Overview
Opinions across the business world differ about the current state of enterprise AI adoption and how alarmed organizations should be—if at all—by signs of somewhat underwhelming returns on adoption initiatives thus far. As organizations prepare to double down on AI next year, though, it's clear that many will also be looking to get more disciplined with where and how they try to use it. And I think that organizations can separate themselves by thinking with clear, precise eyes about delegation—inviting employees into an honest, ongoing conversation about what's worth doing with AI in the present moment and what’s not.
What is delegation?
Anthropic's online training, AI Fluency: Framework & Foundations, provides a "4D" framework to frame the challenge of helping people collaborate with AI systems. Though the course is tailored more for individual adoption, I find the framework a helpful one to think about the general dimensions of AI adoption. They are:
- Delegation: Thoughtfully deciding what work to do with AI vs. doing yourself
- Description: Communicating clearly with AI systems
- Discernment: Evaluating AI outputs and behavior with a critical eye
- Diligence: Ensuring you interact with AI responsibly
Thinking about these Ds, it seems to me that Delegation, as a category, does not receive the conversational or strategic weight it deserves. There is plenty of focus on LinkedIn and within organizations on the skill of prompting. The same goes for challenges of quality assurance and ethics. When it comes to questions of how and where teams can best use AI today, however, I think there's a lot more room for precise, practical thinking.
We are still in the very early years of enterprise AI, and there are realities to what the systems can do well today. Recognizing these realities does not make you a skeptic. As Demis Hassabis, leader of Google DeepMind, said it in a June interview:
I think there are missing capabilities right now that all of that have used the latest LLMs and chatbots will know very well, like on reasoning, on planning, on memory—I don't think today's systems can do, you know, true invention, true creativity, hypothesize new scientific theories. They're extremely useful, they're impressive, but they have holes … Performance across the board isn't consistent enough.
— Demis Hassabis, Google DeepMind
Between a Rock and a Hard Project
Leaders working to drive team- and enterprise-wide AI adoption initiatives are caught in a difficult wedge. On one side they have the people they lead, those who will have to reckon with the day-to-day limits Hassabis alludes to. On the other side they have the talk of grandiose use cases that swirls across LinkedIn and podcasts, and the pressure from boards on down to show signs of transformative adoption. It's easy to see how, caught in this wedge, precise, practical questions of where and how you’re going to use AI can get overlooked. But the first group, the employees, will disengage if they’re asked to adopt technology that doesn’t do what was promised.
AI can still, of course, be immensely useful for companies today. You just don’t get there by pretending these limits don't exist, or ignoring employees who try to point them out. The smartest companies will make space to talk about delegation—about what you can really do with today's systems and what's worth the effort currently required. They’ll bring in employees directly using the technology both at the start of and through initiatives to help guide their thinking.
These conversations should focus a lot on complexity and memory requirements. Though reasoning capabilities are advancing, current AI systems struggle with highly complex, ambiguous scenarios and excel at structured tasks with clear inputs and outputs: summarizing documents, drafting standard communications, extracting data from forms. And while AI can process large volumes of information, it has limited "working memory"; it frequently loses exact details or struggles to connect information scattered across extensive document sets. These constraints can often be worked around, but doing so typically requires additional human labor—reviewing outputs more carefully, feeding information in smaller chunks, culling data differently on the front end, providing more detailed prompts… you get the idea. The question isn't just what's technically possible. It’s also what's actually practical given the real human effort involved.
Looking Through the Lens of Chatbots
Consider the example of internal chatbots, a common focus of early adoption initiatives. Not all chatbot ideas are created equal, and precise delegation can make or break success. An AI assistant answering employee questions about leave policies from one clear, concise handbook can work well—the rules are structured, the information set is contained and the quality bar is straightforward for employees to define and train around. But deploy that same chatbot to field questions about benefits eligibility across multiple employee types, pulling from dense plan documents with countless exceptions and edge cases, and you're likely to see frustration mount quickly. The information set is too sprawling, the exceptions too numerous and the potential cost of inaccuracy too high.
This is just one example, of course, and the specific questions an organization should ask will vary with context. The most important thing may be simply creating the space to talk about and think through these details as they relate to your own workflows and challenges. This means engaging the people who will really get their hands on the work each day—inviting them to share about the ways they work, what uses they would find most useful and how things are going with the initiatives so far.
Because in addition to more direct return on investment, that’s what all this is about—making sure employees are being empowered and brought along by AI. When people are pushed to use technology that can’t do what was promised, disengagement follows. But done right, AI adoption can be a positive force for retention and culture. And United States companies in particular have an opportunity to improve, with international employees reporting they participate in activities like providing feedback, early testing and requesting specific features at least 10% more often than U.S. employees.
AI adoption is a race, but it’s set to be a long one, run in parallel with the development of the technology itself. Soon the organizations that make thoughtful delegation decisions—that are purposeful, practical and collaborative, bringing in people every step of the process—will start to pull ahead.
