AI-driven tools are proliferating—but still not where we want them to be. NinjaTech AI now offers a tool called Atlas, for example, which can schedule appointments, do web research and write code. Calendar scheduling seems to be a common task for these assistants, but I want more.

My day involves a little scheduling, jumping into Excel to update some spreadsheets with the latest data from statistical agencies, going into QuickBooks to create an invoice, heading into Constant Contact to send out my newsletter. In each of these examples, the stuff I’m doing is routine for me; what I do every month or week is what I’ve done over and over many times before. Some of these apps have productivity enhancements I have not bothered to learn, because I would have to invest a good bit of time to learn about one narrow tool. What I’d like is an AI assistant that can watch what I do and say, “Hey, I can do that next time you need it done.”

I shared my frustration with the CEO of NinjaTech AI, Babak Pahlavan. He described in an email four “parallel waves” needed for my kind of personal assistant. The first was self learning. Teaching the AI to win a game is relatively easy because of the well-defined rules. But finding the rules for human decisions is pretty hard. The AI needs to get ideas, try them out, and learn from them. Starting small, with narrow applications such as calendar management, will build technical skills to—hopefully—create the full-blown personal assistant.

Autonomy is Pahlavan’s second wave needed for better assistants. A co-pilot works alongside a person so the human can direct it and correct errors on the fly. A fully autonomous personal assistant needs to do things—and do them right—when no human is watching. Large language models are not there yet.

The third wave, skill set convergence, describes moving away from a huge set of very narrow tools to a small set of very broad tools. Every tool will have its own rules and style. Workers don’t want to learn how to employ multiple tools; they want to learn how to use one tool and have that tool do everything.

Finally, Pahlavan believes we need real-time streaming: “… true assistants are always-on, omnipresent and available across many channels, … available to iMessage, Slack, Call, FaceTime, Email etc. just like a human assistant.” Actually, the AI assistant will be more available than a human, so long as AI doesn’t need breaks for the bathroom, eating or sleeping.

On the optimistic side, Pahlavan believes that progress is coming very quickly, and that in a few years NinjaTech’s assistants will be self-learning and omnipresent.

I ran Pahlavan’s take on AI-powered assistants past a person who works at the cutting edge of AI (who prefers to go nameless). He wrote back, “The fundamental issue is the model makes too many mistakes. It's just not smart enough right now. I agree that there are other issues (Babak points out some), but they all seem approachable compared to the model just giving the wrong answer too often.”

Current AI models could watch what I do continuously and jump in to take over tasks, but “the model is going to do silly things way too often,” my contact wrote. This is the well-known hallucination problem. The models simply make stuff up, and sound very credible when they do. Vectara, an AI company, recently evaluated prominent AI models and found hallucination rates ranging from 3% to 27%. The problem is not simply, as often reported, because the AI is repeating falsehoods it found on the web. It makes up its own falsehoods regularly.

Top minds are working on the problem and will probably find a way to solve it. The solution could prove to be pretty expensive, but we don’t know yet. I’m sure they will eventually get accurate results, or pretty close.

The versatile and ubiquitous AI-powered assistant will make anyone who works with a computer more productive, as well as some who don’t even use a computer in their work today. A factory machine operator, for example, might say to the tablet, “I’m shutting down to calibrate the Number 3 cutting head.” And the tablet replies, “We’re three hours from a belt replacement; do you want to do that at the same time?”

AI will boost productivity for a wide range of workers, but we’re a few years away from that.