A Sacramento brokerage’s program led by Mukesh Kumar shows how to turn AI into a controllable operating asset with guardrails, logs, and P&L‑tied KPIs.

Artificial intelligence in logistics suffers from a credibility gap. Vendors promise transformation, but few can demonstrate how autonomous systems actually operate inside a live supply chain, or show boards what happens when the software makes a mistake. At T3RA Logistics, a Sacramento-based freight brokerage managing $30 million in annual volume, Chief Growth and Technology Officer Mukesh Kumar has built something different: an AI deployment that answers to finance committees rather than impressing them.

Kumar’s approach replaces sweeping transformation narratives with narrow, auditable agents that execute specific tasks under strict supervision. A pricing agent standardizes target rates across lanes, cutting pricing cycles from hours to minutes and delivering roughly $40,000 in monthly cost avoidance. Operational agents manage tender checks, appointment scheduling, and status updates, but only within predefined boundaries. The structure produces results a CFO can track: gross margins have climbed from approximately 11% to 15% year over year.

Policy Before Code

Kumar’s program begins not with models but with restrictions. Before any agent touched a production shipment, T3RA’s leadership documented what the software could never do: edit timestamps, commit to penalties, or negotiate without human approval. Every automated action must include a reason code. Any scenario requiring judgment routes to a named manager. “The rules make the system legible to risk and audit teams,” Kumar explains. “Accountability stays with managers, not tools.”

The human-in-the-loop design is intentional. Agents process the predictable 80 percent of activity — routine confirmations, data transfers, and status pings. People own the remaining 20% where money, relationships, or brand reputation are at stake. The goal is not workforce reduction. It is cleaner exception handling and steadier service levels.

Five Deliverables Boards Should Demand

For directors evaluating an AI budget request, Kumar’s framework offers a checklist. Every proposal should include:

  • A one-page AI policy written for non-technical board members, specifying what agents can and cannot do, consent rules for contacting external parties, escalation procedures, and data retention standards.
  • A guardrails matrix defining confidence thresholds, time limits, and hard stops for each task. If a facility portal times out three times, the playbook states who gets paged and what happens next.
  • Immutable logs capturing inputs, outputs, reason codes, and final human decisions in a system that supports forensic review. “If a regulator or enterprise customer asks for proof,” Kumar notes, “the artifacts exist.”
  • A KPI dashboard linked to P&L, tracking touches per load, exception rates, on-time performance, cycle time, and margin contribution. Every agent in production must show before-and-after deltas with trend lines.
  • A blameless incident playbook documenting what triggered each error, which guardrail succeeded or failed, the corrective action taken, and steps to prevent recurrence. Status reports flow to the audit or risk committee until closure.

Pilot What You Can Measure

Platform vendors often promise end-to-end automation. Kumar favors the opposite: small, bounded agents that do one job well. A contained scope reduces integration risk across the fragmented reality of email threads, carrier portals, and facility-specific protocols. It also makes outcomes measurable. An agent either reduces manual touches and exceptions, or it does not.

Boards can enforce this discipline by approving discrete 30-day pilots with named owners, fixed budgets, and explicit success criteria. Monthly reporting goes to finance; quarterly reviews go to audit or risk. Anything that fails to move the metrics gets rolled back.

Accountability as Operating Discipline

Clear ownership prevents the drift that kills AI initiatives. At T3RA, responsibility is assigned like any other capability:

The COO owns rollout and operational outcomes. The CFO owns KPI integrity and P&L linkage. The general counsel or the chief risk officer owns policy, privacy, and vendor contracts. A lean automation team — one product manager, one engineer, one data lead, and a rotating operations specialist—delivers changes on a weekly cadence.

Compensation rewards subtraction as much as addition. Deleting underperforming agent behaviors keeps the system maintainable and risk-bounded.

The Dual-Use Stress Test

Kumar’s reliability standard comes from trialing new agents on military freight first. Requirements are unforgiving: immutable logs, deterministic fallbacks, and clear override paths. Only after an agent performs under those constraints does it graduate to retail or food logistics. Government procurement teams demand auditability. Boards should want the same.

Questions for the Next Meeting

Directors evaluating AI spending should ask: Which tasks are in the next 30-day pilot, and why? What are the red lines, and who approves exceptions? Where are logs stored, and who audits them? Which KPIs will change in the first month, and how will the team know? What gets rolled back if metrics do not improve?

Most AI programs stall under tool sprawl and vague accountability. The test for directors is whether management can govern AI the way it governs safety, quality, and finance — with policy, logs, numbers, and names. In a soft freight market where inches of margin separate survivors from casualties, treating AI as a controllable operating asset rather than a moonshot may be the edge that matters.