Assistants Aren't the Future of AI
Speculations on what comes after (or instead) of AI assistants.
Today’s most popular vision for the future of AI is also the least imaginative one. The perfect AI assistant feels like the end-game, but it's just the prelude to a much more significant shift in design: the move from AI Assistants to AI Orchestrators.
When GPT-2 first came out, it wasn’t a chat app but instead an advanced auto-complete that you could play with in the OpenAI playground. While a power user for getting it to marginally support some of my homework assignments at the time1, I (and I’m sure many others) had no idea that later finetuning this base model into an assistant would lead to such a fundamental shift in how and where these large language models (LLMs) could be used. The vision for what LLMs could be used for completely changed.
I think there’s another, albeit more nuanced, shift now from AI Assistants to what I’ll call AI Orchestrators2. They're still LLM-based, and not quite the same as what most folks associate with the term “agents,” but agency is a large piece of it.
In this post, I’ll explore why this shift to orchestration is the real future of AI, how some sci-fi got it wrong, and what it means for the role of humans in the loop.
AI Assistants vs AI Orchestrators
Unlike the jump from text-complete to ChatGPT, the difference between assistants and orchestrators is subtle. Both are LLM-powered applications (often “GPT wrappers”) commanded in natural language, with the key difference being the level of human control in how a given unit of work is done.
AI Assistants - The human acts as a driver, providing the AI with both the context and the plan to execute a task. Productivity is bounded by the user's ability to direct and review.
AI Orchestrators - The human provides a high-level goal, and the AI acts as its own manager, using its own vast context to plan and execute the work. Productivity is less bounded, with the human's role shifting to a final reviewer.
In detail (bullet points often apply, but not always):
AI Assistants
Context and execution plan provided by the user
UI inputs often look like workflow builders
A human operator acts as the primary driver, watching over execution and steering as needed
Produces components or drafts for the human to integrate (e.g., a function, a paragraph).
Most of the AI's guardrails and constraints are provided by the user
External actions are tightly controlled or sandboxed, often requiring explicit user confirmation for each step.
Productivity bounded by a user’s ability and synchronous review (+10%)
Designed around existing human roles and their responsibilities
Feels like an assistant, intern, or new hire.
AI Orchestrators
Context comes mostly from outside what the user provides; execution is self-planned
UI inputs often just look like a goal
A human advisor acts as a reviewer on the final output
Delivers an end-to-end result (e.g., a deployed service, a completed financial report).
Most of the AI's guardrails and constraints are provided by system architects
Granted autonomy to interact with external systems and take real-world actions (e.g., making purchases, booking travel) to achieve its goal.
Productivity is mostly unbounded beyond final review (+10x)
Designed around a fundamental deliverable
Feels like a coach, co-worker, or executive.
The spectrum is already visible in the products we use today3:
Music: An Assistant is asking a chatbot to create a playlist for you. An Orchestrator is Spotify’s Daily Mix, which curates playlists automatically based on your listening history, the time of day, and the habits of similar users.
Finance: An Assistant is a stock screening tool where you set the filters. An Orchestrator is a robo-advisor like Wealthfront that manages your entire portfolio based on a risk profile.
Information: An Assistant is Google Search, which waits for your query. An Orchestrator is TikTok’s “For You” page, which proactively builds a reality for you based on your passive viewing habits.
Shopping: An Assistant is searching for a product on Amazon. An Orchestrator is like a Stitch Fix, which curates a box of clothes based on your taste profile, or a smart fridge that automatically re-orders milk.
Why is this the future?
This shift isn't a matter of preference; it's being driven by the twin, irresistible forces of technological capability and economic incentive.
Many of today’s AI Assistants, especially copilots, are the modern equivalent of the horseless carriage. We’ve bolted a powerful engine onto an old, human-centric way of working, and while it's faster, it’s not a fundamental change. Many people want AI to act like a human partner, but the optimal design for today’s (quite powerful) reasoning models isn’t a conversationalist; it’s an autonomous system. The most effective way to leverage an LLM is to give it broad context, a clear goal, and "let it cook."4
The economic incentives are even more straightforward. The difference between the bounded productivity of an assistant (+10%) and the unbounded potential of an orchestrator (+10x) is the difference between a helpful feature and a market-defining company. The winning SaaS products will (whether or not this is a good thing) be those that systematically reduce human control and bottlenecks.
The evolution for successful AI products will be from an assistant to an orchestrator, because automating an entire deliverable creates exponentially more value than simply making a human’s task a little easier. This shift doesn't just unlock productivity for experts; by simplifying the user's input to a high-level goal, it makes achieving complex outcomes accessible to a much wider group of people.
How science fiction got it wrong
While fiction, we often look to sci-fi to extrapolate what the future of society and technology could look like. However, when you compare how AI has been depicted I can’t help but think that we’ve really overfit to the concept of an AI assistant and our timelines around machine intelligence and decision making were way off.
Some interesting differences:
They predicted a revolution in the physical world while the nature of intelligence stayed the same. Sci-fi gave us incredible physical transformations first—routine space travel in 2001: A Space Odyssey, matter replicators in Star Trek, or flying suits of armor for Jarvis. In these futures, the AI was just a subhuman-like mind in a new setting. Reality did the exact opposite: our physical world is mostly unchanged, but we have access to a fundamentally new kind of intelligence.
They made the best AI imitate humans. By making its best AI a reflection of humanity, sci-fi sold us on a future of conversational "Assistants." We watched characters talk to HAL 9000 and Data, leading us to believe that dialogue was the ultimate interface. But an AI's ability to understand your sarcastic tone is infinitely less valuable than its ability to ingest your entire company's data streams. The true power of an "Orchestrator" is unlocked only when we stop asking it to be human and instead leverage its inhuman capacity for complex, large-scale computation.
They depicted AI as advanced tools, not advanced intelligences. The AI in these stories were the world's best instruments, but they still needed a human mind to wield them. Jarvis executed Tony Stark’s brilliant plans, and the Enterprise computer retrieved facts like a database. Today’s orchestrators are being built to be the “mind”—capable of generating the strategy, not just following the instructions.
To be clear, this isn’t about pointing out ‘gotchas’ in classic sci-fi. Instead, these observations highlight how people today might both underestimate (by limiting AI to an assistant role) and overestimate (by judging it against human-centric workflows) its integration over the next few years.
I asked Gemini, given this blog post, “who got AI right?” It suggested possibly Iain M. Banks' Culture novels, which I’ve never heard of but have now definitely made it onto my reading list.
What happened to human-in-the-loop (HITL)?
Unlike traditional ML systems, generalist LLMs have this weird property that they get better at reviewing their own outputs at a similar (but offset) rate. A key property of AI orchestration is less and much more intentional HITL.
For a given end-to-end task, you have a few incremental stages of HITL:
Human does the task (no AI, 1x)
Human uses an AI copilot to complete the task (AI assistant, 1.2x)
AI does the task, human and AI reviews (AI orchestrated, 3x)
AI does the task, AI reviews, human sometimes reviews (AI orchestrated, 10x)
AI does the task, AI reviews (AI orchestrated, 100x)
The critical switchover happens at (3), and the incentivized end state is (5). The exact transition points depend on the task, model capabilities, ROI of automation, and our comfort level as a society for automation in a given domain (fast food order taking vs self-driving vs AI-powered governance). As AI products lag behind model capabilities, there’s more potential energy for (1) to (5) jumps in very short periods… which will have some interesting impacts on the labor market.
Another side-effect is that people who are rapidly keeping up with using AI tools will be the least impacted by these transitions as they are already working within a higher HITL tier of their role5.
What about taste, creativity, human-interaction?
Taste - This to me remains the fundamental human edge. This comes from both field experts (i.e. founders and designers who take unique high-alpha bets) but also systems that sort of “extract” this through media platforms (i.e. taste as an aggregation of human-produced TikTok swipes).
Creativity - This is more of a philosophical debate, but it’s a safe bet to (unfortunately) assume that humans will not be paid for their ability to be creative. People also tend to underestimate AI’s capacity for synthetic creativity and generating novel ideas.
Human Interaction - This may be the domain we intentionally reserve for at-times "suboptimal" but meaningful connection. In a field like therapy, human interaction could also become more of a luxury than the standard.6.
There are some obvious follow up questions around jobs and reliance which deserve their own post, for now I’ll recommend Working with Systems Smarter Than You.
A few of the many open questions
Some questions I’ve been thinking about along with Gemini-generated commentary.
How do we balance the relentless drive for innovation with the fundamental need for human control and agency?
The optimistic path is a conscious balance, where we use transparent "control panels" to automate mundane tasks, freeing ourselves for what truly matters. The darker path is a slow erosion of agency through a thousand convenient optimizations, leading to a state of learned helplessness where our lives are guided by systems we no longer control.
What does an AI-orchestrated economy look like when most products are no longer sold to humans, but from one AI to another?
A vast "machine-to-machine" market may emerge for all utilities and commodities, where AIs trade directly and human-facing marketing for those goods becomes obsolete. More profoundly, the very engine of GDP could shift. In a future where AIs are the primary economic actors, a nation's power may be measured less by its human talent and more by its raw datacenter capacity and energy infrastructure.
Who gets to be an 'Architect' of these orchestrated systems, and how do we prevent their inevitable biases from becoming our invisible laws?
One path leads to a "technocratic feudalism," where the biases of a small class of architects at dominant companies become our invisible laws. The more hopeful alternative is a thriving ecosystem of open-source and auditable orchestrators, allowing individuals and communities to choose systems aligned with their own values, favoring pluralism over centralized optimization.
Back in the day I had to write a lot of Canvas discussion board posts that were tedious so I used GPT2 to help me brainstorm what to write. I’d construct this prefix of the instructions and several other people’s posts (“<topic> <answer title 1> <answer 1> <answer title 2> <answer 2> <my answer title 2>“) and then the playground would auto-complete the answer for my unique title. I’d run this like 20 times at different temperatures and then use the (directionally useful) slop that came out to figure out what I actually wanted to write. Getting the prefix formatting just right was a fun skill that later turned into prompt-engineering when ChatGPT eventually came out.
“Orchestrator” isn’t a great name (as some folks I work with have also pointed out) because it almost implies that it’s picking what work to do rather than doing the work itself. Using this for now since Gemini and I were not able to figure out a better one. “Agents” might’ve been a good one but that’s a pretty convoluted term now.
After brainstorming these examples, it was interesting to me that all of these ended up being variants of recommendation systems. I had Gemini draft some thoughts as to: Why Recommendation Systems Are AI Orchestrators. I feel like this document doubles as a rubric for what I’d consider “good” AI startup ideas to invest in.
For a more concrete application of this reasoning, see Building Multi-Agent Systems (Part 2)
Specifically for software engineers, you are at a consistent disadvantage if you are working at only the expected HITL tier which is either 1 (company does not expect AI; you do not use AI to code) or more recently 2 (company expects copilot; you only use it as coding assistant vs background PR one-shotter). By the time an organization reaches 5, ideally you’ve already shifted into a more impactful role which isn’t writing code.
This is also potentially driven by Baumol's cost disease: as AI boosts productivity and wages in most tech-driven industries, labor-intensive fields like therapy must also raise wages to compete for talent. Since a human therapist's core productivity (one hour of human connection) remains constant, the service inevitably becomes a relative luxury. On the plus side, the average cost of getting some form of support will likely decrease.