Money is the Utility Function: Thinking about Return on AI

2024-06-06

The attraction of AI, one of the fastest-moving technologies in recent memory, shouldn't obscure the fact that in a business context the tech stack is simply the how and the bottom line is the why. So far most of the discussion about the impact of AI in business has been driven by technologists and futurists: descriptions of capabilities (real, future, or neither) and their direct effect on activities as seen through the eyes of CTOs and end users instead of CFOs and COOs. This is a good way to develop new technologies and a poor way to deploy them. The business impact of using AI in a business, just like every other technology, can only be understood and optimized by combining technological know-how with an analytical framework focused on the business outcomes, not just technological capabilities or specific processes.

A first way to conceptualize this is as return on AI, where "AI" can be read as "artificial intelligence," but in a more general context it should be understood as "augmented intelligence." In a world where the cognitive capabilities of organizations, people, and even objects can increase month after month, understanding not just what technologies are available but what the financial return on the economic, opportunity, and adaptation costs of deploying them is a critical aspect of management at all scales.

RoAI is not fundamentally different from other forms of investment analysis. Given a model of, say, company valuation as a function of endogenous decisions and probabilistically known exogenous factors, it's conceptually straightforward — and in practice fascinatingly complex — to understand the impact and risk profile of different investment choices. What makes RoAI different is that our collective experience and analytical toolset is much newer and more tentative than for types of investments that in some cases have been undertaken for thousands of years. We are dealing with new investment possibilities that impact our organizations through new and still only partially understood channels. Right now RoAI is as much an art as a science.

Although there's no manual or software package for performing this kind of analysis, and therefore no substitute for experience and creative judgment, it can be useful to look at specific ways in which the analysis of investment in contemporary AI tools has nuances that are conceptually very familiar to managers and investors but are under-discussed in technology-driven narratives:

Process heterogeneity

Most processes in manufacturing or finance have very precise operational definitions: shorting a stock or assembling a laptop can in practice be complex activities involving multiple steps, but they are homogeneous enough that it's possible to either anticipate or measure the impact of a new technology to a sufficiently high level of accuracy to guide investment choices.

Although often falling under the umbrella of finance and engineering, there are other processes — generally speaking those closer to knowledge work — that hide enormous variability under an apparently identical label. Ignoring this heterogeneity adds underestimated or just blind risks to the deployment of AI technologies in those projects.

Consider for example the use of coding assistants like Microsoft Copilot. It's clearly demonstrable, even notorious, that these tools increase the speed of coding. It's less often noted that "coding" is a single activity in the same sense than "writing" is a single activity. A developer typing code in an IDE can be implementing a well-known typical application they have done multiple times, spending most of the time familiarizing themselves with a new library or framework. Or they could be working out the design of a first-of-its-kind high performance system, in which case the amount of coding involved might be very small compared with the time spent reading the relevant literature and performing abstract analysis, research, simulations, and experiments.

The use of a coding assistant will have quantitatively and even qualitatively different impacts on the business characteristics of these development processes. For a typical application using popular libraries automated code generation might reduce enormously total development time without diminishing the quality of the end product; anecdotally, this might be by 50% or more, with corresponding reductions in the all-important labor costs per project.

For a truly innovative system the situation is entirely different. Coding in this sort of project can be a very small component of its time and labor costs compared with other forms of R&D. Even the MVP phase of a project can come much later than the true beginning of the work: both Google and Bitcoin were pen-and-paper mathematics before there was a single line of code.

There's a similar impact from specialized requirements in performance or robustness. The more a system needs to push back the technological frontier, the less of its code, be it large or small, can be generated by an AI, because it will look less and less like any code the AI has been trained with.

Of course, most projects live somewhere in the spectrum between a cookie-cutter web page and control code for a space telescope. But companies and even projects within the same companies are indeed different. Without an evaluation process that's both technically savvy and brutally honest a business faces the twin risks of underinvestment — spending less on AI for a given process assuming it's more idiosyncratic than it really is — or overinvestment, deploying coding assistants and similar tools in other domains to perform work that's too specialized and unique to the project to be meaningfully helped along by that sort of AI..

The second risk implies a subtler one related to scheduling and budgeting. Assuming that a project will be sped up by a coding or writing assistant more than it can be given its nature can lead to massively underestimating time and labor costs or, even worse, putting project leaders under pressure to change the project into something that can be sped up by a coding assistant. This is perhaps the most insidious danger of applying AI without a careful consideration of process and business context because it can lead to a vicious circle where obvious fast turnaround times coupled with unrecognized losses in delivered innovation nudge a company away from its strategic plans.

Downstream scalability

Deploying AI in all processes where it can improve productivity and only in those is only part of a business-driven implementation strategy. The fact that AI tools are relatively inexpensive and easy to deploy leads to implementation choices being made at a very low level, often by small teams or even individuals. We've seen above how this can backfire at the process level, but it can also have unintended negative impacts at the company level.

A good analogy would be a manufacturing process and an equipment upgrade that multiplies the quantity and speed with which a certain component is built. Absent equivalent upgrades or capability investments in downstream processes, the optimal use of this upgrade is to build the same number of components and benefit from time and cost savings rather than to overwhelm the system with a glut of unusable components..

Knowledge work in a corporate context both is and isn't the same. Mails, reports, slides, graphs, presentations, or models can all be seen as informational components of downstream decision-making processes. In that sense, tools that help automate parts of their creation — and AI tools are already very good at it — reduce enormously the time it takes to "manufacture" a comparable unit of content.

Companies face the same sort of optimization questions for internal content flow in decision-making processes as they do in factory lines. Twice the number of slides requires twice as much time to analyze: is the extra opportunity cost to highly-paid, time-restricted managers justified by the improved decision making? Is decision-making improved at all by the new material? AI-assisted writing carries in itself a strong probability of more writing requiring more time to read and digest. This might have a benefit and certainly has a cost. The most direct benefit accrues to the person generating the context but the time and attention cost is diffused across the company as a whole. Without a clear-eyed company-wide deployment strategy designed to maximize the benefit to the bottom line, the use of productivity-enhancing tools at the individual and team level can have downstream impacts that more than negate those benefits.

(It's always possible to use AI tools downstream to summarize the input coming from upstream processes. This is a surprisingly frequent and financially nonsensical setup resulting from the sum of individually rational deployment choices: the company is paying for two AI services —in direct and indirect ways— simply to create things on one step and remove them on the other.)

The pattern of subtler risks coming from individuals taking tactically optimal decisions that result in strategically poor outcomes can be even more dangerous than this: without very solid thresholds for the marginal impact in decision-making quality of the information in a given piece of internal content or presentation (and this is the foundational concept of RoAI analysis) individuals will generate content that's easier for the AI tool and therefore less costly for them in time and resources, not the internal content that most benefits the company.

Every company has a sense of what types of information are most important to its decision-making. For most companies this sense has become too imprecise, or plain outdated, as the nature of the tools generating and processing content has changed.

None

Paradoxical equilibria

The danger mentioned above is a form of paradoxical equilibrium where the individual's decision-making process, absent an updated collective framework, leads to negative organizational outcomes. In order to understand the longer-term strategic impact of AI tools it's important to keep in mind that this happens as well at the level of whole industries.

AI tools are extremely expensive to build: there are very few organizations in the world capable of building the largest and most advanced types of foundational models. But the deployment of AI tools, at least in the purely technological sense, is extremely simple and cheap. Many of the issues mentioned in this piece wouldn't happen very often or at all if these tools were expensive. Nobody buys and deploys an expensive suit of large-scale robotic machinery without a careful analysis of their impact across the whole company.

Systemically, their low cost of access and implementation means that AI tools have to be considered ubiquitous. They might be cutting-edge in a technological sense but they are not competitive advantages by themselves. Whatever benefit a company will gain from using, for example, a coding assistant, it will not be by comparison with its competitors, all of which will also be using coding assistants, and very often the same one.

None

It's easy to construct a plausible narrative for any industry where coding or writing assistants have a large impact on productivity that follows this pattern:

This sort of path is very sensitive to the particulars of each industry or market yet it's at least as plausible as the alternative. In the specific case of the software factory industry, far from a world of fewer developers and larger margins, AI coding assistants have a significant likelihood of pushing companies towards much larger headcounts and much slimmer margins as they attempt to survive in a context of universally lower unit labor costs. This is unless they shift to the sort of bespoke development that cannot be helped much by AI assistants.

Either way, the tech demo and the business future need not look even remotely similar.

The path from prompt to profit is not obvious

The main takeaway is not that any of these things is necessarily going to happen. Instead it's that the impact of an AI tool at small scale and short term is much different from its impact over the longer term and at the level of companies and industries. Technological know-how is an institutional prerequisite to understanding these possibilities and using AI in a way that most benefits the business goals of a company yet it's far from sufficient.

An analytical process that takes as foundational the bottom line and not the neural network is the keystone of a successful strategy, and until enough time has passed for us to fully understand the technology and its applications — in other words, until it has become boring — understanding the return on AI will remain both difficult and necessary.