalent way today to keep agents honest. The best solution we have is to have humans in the loop. Granted it doesn’t need every human to stay around – but many will be needed to keep AI compliant
6. Jevons paradox can’t be forgotten . For foreseeable future, there are plenty of use cases for AI and I don’t see Jevons paradox failing. So as AI becomes more and more efficient, companies will push even more use cases into production which will need even more humans to be around.
7. GDPR and DPDP type laws won’t allow seamless cross border autonomous workflows. Sovereign AI is important and is here to stay. That essentially means there will a bunch of cross border workflows that still need humans on either end to make it flow
8. BPO has a lot of last mile aspects that are not yet AI friendly. While a lot of the work is repetitive and easy to automate via AI, most enterprise workflows have a last mile part that needs soft skills to navigate the enterprise nuances. Maybe it’s possible to automate some of it over time by reimagining from scratch – but we are not taking about 5 years in this case
9. As long as LLM hallucinates, some human needs to stay in the loop. Yes you can reduce hallucination to some degree – but at its core LLM is autoregressive and unless a very different architecture emerge from research community, we can’t put too many things into production without humans in the loop. Reducing hallucinations is not cheap – the very big models have high inference cost. RAG type solutions have limitations and are expensive to maintain as enterprises evolve.
10. LLMs are static learner’s unlike humans . LLMs are not sample efficient like humans when it comes to learning – we don’t need thousands of cat pictures to know what a cat looks like. Once they are trained, they don’t learn on the job the way a human does. A human BPO agent can be told that they are wrong and need to do a task another way. The way to change an AI agent to act like that is not compute and memory efficient today. When we say an LLM remembered what we said earlier, what we mean is that a layer around it feeds the content of past conversations back to it behind the scenes every time which is quite inefficient and expensive. Short of fundamental research breakthroughs, we will need to keep coming up with better engineering hacks for efficiency.