New research shows how LLMs can balance organizational performance with individual priorities.
Listen to this article:
If AI is integrated into operational and decision-making systems within organizations, it could in effect start to act like a manager. Think of AI in charge of assigning workers to tasks, sequencing work, or distributing resources. Executives who oversee this type of AI transformation face an important design challenge: can AI be beneficial and effective for your firm, while also being aligned with the interests of your employees, your customers, and others affected? The new paper, “Guiding Worker Self-Selection in Crowdsourcing Contests: An LLM-Augmented Algorithmic Approach,” co-authored by HBS AI Institute chair and co-founder Karim R. Lakhani, examines this dynamic in the context of crowdsourcing contests, specifically exploring whether AI can power a recommendation system intended to increase productive effort while limiting workers’ dissatisfaction.
Key Insight: Give the LLM the Hard Design Problem
“The difficulty is the ordering: When workers differ in their valuations, production efficiency, and effort costs, choosing a good score vector by hand is far from obvious.” [1]
On crowdsourcing platforms (such as Topcoder), clients post tasks (like software or logo design) as online contests. Workers enter by building and submitting complete solutions. Clients want substantial, high-quality submissions, so platforms benefit when workers invest more effort. But typically only a single winner or small number of winners receive rewards for a contest (typically monetary compensation), so workers will likely regret having competed if they didn’t win, or if, theoretically, another contest would have offered a better outcome relative to the effort they invested.
To address this tension, the researchers introduce GRAF, a greedy algorithmic framework* for making recommendations about which contests workers should enter. GRAF considers workers sequentially, recommending a contest to each one based on the selections it has already proposed for earlier workers. Devising an optimal solution to this is quite complex because workers differ in how much they value each contest, how productive they are, and how costly effort is for them. This complexity motivates the researchers to use an LLM to develop GRAF’s worker-priority rule, which determines the order in which GRAF will consider workers.
Key Insight: Test What Works
“It adopts an evolutionary search procedure, iteratively searching for algorithms that yield better trade-offs between OE [Overall Effort] and regret.” [2]
Rather than asking the LLM to invent one worker-priority rule and accepting its answer, the researchers created LLMScore, a system that repeatedly generates and tests possible rules. LLMScore produces candidate scoring algorithms as executable Python code, which are then evaluated on simulated crowdsourcing contests. Rules that maximize total worker effort earn high scores, but rules causing excessive worker regret are penalized. The best rules are mutated or combined to create stronger rules. If improvements stall, LLMScore automatically rewrites its own prompt to discover fresh rules.
The researchers trained this process on 50 small simulated contests, then tested the resulting rules on larger and structurally different scenarios. The LLM-generated rules generally outperformed the researchers’ handcrafted alternatives. For example, in the most complex setting with 15 workers and three contests, the best LLMScore rule produced about 7% more median overall effort while reducing median maximum worker regret by 55%. And because LLMScore outputs code, operators can inspect, audit, and modify what the model has produced rather than deploying an opaque model directly into their system.
Why This Matters
It is easy for an organization to say that an AI-enabled system should “balance stakeholder interests,” but that phrase avoids the hardest implementation questions. Which interests? How will each be measured? What mechanisms will actually enforce the trade-offs? For executives and business leaders deploying AI across operations, workforce systems, or customer interactions, alignment becomes a design choice that has to be specified, weighted, tested, and governed. This research highlights how transparent, iterative systems offer a playbook for letting AI direct human workers in a sophisticated and responsible manner, and crucially, because workers retain full autonomy, the AI recommendations serve as advisory guidance rather than top-down directives. Furthermore, by stress-testing these alignment mechanisms in synthetic simulations first, organizations can validate trade-offs across diverse scenarios before deploying AI in live workforce environments.
References & Notes
[1] Thach, Nguyen, Hau Chan, David Parkes, and Karim Lakhani, “Guiding Worker Self-Selection in Crowdsourcing Contests: An LLM-Augmented Algorithmic Approach,” arXiv preprint arXiv:2609.07749 (2026): 2. https://doi.org/10.48550/arXiv.2609.07749
[2] Thach et al., “Guiding Worker Self-Selection in Crowdsourcing,” 5.
* A greedy algorithm is a method that makes the best immediate choice at each step, building up a complete solution one decision at a time without recalculating past steps or searching through every possible combination. In the paper’s context, GRAF doesn’t try to solve the entire assignment problem for all workers at once. Instead, it processes workers sequentially in a specific order and recommends to each worker their best option based on the choices already made by earlier workers.
This article was created through a human-led, AI-assisted process.
Meet the Authors
Nguyen Thach is a PhD student at the University of Nebraska-Lincoln.

Hau Chan is Associate Professor at the University of Nebraska-Lincoln.

David Parkes is John A. Paulson Dean and George F. Colony Professor of Computer Science at the Harvard John A. Paulson School of Engineering and Applied Sciences.

Karim R. Lakhani is the Dorothy & Michael Hintze Professor of Business Administration at Harvard Business School. He specializes in technology management, innovation, digital transformation, and artificial intelligence. He is also the Co-Founder and Faculty Chair of the HBS AI Institute and the Founder and Co-Director of the Laboratory for Innovation Science at Harvard (LISH).