
What do we know so far from the gazillion POCs on GenAI from the last couple of years?
- Hallucination is a feature of LLM, not a bug . Thanks to the POCs – yes I know no one likes POCs and want very quick payback – we now know enough techniques to mitigate hallucination to use LLMs in enterprise workflows – like context restrictions (like RAG) , verification (CoVe, CoT etc) , specialisation (fine tuning, SLM etc), and of course human in the loop.
- Security is a big deal and probably under invested across the board
- Governance is boring and expensive and yet you can’t scale AI without it
- AI is still appears quite expensive and unpredictable for massive deployments – and much like how FinOps happened with public cloud scaling, AI needs to make friends with the office of the CFO
So what next ?
Duh … agentic AI .. what else !
GenAI proved that human workers have an awesome tool now to make them WAY more productive in what they already do. GenAI still is limited in the sense that it can’t act autonomously – so naturally our productivity obsessed world moved quickly to agentic solutions which can act independently to various degrees.
Will it work as advertised ?
The basic building blocks are all there – and will keep improving like technology has been progressing all these years. But that doesn’t translate to agents ruling the whole world in 2026 !
At its most basic level – agents need to talk to other agents and should be able to work with tools to get work done. Those protocols like A2A and MCP are already available and a quarter gazillion POCs have been done on those. We now have a better idea of what needs to get fixed to scale deployments.
For example, MCP standardized how agents talk to tools – which is enough for a POC. But in an enterprise deployment, you also need consistent handling of security. So now we use MCP gateways so that AI models don’t need to worry about raw credentials etc.
Another example – A2A standardized how an agent can talk to another, which is enough to do a POC. But any respectable enterprise workflow needs a lot of agents talking to each other – which leads to all kinds of orchestration overheads. What’s the point in handing off tasks back and forth without solving it? So now we use registries that can identify the best agent for a given task, and we set a limit on how many hops can happen before a human gets involved. In the same vein, we now have better approaches on scaling performance and handling security.
Even if all those things worked – we still can’t let agents do lose in enterprise workflows without having a solid audit solution in place. Similar to OpenTelemetry traces, now we can use observability headers to leave the trail to look back at which agents did what actions.
One of the best things that happened last year was that A2A and MCP both were donated to Linux foundation – which helps cross vendor collaboration that benefits everyone !
What is the bear case here?
The bull case is obvious – there is plenty of breathless commentary on that and I won’t rehash it 🙂
Having been in the tech industry for a while, I am sure we will see some high profile failures and the associated doomsday predictions. That happened for ERP, Mobile, Cloud and so on and I am sure AI will be no exception either.
If I were to predict – my top 3 reasons for AI project failures will be
- High costs – most likely because cost estimation models don’t scale from pilot to production
- Poor data quality
- Lack of clarity in business case
I also think there will be plenty of massive success stories from the companies who go about the deployments thoughtfully. But of course bad news sells more clicks !
Agentic AI deployment is about complex integration
One thing is clear – there won’t be “one big AI” to rule the enterprise like we see in sci-fi movies. The future is a complex system of highly specialized agents that will be hired and fired as needed for each task – and the primary challenge will be to make that integration work.
Winners and Losers in CIO offices and service providers
Agentic AI and associated technologies will disrupt “business as usual” significantly – that much is a given at this point. The question is who will win and who will lose.
Vast majority of IT work today – whether it’s a CIO team or a service provider – is some type of development activity. The default operating model is labor based with standardized processes. The skills levels are a mixed bag. Every year, there is some productivity improvement with better automation and tooling improvements- but those are incremental changes.
This is an easy area for LLMs and agents to make drastic changes and create massive productivity gains and low risk. The highly skilled engineers will always be needed and probably will get paid even more – but that is a modest subset of the labor pool. This is where CIOs will find the most budget to implement agentic AI.
So who will thrive?
In the very near term – technical proficiency might be the biggest differentiator. The initial deployments will have a lot of technical challenges like the ones I mentioned above and perhaps many more, and that will need really good engineering skills to mitigate. The use cases will generally be less ambitious till the underlying plumbing is in place.
But then it will change rapidly and the differentiator will be the process and industry knowledge – especially to deal with last mile problems that cannot be solved by great engineering alone.
Post script
AI – and agentic AI – is here to stay. We will probably vastly over estimate its impact in the short term and suffer the disappointment. We will also largely under estimate its impact over the long term. The popular commentary is around AGI and ASI and so on and that is all very worthy as future goals. We know by now that we don’t need AGI or ASI to have massive ROI in most enterprise use cases.
My hope – and prediction – is that several smart companies will start thoughtfully deploying agentic AI in production this year with realistic business cases justifying them.
Happy new year !












