Friday, May 26, 2023

Chat-Based Development Changes the Build vs Buy Equation

The size of most markets is governed by a combination of supply and demand. Typically one or the other is limited: while there may be a bottomless appetite for chocolate, there is only so much cocoa in the world; conversely, while there is a near-infinite supply of Internet content, there are only so many hours available to consume it. The marketing technology industry has been a rare exception with few constraints on either factor. The supply of new products has grown as software-as-a-service, cloud platforms, low-code development tools, and other technical changes reduce development cost to something almost anyone with a business idea can afford, and widely available funding reduces barriers still further. Meanwhile, the non-stop expansion in marketing channels and techniques has created an endless demand for new systems. Further accelerating demand, the spread of digital marketing to nearly all industries has powered development of industry-specific versions of many product types.

Despite the visibility of these factors, the uninterrupted growth of martech has been accompanied almost from the start by predictions it will soon stop. This based less on any specific analysis than a fundamental sense that what goes up must come down, which is generally a safe bet. There’s also an almost esthetic judgement that a market so large is just too unruly, and that the growing number of systems used by each marketing department must indicate substantial waste and lost control.

One common metric cited as evidence for excess martech is utilization rate: Gartner, the high priests of rational tech management, reported last year that martech utilitzation rates fell from 58% in 2020 to 42% in 2022 and ranted in a recent press release that “The willingness to let the majority of their martech stack sit idle signifies a fundamental resource disconnect for CMOs. It’s difficult to imagine them leaving the same millions of dollars on the table for agencies or in-house resources. This trade-off of technology over people will not help marketing leaders accelerate out of the challenges a recession will bring.” They were especially incensed that their data showed CMOs are increasing martech’s share of the marketing budget, comparing them to “gamblers looking to write-off their losses with the next bet.” (They probably meant “recoup”, not “write-off” their losses.)

This isn’t just a Gartner obsession. Reports from Integrate, Wildfire, and Ascend2 also cite low utilization rates as evidence of martech overspending.

It aint necessarily so.

For one thing, underutilization is common in all departments, not just marketing.  Nexthink found half of all SaaS licenses are unused. Zylo put the average company-wide utilization at 56% and Productiv put it at 45%. (These studies measure app usage, not feature usage. But you can safely bet that feature usage rates are similarly low through the organization.)

More fundamentally, there’s no reason to expect people to use all the features of the products they buy. What fraction of Excel or Powerpoint features do you use? What’s important is finding a system with the features you need; if it has other features you don’t need, that’s really okay so long as you’re not paying extra or finding they get in your way.  Software vendors routinely add features required by a subset of their users. Since that helps them serve more needs for more clients, it’s a benefit, not a problem.

The real problem isn’t presence of features you don’t need, but the absence of features you do. That’s what pushes companies to buy new systems to fill in the gaps. As mentioned earlier, the great and on-going growth of the martech industry is due in good part to new technologies and channels creating new needs which existing systems don’t fill. That said, it's true that some purchases are unnecessary: buyers don’t always realize that a system they own offers a capability they need. And, since vendors add new features all the time, a specialist system may become redundant if the same features are added to a company’s primary system.

In both of those situations, avoiding unnecessary expense depends on marketers keeping themselves informed about what their current systems can do. This is certainly a problem: thanks to the miracle of SaaS, it’s often easier to buy a new system than the fully research the features of systems already in place. (Installing and integrating the new system will probably be harder, but that comes later.) So we do see reports of marketers trying to prune unnecessary systems from their stacks: for example, the previously-cited Integrate report found that 26% of marketers expected to shrink their stacks. Similarly, the CMO Council found 25% were planning to cut martech spend and Nielsen said 24% were planning martech reductions. (Before you sell all your martech stock, you should also know that each report found even more marketers were planning to increase their martech budgets: 32% for Integrate, 36% for CMO Council, and 56% for Nielsen.)

Let’s assume, if only for argument’s sake, that marketers are reasonably diligent about buying only products that close true gaps. New gaps will continue to appear so long as innovation continues, and there’s no reason to expect innovation will stop. So can we expect the growth in martech products will also continue indefinitely?

Until recently I would have said yes (unless budgets are severely crimped by recession). But the latest round of AI-based tools has me reconsidering. 

Specifically, I wonder whether marketers will close gaps by building their own applications with AI instead of buying applications from someone else. If so, industry expansion could halt.

Marketers have been using no-code technologies to build their own applications for some time. Some of these may have depressed demand for new martech products but I don’t think the impact has been substantial. That’s because no-code tools are usually constrained in some way: they’re an interface on top of an existing application (drag-and-drop journey builders), a personal productivity tool (Excel), or limited to a single function (Zapier for workflow automation). Building a complete application with such limited tools is either impossible or not worth the trouble.

The latest AI systems change this. Chat-based interfaces let users develop an application by describing what they want a system to do. This enables the resulting system to perform a much broader set of tasks than a drag-and-drop no-code interface. That said, the actual capabilities depend on what the model is trained to do. Today, it still takes considerable technical knowledge to include the right technical details in the instructions and to refine the result. But the AIs will quickly get better at working out those details for themselves, drawing on larger and more sophisticated information about how things should be done. Microsoft’s latest description of copilots and plug-ins points in this direction: “customers can use conversational language to create dataflows and data pipelines, generate code and entire functions, build machine learning models or visualize results.”

What’s important is that the conversational interface will drive a system that automatically employs professional-grade practices to ensure the resulting application is properly engineered, tested, deployed, and documented. In effect, it pairs the business user with a really smart developer who will properly execute what the user describes, let the user examine the results, and tweak the product until it meets her goals. Human developers who can do this are rare and expensive.  AI-based developers who can do this should soon be common and almost free.

This change overcomes the fundamental limitation of most user-built apps: they can only be deployed and maintained by the person who built them and are almost guaranteed to violate quality, security and privacy standards. This issue – let’s call it governance – has almost entirely blocked deployment of user-built systems as enterprise applications. Chat-built systems remove that barrier while fundamentally altering the economics of system development.  More concretely: building becomes cheaper so buying becomes less desirable.  This could significantly reduce the market for purchased software.

Anyone who has ever been involved in an enterprise development project will immediately recognize the flaw in this argument: there’s never just one user and most development time is actually spent getting the multiple stakeholders to agree on what the system should do. Agile methodologies mitigate these issues but don’t entirely eliminate them. 

Whether chat-driven development can overcome this barrier isn’t clear.  It will certainly speed some things up, which might enable teams to build and compare alternative versions when trying to reach a decision. But it might also enable independent users to create their own versions of an application, which would probably lead to even stiffer resistance to adopting someone else’s approach. 

One definite benefit should be that chat-based applications will learn to explain how they function in terms that humans can understand. Responsible tech managers will insist on this before they deploy systems those applications create.

In any event, I do believe that chat-built systems will make home-built software a viable alternative to packaged systems in a growing number of situations. This will especially apply to systems that fill small, specialized gaps created as new marketing technologies develop. Since filling these gaps has been a major factors behind the continuous growth of martech, industry growth may slow as a result.

Incidentally, we need a name for chat-based development. I’ll nominate “vo-code”, as shorthand for voice-based coding, since it fits in nicely with pro-code, low-code, and no-code. I could be talked into “robo-code” for the same reason.

Monday, May 15, 2023

Let's Stop Confusing LEGO Blocks with Computer Software

Can we retire Lego blocks as an analogy for API connections?  Apart from being a cliché that’s as old as the hills and tired as a worn-out shoe, it gives a false impression of how easy such connections are to make and manage.  

Very simply, all Lego block connectors are exactly the same, while API connections can vary greatly.  This means that while you can plug any Lego block into any other Lego block and know it will work correctly, you have to look very carefully at each API connector to see what data and functions it supports, and how these align with your particular requirements.   Often, getting the API to do what you need will involve configuration or even customization – a process that can be both painstaking and time consuming.  When was the last time you set parameters on a Lego block?

There’s a second way the analogy is misleading.  Lego blocks are truly interchangeable: if you have two blocks that are the same size and shape, they will do exactly the same thing (which is to say, nothing; they’re just solid pieces of plastic).  But no two software applications are exactly the same.  Even if they used the same API, they would have different internal functions, performance characteristics, and user interfaces.  Anyone who has tried to pick a WordPress plug-in or smartphone app knows the choice is never easy because there are so many different products.  Some research is always required, and the more important the application, the more important it is to be sure you select a product that meets your needs.  

This is why companies (and individuals) don’t constantly switch apps that do the same thing: there’s a substantial cost to researching a new choice and then learning how to use it.  So people stick with their existing solution even if they know better options are available.  Or, more precisely, they stick with their existing solution until the value gained from changing to a new one is higher than the cost of making a switch.  It turns out that’s often a pretty high bar to meet, especially because the limiting factor is the time of the person or people who have to make the switch, and very often those people have other, higher priority tasks to complete first.

I’ve made these points before, although they do bear repeating at a time when “composability” is offered as a brilliant new concept rather than a new label for micro-services or plain old modular design.  But the real reason I’m repeating them now is I’ve seen the Lego block analogy applied to software that users build for themselves with no-code tools or artificial intelligence.  The general argument is those technologies make it vastly easier to build software applications, and those applications can easily be connected to create new business processes.

The problem is, that’s only half right.  Yes, the new tools make it vastly easier for users to build their own applications.  But easily connected?  

Think of the granddaddy of all no-code tools, the computer spreadsheet.  An elaborate spreadsheet is an application in any meaningful sense of the term, and many people build wonderfully elaborate spreadsheets.  But those spreadsheets are personal tools: while they can be shared and even connected to form larger processes, there’s a severe limit to how far they can move beyond their creator before they’re used incorrectly, errors creep in, and small changes break the connection of one application to another.  

In fact, those problems apply to every application, regardless of who built it or what tool they used.  They can only be avoided if strict processes are in place to ensure documentation, train users, and control changes.  The problem is actually worse if it’s an AI-based application where the internal operations are hidden in a way that spreadsheet formulas are not.

And don’t forget that moving data across all those connections has costs of its own.  While the data movement costs for any single event can be tiny, they add up when you have thousands of connections and millions of events.  This report from Amazon Prime Video shows how they reduced costs by 90% by replacing a distributed microservices approach with a tightly integrated monolithic application.  Look here for more analysis.  Along related lines, this study found that half of “citizen developer” programs are unsuccessful (at least according to CIOs), and that custom solutions built with low-code tools are cheaper, faster, better tailored to business needs, and easier to change than systems built from packaged components.  It can be so messy when sacred cows come home to roost.

In other words, what’s half right is that application building is now easier than ever.  What’s half wrong is the claim that applications can easily be connected to create reliable, economical, large-scale business processes.  Building a functional process is much harder than connecting a pile of Lego blocks.

There’s one more, still deeper problem with the Lego analogy.  It leads people to conceive of applications as distinct units with fixed boundaries.  This is problematic because it creates a hidden rigidity in how businesses work.  Imagine all the issues of selection cost, connection cost, and functional compatibility suddenly vanished, and you really could build a business process by snapping together standard modules.  Even imagine that the internal operations of those modules could be continuously improved without creating new compatibility issues.  You would still be stuck with a process that is divided into a fixed set of tasks and executes those tasks either independently or in a fixed sequence, where the output of one task is input to the next.  

That may not sound so bad: after all, it’s pretty much the way we think about processes.  But what if there’s an advantage to combining parts of two tasks so they interact?  If the task boundaries are fixed, that just can’t be done.  

For example, most marketing organizations create a piece of content by first building a text, and then creating graphics to illustrate that text.  You have a writer here, and a designer there.  

The process works well enough, but what if the designer has a great idea that’s relevant but doesn’t illustrate the text she’s been given? Sure, she could talk to the writer, but that will slow things down, and it will look like “rework” in the project flow – and rework is the ultimate sin in process management.  More likely, the designer will just let go of her inspiration and illustrate what was originally requested.  The process wins.  The organization loses.

AI tools or apps by themselves don’t change this.  It doesn’t matter if the text is written by AI and then the graphics are created by AI.  You still have the same lost opportunity -- and, if anything, the AIs are even less likely than people to break the rules in service of a good idea.

What’s needed here is not orchestration, which manages how work moves from one box to the next, but collaboration, which manages how different boxes are combined so each can benefit from the other.

This is a critical distinction, so it’s worth elaborating a bit.  Orchestration implies central planning and control: one authority determines who will do what and when.  Any deviation is problematic.  It’s great for ensuring that a process is repeated consistently, but requires the central authority to make any changes.  The people running the individual tasks may have some autonomy to change what they do, but only if the inputs and outputs remain the same.  The image of one orchestra conductor telling many musicians what to do is exactly correct.  

Collaboration, on the other hand, assumes people work as a team to create an outcome, and are free to reorganize their work as they see fit.  The team can include people from many different specialties who consult with each other.  There’s no central authority and changes can happen quickly so long as all team members understand what they need to do.  There’s no penalty for doing work outside the standard sequence, such as the designer showing her idea to the copywriter.  In fact, that’s exactly what’s supposed to happen.  The musical analogy is a jazz ensemble although a clearer example might be a well-functioning hockey or soccer team: players have specific roles but they move together fluidly to reach a shared goal as conditions change.

If you want a different analogy: orchestration is actors following a script, while collaboration is an improv troupe reacting to each other and the audience.  Both can be effective but only one is adaptable.

Of course, there’s nothing new about collaboration.  It’s why we have cross-functional teams and meetings.  But the time those teams spend in meetings is expensive and the more people and tasks are handled in the same team, the more time those meetings take up.  Fairly soon, the cost of collaboration outweighs its benefits.  This is why well-managed companies limit team size and meeting length.

What makes this important today is that AI isn’t subject to the same constraints on collaboration as mere humans.  An AI can consider many more tasks simultaneously with collaboration costs that are close to zero.  In fact, there’s a good chance that collaboration costs within a team of specialist AIs will be less than communication costs of having separate specialist AIs each execute one task and then send the output to the next specialist AI.

If you want to get pseudo-mathy about it, write equations that compare the value and cost added by collaboration, with the value and cost of doing each task separately.  The key relationship as you add tasks is that collaboration cost grows exponentially while value grows linearly.*    This means collaboration cost increases faster than value, until at some point it exceeds the value.  That point marks the maximum effective team size.   

We can do the same calculation where the work is being done by AIs rather than humans.  Let’s generously (to humans) assume that AI collaboration adds the same value as human collaboration.  This means the only difference is the cost of collaboration, which is vastly lower for the AIs.  Even if AI collaboration cost also rises exponentially, it won’t exceed the value added by collaboration until the number of tasks is very, very large. 


Of course, finding the actual values for those graphs would be a lot of work, and, hey, this is just a blog post.  My main point is that collaboration allows organizations to restructure work so that formerly separate tasks are performed together.  Building and integrating task-specific apps won’t do this, no matter how cheaply they’re created or connected.  My secondary point is that AI increases the amount of profitable collaboration that’s possible, which means it increases the opportunity cost of sticking with the old task structure.  

As it happens, we don’t need to imagine how AI-based collaboration might work.  Machine learning systems today offer a real-world example the difference between human work and AI-based collaboration.  

Before machine learning, building a predictive model typically followed a process divided into sequential tasks: the data was first collected, then cleaned and prepped for analysis, then explored to select useful inputs, then run through modeling algorithms.  Results were then checked for accuracy and robustness and, when a satisfactory scoring formula was found, it was transferred into a production scoring system.  Each of those tasks was often done by a different person, but, even if one person handled everything, the tasks were sequential.  It was painful to backtrack if, say, an error was discovered in the original data preparation or a new data source became available late in the process.

With machine learning, this sequence no longer exists.  Techniques differ, but, in general, the system trains itself by testing a huge number of formulas, learning from past results to improve over time.  Data cleaning, preparation, and selection are part of this testing, and may be added or dropped based on the performance of formulas that include different versions.  In practice, the machine learning system will probably draw on fixed services such as address standardization or identity resolution.  But it at least has the possibility of testing methods that don’t use those services.  More important, it will automatically adjust its internal processes to produce the best results as conditions change.  This makes it economical to rebuild models on a regular basis, something that can be quite expensive using traditional methods.

Note that it might possible to take the machine learning approach by connecting separate specialist AI modules.  But this is where connection costs become an issue, because machine learning runs an enormous number of tests.  This would create a very high cumulative cost of moving data between specialist modules.  An integrated system will have fewer internal connections, keeping the coordination costs to a minimum.

I may have wandered a bit here, so let me summarize the case against the Lego analogy:

  • It ignores integration costs.  You can snap together Lego blocks, but you can’t really snap together software modules.
  • It ignores product differences.  Lego blocks are interchangeable, but software modules are not.  Selecting the right software module requires significant time, effort, and expertise.
  • It prevents realignment of tasks, which blocks improvements, reduces agility, and increases collaboration costs.  This is especially important when AI is added to the picture, because expanded collaboration is a major potential benefit from AI technologies.

Lego blocks are great toys but they’re a poor model for system development.  It’s time to move on.
 

___________________________________________________________
* My logic is that collaboration cost is essentially the amount of time spent in meetings.  This is a product of the number of meetings and the number of people in each meeting.  If you assume each task adds one more meeting and one more team member, and each meeting last one hour, then a one-task project has one meeting with one person (one hour, talking to herself), a two-task project has two meetings with two people in each (four hours), a three-task project has three meetings with three people (nine hours), and so on.  When tasks are done sequentially, there is presumably a kick-off at the start of each task, where the previous team hands the work off to the new team: so each task adds one meeting of two people, or two hours, a linear increase.  

There’s no equivalently easy was to estimate the value added by collaboration, but it must grow by some amount with each added task (i.e., linearly), and it’s likely that diminishing returns prevent it from increasing endlessly.  So linear growth is a reasonable, if possibly conservative, assumption.  It's more clear that cumulative value will grow when tasks are performed sequentially, since otherwise the tasks wouldn't be added.  Let's again assume the increase is linear.  Presumably the value grows faster with collaboration than sequential management, but if both are growing linearly, the difference will grow linearly as well.
 










Friday, May 05, 2023

Will ChatGPT Destroy Martech?

Like everyone else, I’ve been pondering what generative AI means for martech, marketing, and the world in general.  My crystal ball is no clearer than yours but I’ll share my thoughts anyway.

Let’s start by looking how past technology changes have played out.  My template is the transition from steam to electric power in factories.  This happened in stages: first, the new technology was used in exactly the same way as the old technology (in factories, this meant powering the shafts and belts that previously were powered by waterwheels or steam engines).  Then, the devices were modified to make better use of the new technology’s capabilities (by attaching motors directly to machine tools).  Finally, the surrounding architecture was changed to take advantage of the new possibilities (freed from the need to connect to a central mechanical energy source, factories went from being small, vertical structures to large horizontal ones, which allowed greater scale and efficiency).  We should probably add one more stage, when the factories started to produce new products that were made possible by the technology, such as washing machines with electric motors.

During the earliest stages of the transition, attention focused on the new technology itself: factories had “chief electricians” and companies had “chief electricity officers”, whose main advantage was they were among the few people who understood new technology.  Those roles faded as the technology became more widely adopted.  The exact analogy today is “prompt engineer” in AI, and will likely be even shorter-lived as a profession.  

Most of the discussion I see today about generative AI is very much stuck in the first phase: vendors are furiously releasing tools that replace this worker or that worker, or even promising a suite of tools that replace pretty much everyone in the marketing department.  (See, for example, this announcement from Zeta Global https://zetaglobal.com/press-releases/zeta-introduces-generative-ai-agents-powered-by-zoe/  .)  Much debate is devoted to whether AI will make workers in those jobs more productive (hurrah!) or replace them entirely (boo!)   I don’t find this particular topic terribly engaging since the answer is so obviously “both”: first the machines will help, and, as they gradually get better at helping, they will eventually take over.  Or, to put it in other terms: as humans become more productive, companies will need fewer of them to get their work done.  Either way, lots of marketers lose their jobs.  

(I don’t buy the wishful alternative that the number of marketers will stay the same and they’ll produce vastly more, increasingly targeted materials.  The returns on the increasing personalization are surely diminishing, and it’s unrealistic to expect company managers to pass up an opportunity to reduce headcount.)

While the exact details of the near future are important – especially if your job is at stake – this discussion is still about the first stage of technology adoption, a one-for-one replacement of the old technology (human workers) with the new technology (AI workers).   The much more interesting question is what happens in the second and third stages, when the workplace is restructured to take full advantage of the new technology’s capabilities.

I believe the fundamental change will be to do away with the separate tasks that are now done by specialized individuals (copywriters, graphic designers, data analysts, campaign builders, etc.).  Those jobs have evolved because each requires complex skills that take full time study and practice to master.  The division of labor has seemed natural, if not inevitable, because it mirrors the specialization and linear flow of a factory production line – the archetype for industry organization for more than a century.  

But AI isn’t subject to the same constraints as humans.  There’s no reason a single AI cannot master all the tasks that require different human specialists.  And, critically, this change would bring a huge efficiency advantage because it would do away with the vast amount of time now spent coordinating the separate human workers, teams, and departments.  There would be other advantages in greater agility and easier data access.  Imagine that the AI can get what it needs by scanning the enterprise data lake, without the effort now needed to transform and load it into warehouses, CDPs, predictive modeling tools, and other systems.  Maintaining those systems takes another set of specialists whose jobs likely to vanish, along with all the martech managers who spend their time connecting the different tools.  

Of course, the vision of “citizen developers” using AI to create sophisticated personal applications on the fly is entirely irrelevant when the citizen developers themselves no more have jobs.  Thousands of independent applications that make up today’s martech industry may vanish, unless the marketing Ais build and trade components among themselves – which could happen.

So far, I’ve predicted that monolithic AI systems rather than teams of (human or robotic) specialists will create marketing programs similar to today’s campaigns and interactions.  But that assumes there’s still a demand for today’s types of campaign and interactions.  This brings us to the final type of change: in the outputs themselves.

Again, we may be in for a very fundamental transformation.  The output of a marketing department is ultimately determined by the how people buy things.  It’s a safe bet that AI will change that dramatically, although we don’t know exactly how.   For sake of argument, let’s assume that people adopt AI agents to manage more of their personal lives for them (pretty likely) and that they delegate most purchasing decisions to those agents (less certain but plausible, and again already happening to a limited degree).  If that happens, our AI marketing brains will be selling to other Ais, not to people.  To imagine what that looks like, we again have to move beyond expecting the AI to do what people do now, and look at the best way for an AI to achieve the same results.

If you think about marketing today – and for all the yesterdays that ever were – it’s based on the fundamental fact that humans have a limited amount of attention.  Every aspect of marketing is ultimately aimed to capturing that attention and feeding the most effective information to the human during the time available.  

But AIs have unlimited attention.

If an AI wants to buy a product, it can look at every option on the market and collect all the information available about each one.  Capturing the AI’s attention isn’t an issue; presenting it with the right information is the challenge.  This means the goal of the marketing department is to ensure product information is available everyplace the AI might look, or maybe in just one place, if you can be sure the AI will look there.   Imagine the world as a giant market with an infinite number of sellers but also buyers who can instantly and simultaneously visit every seller and gather all the information they provide.  The classical economist’s fantasy – perfect market, perfect information, no friction – might finally come true.

And, also as the classical economists dream, the buyers will be entirely rational, not swayed by emotional appeals, brand identity, or personal loyalties.  (At least, we assume that AI buyers are rational and objective, although it won’t be easy to ensure that’s the case. That relates to trust, which is a topic for another day.)

If the role of marketing is to lay out virtual products on a virtual table in a virtual market stall, there’s no need for advertising: every buyer will pass by every stall and decide whether to engage.  With no need for advertising, there’s no need for targeting or personalization and no need for personal data to drive that targeting or personalization.  Privacy will be preserved simply because advertisers will no longer have any reason to violate it.

The key to business success in this world of omniscient, rational buyers is having a superior product, and, to a lesser extent, presenting product information in the most effective way possible.  There’s still some room for puffery and creativity in the presentation, although presumably mechanisms such as consumer reviews and independent research will keep marketers reasonably honest.  (Trust, again.)   There’s probably more room for creativity in developing the products themselves and constructing a superior experience that extends beyond the product to the full package including pricing, service, and support.

We can expect the AIs to play a major role in developing those new and optimal products and experiences, although I suspect the pro-human romantic in everyone reading this (except you, Q-2X7Y) hopes that people will still have something special to contribute.  But, wherever the products themselves come from, it will be up to the marketing AI to present them effective to the AI shoppers.

(Side note: today’s programmatic ad buying marketplace comes fairly close to the model I’m proposing.  The obvious difference is the auction model, where buyers bid for a limited supply of ad impressions.  It’s conceivable that the consumer marketplace would also use an auction.  Again, just because most of today’s shopping is based on a fixed price model, we shouldn’t assume that model will continue in the future.  Come to think of it, an auction would probably be the best approach, since buyers could adjust their bids based on their current needs and preferences, and sellers could adjust them based on inventory and current demand.  In the traditional marketplace, this would be called haggling, or negotiating, and it's the way buying has been done for most of history.  With perfect information on both sides, the classical economists would be pleased yet again. It could be fruitful to explore other analogies with the programmatic marketplace when trying to predict how the AI-to-AI marketplace will play out.)

(You could also argue that Amazon, Expedia, and similar online marketplaces already offer a place for virtual sellers to offer their virtual wares to all comers.  Indeed they do, but the exact difference is that searching on Amazon requires a painfully inefficient use of human time.  If Amazon evolves a really good AI-based search method, and can convince users to share enough data to make the searches fully personalized, it could indeed become the basis for what I’m proposing.  The biggest barrier to this is more likely to be trust than technology.  It’s also worth noting that traditional marketing barely exists on those marketplaces.  The travel industry, where most marketing is centered on loyalty programs, may be an early indicator of where this leads.)

So what role, exactly, do humans play in this vision?  

As consumers, humans are no longer buyers.  Instead, they receive what’s purchased on their behalf.  So, their main role is to pick an AI and train it to understand their needs.  Of course, most of that training will happen without any direct human effort, as the AI watches what its owner does.  (I almost wrote “master”, but it’s not clear who’s really in charge.)  For people with disposable income, purchases are likely to move away from basic goods to luxury goods and experiences.  Those are inherently less susceptible to purely rational buying decisions, so there’s a good chance that conventional buying and attention-based marketing will still apply.

As workers, humans are in trouble.  Farming and manufacturing have been shrinking for decades and AI is likely to take over many of the remaining service jobs.  Some conventional jobs will remain to do research and supervise the AI-driven machines, and there may more jobs where it matters that the worker is human, such as sports and handcrafts, and where human interaction is part of the value, such as healthcare.  But total employment seems likely to decrease and income inequalities to grow.  It’s possible that wealthy nations will provide a guaranteed annual income to the under-employed.  But even if that happens, meaningful work will become harder to find.

I’ll admit this isn’t a terribly pleasant prospect.  The good news is, predictions are hard, so the odds are slim that I’m right.   I’m also aware that we’re at the peak of the hype cycle for ChatGPT and perhaps for AI in general.  Maybe what I’ve described above isn’t technically possible.  But, given how quickly AI and the underlying technologies evolve, I wouldn’t bet on technology bottlenecks blocking these changes indefinitely.  Quantum AI, anyone?

All that said, of the three major predictions, I’m most confident about the first.  It’s pretty likely that a monolithic marketing AI will emerge from the specialized AI bots that are being offered today.  The potential benefits are huge, the path from separate bots to an integrated system is an incremental progression, and some people are already moving in that direction.   (Pro tip: it’s easier to predict things that have already happened.)

The emergence of an AI-driven marketplace to replace conventional human buying is much less certain.  If it does happen, the delegation will emerge in stages.  The first will cover markets where the stakes are low and buying is boring.  Groceries are a likely example.  How quickly it spreads to other sectors will depend on how much time people have and how much intrinsic enjoyment they derive from the shopping itself.

The role of humans is least predictable of all.  Mass under-employment probably isn’t sustainable in the long run, although you could argue it’s already the reality in some parts of the world.  The range of possible long-term outcomes runs from delightful to horrific.  Where we end up with depend on many factors other than the development of AI.  The best we can do is try to understand developments as they happen and try to steer things in the best directions possible.