Agentic AI in Sourcing: Solving the “Optimal Stopping” Challenge

The procurement sector is navigating a phase shift from "Predictive" tools to the Agentic era. This evolution, driven by the maturation of Large Language Models (LLMs) and Reinforcement Learning, allows software to not just recommend actions, but to execute them autonomously.

The Evolution of Agency

Current enterprise systems are largely deterministic, following rigid "if-then" logic. True agency—or "Level 5" autonomy—requires a system to operate with goal-oriented behavior.

An agentic system perceives its environment (the supplier market), maintains an internal state (current best offers, time elapsed, requester urgency), and selects actions that maximize a long-term reward function. The distinction is critical:

• Deterministic Systems: Stop when a timer expires.
• Agentic Systems: Stop when the marginal cost of waiting exceeds the expected gain from further solicitation.

The Stopping Problem as a Pillar of Autonomy

In fast-paced sourcing—such as spot buying or logistics—the decision to stop is the primary lever of value.

• Stopping too early: Results in "money left on the table" due to premature closure.
• Stopping too late: Results in operational friction, missed delivery windows, and "analysis paralysis."

To engineer a robust solution, Keelvar grounds its algorithms in established mathematical theories of sequential decision-making.

The Secretary Problem & Pandora’s Rule

A famous theoretical analog is the Secretary Problem, which dictates a strategy for picking the best candidate in a sequence. However, sourcing is more complex because we are risk-averse and face "search costs" (time).

Keelvar utilizes Pandora’s Rule, an index policy where the agent calculates a "Reservation Price" for each supplier. The search stops when the maximum value of bids already received exceeds the index of the next best unsearched supplier. This provides a rigorous mechanism for linking "patience" to "stopping time."

Hyperbolic Discounting in Behavioral Economics

Humans rarely value time linearly. We exhibit hyperbolic discounting, valuing immediate resolution disproportionately higher than delayed resolution. An effective agent must utilize a non-linear patience curve to capture the "urgency cliff" where user utility drops to zero.

To solve the stopping challenge, Keelvar uses a Neuro-Symbolic architecture. This hybrid framework combines the data efficiency of mathematical models with the contextual reasoning of modern AI.

• The Predictor (Bayesian Optimization): Using Gaussian Processes, the agent treats the market price as a trend to be modeled. It predicts the likelihood of a better bid arriving—even with sparse data—ensuring decisions are grounded in probabilistic reality rather than guesswork.
• The Strategist (Deep Reinforcement Learning): For high-volume categories like logistics or spot buys, the agent learns through experience. It constantly evaluates the "State" (market velocity and time remaining) to select the "Action" (wait or stop) that maximizes the long-term reward for the business.
• The Empath (Inverse Reinforcement Learning): Because users often can't quantify their own "cost of delay," the agent uses IRL to infer it. By analyzing thousands of historical manual events, it learns the hidden "patience levels" implicit in human behavior, allowing the AI to mimic the intuition of a seasoned category manager.
• The Contextualizer (Large Language Models): Mathematical models are often blind to tone. Our LLM layer monitors unstructured communication to detect critical signals—like a supplier’s hint that they can "sharpen their pencil" or a requester’s growing urgency—and adjusts the stopping parameters in real-time.

Keelvar’s rollout of agentic capabilities follows a tiered phase approach to ensure safety and trust:

1. Phase 1: Smart Timeout (Decision Support): Implementing Gaussian Process models to show users a "probability cone" of future bid improvements.
2. Phase 2: Learned Patience (IRL Layer): Removing manual urgency sliders by clustering users into "Personas" (e.g., The Rusher vs. The Optimizer).‍
3. Phase 3: Neuro-Symbolic Autonomy (Full Agency): The DRL agent takes control of the "Stop/Wait" decision, supervised by an LLM that provides natural language explanations for its actions.

Deploying autonomous agents requires rigorous guardrails:

• Anti-Gaming: To prevent suppliers from "sniping" (waiting until the last second to bid), agents apply randomized stopping policies.
• Explainability (XAI): Every decision is auditable. Users can use "Replay Mode" to see the agent's internal reasoning—how the probability of a better bid declined until it crossed the patience threshold.‍
• Fairness: Reward functions include terms to ensure smaller, diverse suppliers aren't systematically excluded by an agent optimizing purely for velocity.

The stopping problem is a key pillar for Agentic AI in Sourcing that few vendors have addressed. By adopting a Neuro-Symbolic architecture, Keelvar transcends simple automation. Our agents do not just execute instructions; they act with the nuanced judgment of a strategic partner—knowing exactly when to wait for a better deal and when to act to keep the business moving.

‍