AI Makes Context King: New Skills, Org Design, and Learning at Work

Prompting is out, context is in. 

Host Daniel Serfaty and Prof. Joseph Fuller of the Harvard Business School explore why “context engineering”—combining domain knowledge and judgment—is the next essential skill for human-AI teaming. They also discuss how organizational design, retention, and learning must evolve as technology advances faster than traditional upskilling can keep pace.

Transcript

Mini 3: AI Makes Context King: New Skills, Org Design, and Learning at Work

Daniel Serfaty: But talking about this notion that you expended on, replacement versus augmentation of jobs basically, and that's where most of the anxiety in society reside, I believe, that people are asking whether their job can be fully replaced as opposed to transformed by AI and therefore in a smart way. In balance, which job or CASC domains rather, and you started to outline them, those that are rule-based as opposed to knowledge-based perhaps, are more vulnerable to reduction and perhaps even elimination? Could you speculate on that?

Prof. Joseph Fuller: Yeah, so I think we can actually go a little beyond speculation. I think we're beginning to know certainly what we'd call routine cognitive work. Cognitive speaks for itself. Routine just means there's a limited amount of variability to the work I do cognitively. So let's go back to a credit analyst or let's say an accounts payable clerk. I can give the AI the rules that we want to preside over with accounts payable, paying our suppliers when they present a bill, "Is there a claim against the supplier? Do we have contractual obligations to the supplier? Are we trying to preserve cash in a certain geographic area or division or the corporation?" And the AI knows what the new rule is and exactly how to apply it instantaneously.

I don't have to send emails to someone and make sure they read it carefully and didn't misunderstand it or disagree with it even because they have a favorite supplier and they want to get them paid fast, even though under the rule, it might be delayed and then say, "Oh, I hadn't had time to process email." I'm not trying to suggest any malfeasance here. So this generation of AI is very good at those types of routine tasks, which by the way are often the tasks that workers in these jobs describe as the most onerous for them, the least interesting, the ones they sigh when they have to think about it.

"I'm going to take the sales report and the inventory report and write the monthly update for the controller of the division," where you can do a bot to let you write an excellent draft of that. And the bot would probably take an hour to build, and if you're spending three hours a month on that, it will give you a very good draft by it. Having access to the huge databases and your last 20 reports uploaded, you will want to edit it to make the writing flow better and to look for hallucinations. But that discouraging task that you dread every four Thursday of the month goes from a half day to an hour. So it will do a lot of the more burdensome, boring work.

It's going to get very good at multimodal integration. Gemini with multimodal, your audience should understand not only does that mean that it can create in multiple modes, it means now that video and audio is being tokenized and can be part of the context window. And so when you have tokenized video data and alphanumeric and you have context windows already in Gemini that accommodate 1 million tokens, it's a huge unlock into additional learning. Imagine the entire content of YouTube now being in the training set. So that's going to extend it much further.

So rules-based people tend to think financial rules, but we're seeing things like customer analysis where the AI can look at customer behaviors, all the data, create synthetic twins of customers or channels of non-customers and start testing whether or not changes in offer elements will improve market share or gross margin realization or customer satisfaction. It's less good at both in companies and in functions where the data is less clean or accessible. While it's excellent at legal work, it hasn't been very impactful yet in advanced manufacturing often because the physical layout and the way the data is gathered in multiple advanced manufacturing facilities within a single company are often different.

So there isn't as much rich data. It isn't as integratable, and of course, if you're talking about manufacturing, you've got the ISO 9000 mindset of, "I don't want to use a technology that's going to cause a lot of scrap or might even cause the process to get disrupted in a way that damages the capital equipment or something like that." So more caution there. So lots of green shoots of spring, but it tends to be concentrated in a few functions and a few industries right now.

Daniel Serfaty: We'll explore those industries in a few minutes, but you said a few things that I want to dig a little deeper. You mentioned this notion of the context window as opposed to the prompt. Most people, at least with LLMs and with the current way people understand AI at least, see themselves with a dialogue that has a little prompt window. You may enter a PDF or you may enter a sentence, a question or anything you want and then you get the answer. I like very much your generalization into, not just providing a stimulus, an input, we're also providing the context of that input. Can you expand a little bit on that? Let's [inaudible 00:24:57] into that, at least for many of people in our audience.

Prof. Joseph Fuller: So all the models have had what's called the context window, which is, "How many tokens," token being a bit of an image or a bit of a word, "that it remembers in the session?" When AI began to move to the now famous transformer approach where the prompt was not causing the AI to look for the next morphological word based on the previous word, but on the entirety of the context, it could remember, improvement was jaw-droppingly great. Currently, Gemini, the Google AI, has a million token context window. That's a lot of content and so it's getting ever more aware of what you are consistently interested in. And context windows will keep growing as compute becomes available and so we will get better and better, faster and faster resolution.

Now, you mentioned prompts, and for several years, people were focusing on the task of how to write good prompts and there was an emerging profession called prompt engineer. People really understood the models, how they worked, usually add a machine learning background, neural networking background, so they understood those principles. Now, I think that job is giving way to something that may be confusingly, I'm going to call a context engineer. So we're no longer talking about the context window, we're talking about the task. If you're using transformers and have a very rich context window, it's that much more important for the person interacting with the AI to really have domain knowledge.

Because context window, being richer and the process and getting deeper, means the ability for the AI to both generate an important insight that might only be recognized by someone who really knows the field or to create a very compelling story that's just wrong in its base, but that a generalist that doesn't have that rich understanding of the actual dynamics of the phenomena that you're probing may be fooled into accepting. I've always thought hallucination was a misnomer. It doesn't make up stuff that isn't there. It confabulates, it makes up a perfectly logical story based on what it's been trained on, which just happens to be fundamentally flawed.

Daniel Serfaty: But in a sense, it's very comforting, that last start, because you talk about the upheaval of expertise or some people are talking about the death of expertise, which is horrifying and we don't agree, but this notion of it's comforting because you are making in a sense of the implication of your argument is that, as the worker doing cognitive work interact with the AI, the expertise of that worker has a direct implication on the quality of the total work of the human and the AI working together because that worker will know how to put that context better and be able to evaluate what comes out of the AI better. Therefore, a new kind of expertise in a sense is needed, a context expertise. That's right.

Prof. Joseph Fuller: Yeah, exactly. I think that's a very nice description and the context would be everything from how this market has worked historically, how AI is playing into this market, how we in this enterprise do business, what we stand for. The AI may make a very rational decision on granting credit to a customer saying their financial statements indicate that a rule is being violated, but the context engineer may know that, "That customer," "That supplier," rather, "is working on a critical technology for our new product release. We have a 50-year history with them and that we have an archrival who's been trying to get their attention and get more of their engineering talent."

So the context engineer drawing on multiple ... It could just be an expression of corporate values, "There's a longstanding supplier that we value. They're not enjoying a healthy period. We're going to stand by them. That's what we stand for as a company." But I think also, in complex phenomenon, it often requires an expert to actually see something that's genuinely original, that's jumping out. And I think that's another thing that we don't have the term for. Maybe your audience could suggest one that we could all vote on it and I could pick one, but we're soon going to have a new type of error or problem or challenge in AI that it's suggesting something that is genuinely insightful, but the insight is so subtle and advanced that only a true master could detect it.

And this is, of course, detectable in the application of AI to the game of Go where AI won the tournament three consecutive games and in one game, my memory with move 232, but I may have that number wrong. It did something that every expert said was daft, that makes no sense. And then 100-move later, it all came together and it was a brilliant non-linearity. So, that's a encouraging sign and a warning sign that if you can integrate every piece of human knowledge on a topic, you're going to come up with all sorts of things that people haven't thought of before.

Daniel Serfaty: That's such a hopeful perspective, because I think that's very insightful, that notion of recognizing insight or recognizing a unique perspective as opposed to just attributing it to something, "Oh, it's weird," or "He didn't understand," because you are not in the middle of the curve. How do you recognize the three-sigma insight?

Prof. Joseph Fuller: Especially in a rules-based organization?

Daniel Serfaty: Yes. Exactly. That already points to particular sets of perhaps even new skills or new competence that one needs to know how to train in a particular industry or particular organization.