The headlines versus the trenches
"The SaaSpocalypse has begun." "The death of IT services is inevitable."
These framings make for good content. They're also largely written by people who have never had to integrate a 20-year-old ERP with a real-time AI agent while managing a governance framework, a change management programme, and a CFO asking about ROI.
A recent conversation between some of Silicon Valley's most experienced enterprise technology thinkers — on the a16z podcast — cut through the noise in a way I rarely hear. Not pessimistic about AI, but honest about where the actual work lives. Here's what resonated.
"A robot entered an elevator. But it couldn't press the button. So the same company built a separate device just to push the button. Why? Because there is no headless version of the elevator. That is the enterprise AI problem."
1. Silicon Valley and enterprise reality are not the same conversation
Silicon Valley runs on modern tooling, real-time debugging, and flexible architectures built by technical teams for technical users.
Enterprise runs on something different entirely: less technical users across large distributed workforces, fragmented data accumulated over decades, legacy systems that were never designed to be integrated, and governance debt that predates most of the people now tasked with managing it.
Add to that the pace of AI change — where the architecture you choose today may be obsolete in a quarter — and the complexity compounds. The enterprise isn't slow because it's incompetent. It's navigating constraints that Silicon Valley's default assumptions simply don't account for.
2. Agents hit the same integration wall humans do
Any enterprise older than ten years is a collection of systems waiting to be integrated. AI doesn't fix that. It inherits it.
The integration wall that stops human workflows stops agents too. If your procurement system can't talk to your ERP in real time, your procurement agent can't either. If your customer data is split across three CRMs with no single source of truth, your customer-facing agent will reflect that fragmentation in its outputs.
There is no shortcut through the integration problem. Agents are only as connected as the systems underneath them.
3. The access control gap nobody is designing for
Agents that inherit human permissions get blocked at every governance checkpoint that was designed for human workflows. Agents that bypass permissions create security and compliance nightmares.
The gap between these two failure modes is where most enterprise agent deployments get stuck. The answer isn't to give agents more access or fewer restrictions. It's to build a governed layer — where every agent is authenticated, routed, and given precisely what it needs to do its job, within boundaries the business has explicitly defined.
Agents that silently fail because they don't know who to call are not an AI problem. They're an architecture problem.
4. Stop fusing AI into software. Make software consumable by agents.
The framing shift that matters: instead of thinking about AI as software to be integrated, think of it as a user — one that needs identity, permissions, and context to operate.
Give your systems an agent-readable interface. Let agents access your enterprise the way an employee does — authenticated, contextualised, operating within defined boundaries. The enterprises that rebuild their systems with this in mind now are creating infrastructure. The ones that don't are creating legacy.
5. Velocity without governance creates entropy
AI raises the bar for the engineers and teams using it. It does not lower the need for them.
The vibe coding trap — moving fast with AI assistance, shipping quickly, deferring quality — produces short-term velocity and long-term technical debt at a scale humans couldn't generate alone. You're not just creating problems faster. You're creating them at AI speed.
Governance isn't the enemy of velocity. It's what makes velocity sustainable. The enterprises that understand this build faster over time. The ones that don't accelerate into a wall.
6. The scale reality
Here is the most honest thing said in that conversation, and the one most worth repeating:
A large enterprise is going to have to go through change management, systems implementation, and technology integration for agents to actually work. There is no shortcut. You actually do need to do the work.
This is not a counsel of despair. It's the clearest possible articulation of where durable business value gets built. The companies that help enterprises do this unglamorous, essential work — integration, governance, change management, architecture — will build the most defensible businesses in the AI era.
Deepak's Take
The real state of enterprise AI is not the version that gets written about in technology media. It's messier, slower, and more structurally complex than any demo suggests.
The CIOs, CDIOs, and enterprise architects navigating this every day are doing the hardest and most consequential work in technology right now. Not because the problems are unsolvable — they're not. But because solving them requires a depth of understanding about how large organisations actually work that can't be acquired from the outside.
The messy middle is where transformation happens. It's also where the most interesting problems live.
Deepak Choithramani is Co-Founder of Vishleshan AI Solutions. He writes about enterprise AI, agentic systems, and what it actually takes to go from pilot to production.
Follow Deepak on LinkedIn