AI Automation in Operations

The automation curve

Ryan Petersen's Flexport supplies the clearest publicly stated maturity curve for this pattern. At the start of one year the company had automated roughly 20 percent of its operations-heavy work; by the end of that year the figure had reached 50 percent; the original ceiling estimate of 80 percent proved wrong, and the revised expectation is 90 to 95 percent or higher, a number that keeps rising as the underlying models improve.¹ The remaining slice is a stubborn tail of what the team calls the weird crane on truck problem, edge cases that still require a person to physically show up.¹

Three mechanisms explain the jump from the first plateau to the second: classical optimization software for deterministic problems such as routing, large language models for converting unstructured documents like emails, spreadsheets, and PDFs into structured data, and language model agents for communication tasks that were previously too expensive to automate, such as phone calls to verify an address.

Three types of automation

The pattern sorts into three categories. Augmentation lets people do existing work faster, such as natural-language reporting that removes a quarter of an account manager's time. Displacement replaces work a person was doing outright, such as routing an email or translating a booking. Expansion does work that was never economically justified before at all: a solver re-optimizing a canceled shipping container ten times a day, or a phone agent verifying every address proactively because the call now costs nothing to run. Expansion is the type most often left out of automation return-on-investment math, because the saving is not fewer people doing a task but an entirely new activity that creates value nobody was capturing before.

A concrete instance of that gap: Flexport's customs error rate fell from 1.8 percent, already good for the industry, to 0.2 percent, after an AI auditor began reviewing every customs entry before filing, a system that launched in October 2025 with results visible by that November and December.¹ Petersen states the lesson directly: "AI can be 10x better than the human, not just cheaper."¹ A ninefold reduction in defect rate is a different kind of claim than a labor-cost saving, and it is the kind of improvement that wins back an enterprise customer after a quality failure. A second system built the same way, a bill-audit agent assembled in a single week, reconciles every incoming carrier and trucking invoice against the quote, the procurement database, and the shipment's own event and message logs, catching billing errors at a granularity no human team had audited before.¹

Augment before you automate

Eric Glyman supplies the sequencing rule the taxonomy above implies but does not state outright. On the logic that "people tend to jump right to full automation, when really what is incredibly valuable in the short run is augmentation of people on aspects of what they do," Ramp treats augmentation as the first move and full automation as something that follows only where individual sub-steps have proven out.² The method is to decompose a workflow into its actual steps rather than buy a single point solution: instead of purchasing a generic "AI sales development rep," Ramp traced what a rep actually does, find a buying signal, build the list, find the contact's email, write personalized copy, send it, and tooled each step separately. The result was a sales team booking three to four times the meetings of the nearest competitor, with some steps eventually automated fully once they had been proven at the augmentation stage. The same pattern, run by Ramp's applied AI team sitting alongside manual underwriters for a two-week sprint, cut average underwriting time from about two days to a quarter or half a day.

Why it matters

The pattern gives operators a shared vocabulary for what stage of an automation rollout they are actually in, and a warning against skipping straight to full automation before the underlying steps have been decomposed and proven individually. Freight forwarding labor is roughly 10 percent of the total cost an importer or exporter pays, so automating 80 to 90 percent of that work translates to an 8 to 9 percent reduction in total freight cost, a template other operations-heavy industries can apply by identifying their own labor percentage and automation ceiling.