Here come the AI agents: can they conquer bureaucratic “sludge”?
Here come the AI agents: can they conquer bureaucratic “sludge”?
Inferences from Minerva Technology Policy Advisors. Vol.36 - 29 October, 2024
Google is reportedly working on an agentic artificial intelligence product codenamed “Project Jarvis”, and Microsoft CEO Satya Nadella announced the “cornerstones” of the company’s autonomous agent models this week.
What’s happening? The age of the “agents” is upon us. Competing labs are rolling out new, more advanced models that they hope will solve multi-stage problems currently being tackled by human workers.
So what? These models are getting better at function-calling — which allows them to access and use external tools or software — and at breaking larger problems into consecutive tasks.
Yet people equipped with human-level intelligence are routinely stymied by onerous and confusing bureaucratic processes: tedious and time-consuming tasks that the Harvard Law Professor, Cass R. Sunstein, has termed “sludge”.
Inference: Developers will have to focus on ensuring that AI agents can interoperate without making mistakes or causing harm; governments and companies will also need to prime workflows to function with agents, devoting serious time and attention to tackling some of the following challenges:
Friend or foe? Many business and administrative processes that happen on the internet today rely on verifying that human users are humans, rather than automated bots. Agentic systems will blur the line between the two, putting pressure on developers and their platforms to discern when to grant access to legitimate agents, while denying it to bad actors using bots to run scams, or launch denial of service attacks.
Determining when an agent is benign and trying to be helpful, versus when it poses a threat, will be a challenge.
Who is who? Look for identification and digital signature protocols to evolve to accommodate agents while remaining secure. Security-paramount institutions like banks, tax authorities and other organizations that handle sensitive data will also have to consider institutional and process innovations to ensure that they can protect against fraud and impersonation.
Where do agents work most effectively? As workers begin to ask early-stage agents to take on the “sludge” of daily life, errors and issues will likely be common, as they have been with the image, text and video generation capabilities of foundational models. Websites will have to provide interoperability, and may end up competing to be the best destination for agents performing tasks like procurement, booking and scheduling.
Agent friendly? Rating and accrediting websites for agent interactions could become a new cottage industry. Just as the ✨ symbol has become a near-universal indicator that “artificial intelligence was involved here”, users can expect to start seeing “agent-friendly” kite marks on the platforms that develop interoperability measures and want to tout them.
Agents capable of breaking problems into multiple steps and solving them will probably roll out inside companies first, where data inputs and outputs are more constrained, where workflows are well understood, and where risks of a major screw-up are lower. Agents that need to operate between different companies, and across many different digital domains, will pose more of a challenge, and will also likely be of greater concern to regulators.
What unforeseen factors could nudge agent behavior? Decision architecture to affect human purchasing behavior has been at play in stores and online for decades. Subtle cues that can nudge our actions in the direction of a purchase or a click are a big part of how websites are built.
Web developers may come under pressure to build new architectures that have a similar effect on agents; and those that deploy agents will have to learn how and why their systems are being influenced in return.
Devoid of human emotion or experience, these nudges will be of a different character to those that make us book tickets to a particular event, or buy one must-have set of sneakers instead of the other, but agents will still be influenceable.
What still needs a human touch? Scheduling meetings, booking flights and applying for licenses are all menial, although laborious, “sludge”. But if agents eventually become reliable enough to be applied to other, higher-order tasks, more difficult questions may arise.
What role, if any, should AI agents play in hiring and firing? Can agents successfully adopt and adhere to employment law, respect protected characteristics, and reliably deal with sensitive information? Will future generations of more advanced agents eventually be able to take on managerial roles? Should they?
On many of these parameters, human workers are far from perfect. But governments and companies will face a challenging balancing act as they try to enable agents to engage with them, while respecting ethical and legal boundaries.
Homo digitalis. As Yuval Harari writes in a recent op-ed for the FT; “the human world is a latticework of multiple bureaucracies, in which AIs can accrue enormous power.” This may be true, but only if they can successfully navigate the many interlinked domains of information and successfully dock with existing governmental and private sector networks that have been made interoperable.
The upshot is that the integration of agentic artificial intelligence into the working world will not be smooth. It will take more than just clever programs and new breakthroughs in model capabilities for AI to be effective in addressing the “sludge” problem; it will take adjustments to governmental and corporate bureaucracy as well. It will take a form of diffusion (see our volume of Inferences on this subject with Jeffrey Ding here) that gives agents a safe and effective way to dock.
What we’re reading:
Harari’s piece in full; which seems to misread the power current AI systems have to manipulate what he calls “labyrinthine bureaucracies”.
Reporting on Hong Kong’s proposed AI policy.
More on the UK’s mounting debate over providing access to publishers’ data for AI training.
What we’re looking ahead to:
12 - 14 November: IEEE World Technology Summit on AI Infrastructure.
21 November: The Evident AI Symposium, New York.
10 - 11 February 2025: AI Action Summit in Paris, France.
2 - 4 June 2025: AI+ Expo and Ash Carter Exchange in Washington, DC.