Meaningmaking at work.
To build successful AI applications, product management must comprehend meaningmaking.
Hello friends,
tl;dr: AI applications fail when they're not designed to leave meaningmaking — subjective decision-making about relative value — to humans. To build successful AI applications to replace existing business processes, product management must understand what meaningmaking is, unbundle meaningmaking work from non-meaningmaking work, and reflect this unbundling in how products are conceptualized.
Application failure.
I went to Chicago in May to run a workshop on improving the speed and quality of new product development. A few weeks before I got there, I pinged some of the natives to see if they would be free to get together. One of my friends there was testing a new AI application: A scheduler tool that would replace the human work needed to arrange a mutually convenient time and place to meet. (AI application = a deployment of AI technology to solving a problem, which could be an app in the sense of a mobile or desktop app, but doesn’t have to be.)
The AI scheduler tool worked like this: We added the scheduler tool (as an email address) to our email conversation. The tool responded and queried both of us separately about our preferences for times/dates and connected to our calendars so it could schedule the meeting directly.
The AI scheduler failed spectacularly. We both sent a few rounds of emails to the tool about our slot preferences and connected our calendars. As near as I can tell, the schedule tool seriously misinterpreted the meaning and importance of several of my calendar entries. It eventually scheduled a meeting for a day when neither of us would be in Chicago.
The seductive sell here is that the LLMs and other AI systems driving the tool purport to allow it to understand our email conversation, the meeting context, and our existing calendar entries so it can converge on an optimal slot — like a good human assistant would be able to do. The sell is seductive because we really want AI applications that can do everything a human does and AI seems able to do this because it can produce outputs that look like human outputs — but actually AI applications cannot do the meaningmaking work that humans do all the time.
I’ll explain what meaningmaking is below, but I’ve also written previously about why meaningmaking is what distinguishes humans from machines, why AI systems can’t do meaningmaking at all, and the seductiveness of AI’s apparent ability to do meaningmaking.
The failure mode for this AI scheduler tool highlights the importance, the difficulty, and the invisibility of meaningmaking in the work that AI applications are supposed to replace. It also highlights the universal absence of consideration for meaningmaking in the design and conceptualization of AI applications.
This is a big problem and a huge opportunity when building AI applications to replace human work in existing business processes. There’s growing theory and evidence that AI will change how work is done — e.g., this study of knowledge workers using LLMs — and that the effect is likely to be big (e.g., this model of how LLMs could affect labor markets.
Meaningmaking is a special type of sensemaking.
Right now, there’s a lot of investment in building AI systems that do better sensemaking, which is interpreting the meaning of text (or images or sounds or whatever). The type of sensemaking that’s getting attention is the type that lets an AI system identify a thing. This is the type of sensemaking that allows an image-generating AI system to produce an image that a human would recognize as being “jars” or “a painting in the style of Morandi of assorted earthenware jars on a tabletop.”
Meaningmaking is a specific type of sensemaking that is not about identification. Meaningmaking is the act of making subjective decisions about the relative value of things. Thing-identification is important, but meaningmaking comes after and goes well beyond identification.
There are four types of meaningmaking:
Meaningmaking can be trivial and quotidian, as when I decide that I prefer grapefruit to peaches when doing the household fruit order. Meaningmaking can also be extremely material, as when a CEO decides to commit her company’s public reputation and a lot of money and time to developing a controversial and highly uncertain new product category.
The main characteristic of meaningmaking is that it involves decisions that have no absolute or objective verifiability. In consequence, some humans do meaningmaking more mindfully and/or better than others. To be an artist, a judge, a poet, an activist, or a technology innovator depends on meaningmaking — to be successful at any of those things requires mindful and sophisticated meaningmaking. But choosing to take any action requires meaningmaking, so all humans do meaningmaking work all the time.
Though humans do meaningmaking all the time, AI systems cannot do meaningmaking at all. I call this AI’s meaningmaking problem. A good way to think of it is that meaningmaking is what distinguishes humans from machines, and humans must always do the meaningmaking in building or using AI systems (as you can see below).
Meaningmaking at work.
Businesses rely on meaningmaking work all the time.
In an earlier essay, I gave some examples of humans doing meaningmaking work as part of business-critical processes: “An underwriter deciding whether to insure a building project using a new construction method, an entrepreneur choosing what product to focus her startup on, an investment committee structuring an investment-for-equity deal with a startup, a panel of judges ruling on the interpretation of law in a crucial case.”
These all depend on the trained humans involved in the process (the underwriter, the entrepreneur, the IC, the judges) making inherently subjective decisions about the relative value of a thing. There are no a priori objectively correct valuations — whether the thing is the potential liability from an untested construction method, or the potential upside of a startup’s idea. These are judgment calls, and the judgment calls always represent meaningmaking work.
But most everyday business processes also rely on meaningmaking work. These processes are not so obviously business-critical but are nonetheless vitally important. Some examples of humans doing seemingly trivial meaningmaking work include: A customer service representative using their discretion to give an unhappy customer a full refund though the product is working as expected, a production line worker deciding that an assembled part should be rebuilt though it technically passes the inspection criteria, or an executive assistant deciding to reschedule a long-standing internal meeting so the CEO can take a last-minute call with a potential investor.
This is vital meaningmaking work that businesses do “without even thinking about it.” Humans do a lot of meaningmaking work as part of existing business processes, and this meaningmaking work is important but currently mostly invisible.
To build AI applications that can replace existing business processes, we need to understand clearly where and what kind of meaningmaking happens in those processes, then build that understanding into the products we make to replace those processes.
Product that comprehends meaningmaking.
The AI scheduler tool we tried to use in Chicago failed because meaningmaking wasn’t an explicit part of thinking about what the product is and how it is meant to work. The product was built to be an AI system that could replace everything a human assistant does when scheduling meetings
.
The AI scheduler would have looked very different and probably worked much better if the product thinking began instead from investigating what meaningmaking work humans do to schedule a meeting (i.e., decide on the relative value of different slots and existing obligations in a calendar), and then imagining what an AI system can do better than humans to make it easier/faster for humans to do the meaningmaking work to schedule a meeting
. (Here’s how to think about what AI systems can do better than humans.)
The latter approach throws up possibilities for features and UX that trying to build a simplistic human-replacer does not. For instance, a feature I would love in a scheduler application is automated tagging with human validation of calendar entries that conform to user-defined quality criteria
. Examples of these criteria might be: “This meeting was the right length to deal with the agenda,” or “There was enough time between this meeting and the previous meeting for comfortable preparation/travel.”
The application leaves the meaningmaking work of deciding on the “right”-ness of meeting duration and the “enough”-ness of pre-meeting duration to the human user, because both are subjective measures which only a human can decide. This feature would enable the scheduler application to (for example) gradually become more adept at scanning emails for agenda information and scanning calendars, to propose to the human user a range of slots of the “right” duration and which provide “enough” pre-meeting time for travel/prep. This reimagined scheduler application does what humans are relatively bad at doing (rule-following and analyzing lots of data) but explicitly leaves all the meaningmaking work to the human user.
This latter approach to building products is product management that comprehends meaningmaking.
Product management that comprehends meaningmaking.
Product management is systematic, mindful, institutionalized thinking about what products to build and how to build them. I got my exposure to product management in Google’s Product Team, but there are many flavors of product management. What connects all these flavors is that each of them is a framework for thinking about who a product is being built for, what features the product should have, how the product will get built, how it will be resourced, what the business model will look like, etc. But none of these flavors of product management comprehends meaningmaking — they are not sensitized to meaningmaking, nor do they include meaningmaking in how they think about what products to build and how to build them.
This is a problem and an opportunity for product management.
It’s a problem because products that don’t comprehend meaningmaking won’t actually work if meaningmaking is an important part of the work they do. The AI scheduler tool is an example of such a failure mode. This is also why the promised potential of AI systems for making business processes more efficient and effective remains unfulfilled. (Worse, it is why AI applications can make business processes less effective, e.g., when applied to hiring and recruitment.)
But it’s more interesting and productive to think about how meaningmaking is a product management opportunity.
Opportunity in the great unbundling.
Meaningmaking is an opportunity for product management because this approach leads to new ways of thinking about what products can be and how humans interact with them — and especially because no one seems to be doing it systematically yet.
The scope of opportunity is especially enormous when thinking about products that are AI applications intended to replace existing business processes. Nearly every process in every existing business which involves humans bundles meaningmaking work with non-meaningmaking work. And, as Sarah Tavel points out, the way to be successful in selling products to existing businesses is to sell work, not software — by which she means that an AI application is more likely to be sellable if it actually does work that businesses need done. This requires understanding what work is actually being done in the business process the AI application is designed to replace.
Until just a few years ago, we had no good reason to spend any effort understanding work through the lens of meaningmaking. Now, things are different. AI is a plausible candidate for a general purpose set of technologies that seem able to do everything humans can do but actually can’t do the meaningmaking work that humans do all the time but often don’t even realize they are doing.
Suddenly, separating what humans must do from what machines can do better is vital. If humans must do meaningmaking work and machines cannot do meaningmaking work at all, then the key insight is that work must be unbundled into meaningmaking work and non-meaningmaking work. Those who don’t recognize that this unbundling is essential for building good AI applications have been fooled by what I call AI’s seductive mirage.
When a technology like AI appears that prompts this fundamental rethinking of how existing work should be done — previous examples include printing presses leading to a rethinking of handscribed books, or repeatable manufacturing processes prompting the rethinking of piecework production — those who invest in doing that rethinking end up being the value creators.
Unbundling meaningmaking work from non-meaningmaking work is a fundamentally new approach to thinking about what work is, and what work could be. Value creation will come from opening up new avenues for thinking about what AI applications could look like, what kinds of functionality could be useful, how users could interact with AI applications, and what kinds of business models could make sense.
Paying attention to meaningmaking when building AI applications is also a better way to think about alignment in AI deployments. Value is an inherently subjective thing, so the best way to deploy AI so that it is aligned with human values is to build AI applications that are intentionally designed to let their human users make all the subjective value decisions (i.e., do all the meaningmaking work) while the AI applications support their human users by doing the other non-meaningmaking work that machines are better at doing.
Product management which comprehends meaningmaking will be able to build products which allow “AI systems and humans to work interdependently but doing fundamentally different things which they respectively do well.”
What does this kind of product management look like? I have some thoughts, but that’s for next time.
See you here soon,
VT
You're definitely getting at something important here – the need for AI product design to recognize that AIs are (at least for now) better suited for some tasks than others, and good design calls for refactoring the workflow so that the AI can do the things it's good at and not try to do the other things.
But I don't know that it's entirely hopeless for AI systems to ever tackle any of the things you're describing as "meaningmaking"?
First of all, some of the tasks that you're saying are subjective, I think are actually objective in principle, they're just too difficult to evaluate rigorously. For instance:
> These all depend on the trained humans involved in the process (the underwriter, the entrepreneur, the IC, the judges) making inherently subjective decisions about the relative value of a thing. There are no *a priori* objectively correct valuations — whether the thing is the potential liability from an untested construction method, or the potential upside of a startup’s idea. These are judgment calls, and the judgment calls always represent meaningmaking work.
Evaluating liability or startup upside is *very difficult* but I'm not sure it's *subjective*? If I think a startup is worth $50M and you think it is worth $80M, is that because we have different subjective tastes and values, or because one of us did a better job of predicting future outcomes than the other?
(Weather forecasting is an example of something very difficult / fuzzy that deep-learning models are becoming quite good at)
And even for truly subjective choices, it seems like LLMs ought to be able to do a pretty good job of modeling human preferences; there is plenty of relevant material in their training data. I don't have a link handy, but there have been studies showing that when done correctly, it's possible to replace political surveys ("what do you think of this policy proposal") with LLM queries and get results that closely match human responses.
So perhaps, when models make poor choices, it's actually because they're lacking context that is important to the specific situation? If so, then part of the product design challenge will be figuring out how to get that context and present it to the LLMs at the right times.
Wonderful series - I’m really find this to be a helpful lens for thinking about the challenge of getting real value from the advances in AI. However, the “four types of meaningmaking” feels a bit off to me. I was expecting you to anchor the types of meaningmaking to the different types of not-knowing with something like the following:
- Actions: Which actions are worth doing/acceptable in a given situation?
- Outcomes: Which outcomes (and the interconnected outcomes) are desirable/tolerable for the stakeholders of interest?
- Causes: Who is responsible for the outcomes resulting from the environment and a given set of actions?
- Values: What are the acceptable trade offs between different proxies of value?
This framing then gives AI (and tools in general) a clear goal - reduce not-knowing in one or more of these areas in order to make it easier for humans to decide the above questions in the course of their lives.