Nirmal Mukhi is the Chief Architect at ASAPP, and the author of numerous research papers and articles about AI.gettyOver the last few years, the prevailing technology industry narrative has been that we'll soon all have personal AI assistants that will become indispensable in our daily lives.​Offering far more than the ability to compose emails, your assistant will coordinate your calendar, plan a family trip, handle every aspect of your tax return and make the annoying parts of your life go away. Your assistant will become your trusted delegate, your very own version of Tony Stark’s Jarvis for organizing your life. ​If the destination is so obvious, why aren't we there yet? And what has to happen before we are? Getting AI to work well is a far deeper problem than raw reasoning intelligence, so the answer isn't "the models need to get smarter."​ Instead, I want to look at three problems that receive less attention but stand between the assistant we actually want and use.​Even an AGI-level agent will fail without the right environment. Most people assume that once models are intelligent enough, they can do anything. In reality, frontier models already are. Today, agents excel at executing multi-step tasks, especially ones that take humans a few minutes to complete, according to research from METR. The models are also progressing at successfully completing increasingly longer tasks. However, while the models themselves work, they were not built for their current operating environment. This is akin to building Ferraris and asking them to drive on dirt roads. For anything beyond your personal sphere, your agent has to work with enterprises—and the service infrastructure they've built was all designed for humans. An agent can reason brilliantly and still be stranded in front of a login screen and a phone tree. Unlike humans, AI agents shouldn't be calling ​your cable provider and waiting on hold. Instead, the tools and procedures built for agents to use directly need to be developed.But enterprises modernize slowly. If exposing data and tools for AI takes as long as moving to the web or shipping a decent app did, we're looking at years of ingrained systems gumming up the works.​Agent collaboration is in its infancy. End-to-end automation will rarely live inside one agent. For example, the assistant on a phone may be owned by Apple, Google or OpenAI. That agent will need to negotiate with the retailer's agent that holds the account and knows the policies. When a personal assistant asks for a refund on a broken washing machine, the retailer's agent needs to check the warranty, apply the credit and schedule a pickup, maybe with a human signing off in the middle. That's multiple exchanges and a negotiation between two parties who don't trust each other by default. The protocols for it (MCP, A2A and others) are still being worked out. ​​For enterprise leaders, even if the technology is ready, the immediate concern is legal liability. When an autonomous agent executes a financial transaction, who owns the financial downside of a hallucination? When does "my agent authorized it" become a binding commitment? Just as liability questions delayed driverless cars, this uncertainty could hold back useful, ready-to-deploy assistants.​People are not ready to fully trust AI assistants. An assistant is only as good as the human using it, and right now, most people use these systems like search engines: terse queries, no context, no conversation. Remember learning to "speak Google" in the early 2000s? We'll go through that again, except this time the learning curve runs in both directions. The system has to learn us, too—our age, our language, our culture, how much trust we've extended, when to push and when to wait. Software has never had to do this before. In the past, we designed interfaces, and now we're designing a relationship—and we're barely past the first date. Adoption is what drives investment and improvement. The slope of that adoption curve is completely unknown and a huge risk.​ I personally believe these problems will get solved—eventually. So when does calling the cable company become a memory?Here's how I see this trajectory unfolding: 1. The Current Lane: The assistant on your phone stays in its lane—looking up information and conveying it conversationally, managing email, playing music, handling your own data. 2. The Transition: This is when end-to-end transactions start to take real shape, as collaboration standards harden, legal frameworks get worked out and enterprises rebuild their systems for agents instead of humans. 3. The Autonomous Benchmark: This will be when you never have to call your bank, your airline or your insurer again (unless you need to resolve a dispute between the AI agents).​So, given that agentic development is heading in this direction, what can enterprise leaders do to prepare? ​• Invest in your agentic ecosystem, ahead of self-service on the web or your app. Human use of those channels to work with your brand will likely peak soon.• Expose the enterprise via APIs. Treat APIs as the primary interface for agent requests (via frameworks like MCP), not web wrappers for human eyes. Those who architect for machine consumption today will own the customer experience tomorrow.• Review corporate insurance and legal policies. What do they say about AI errors and associated liability? These questions also apply to your AI partners and your agreements with them.• Understand your customers, including that they'll soon be represented by their own agents. Design conversations and a brand personality that solve their problems and build trust through safety, monitoring and fast feedback, so they come to prefer your AI for its competence.​All this matters more than it sounds. Ask around and you'll struggle to find anyone who enjoys calling customer service. For many, the call is more stressful than the problem that prompted it. Don't think of this as automating away a convenience but as automating away the small, daily frustrations of modern life. For the sake of my own sanity, I hope the autonomous future arrives early.​​​​​Forbes Technology Council is an invitation-only community for world-class CIOs, CTOs and technology executives. Do I qualify?