Why Your AI Isn't Showing Up in Your P&L
Author: Pravin Borade, Senior Vice President, Business Systems Engineering, Searce
Article 3 of 5: AI for Logistics that Shows Up in Your P&L
Last time in Part 2, we established that the 40-person manual exception team in your billing operation is not an implementation failure, it is a category boundary that sustaining platforms were never designed to cross.
In this article we will learn why the AI you deployed on top of that operation produced a dashboard your team uses and a P&L impact your CFO cannot see.
Your board mandated AI. You delivered it. You have the implementation, the QBR slide, the executive briefing.
Twelve months later, your CFO's working capital position has not moved. Your DSO is where it was. Your billing team is the same size. The AI is working - the analysts use it, the dashboard is live, the reports are faster. But the P&L does not know any of this happened.
Forrester's 2024 enterprise AI study found that 63% of logistics companies that deployed AI in finance and operations reported no measurable improvement in working capital metrics within 18 months of deployment. The AI was not wrong. The deployment was.
The difference between AI that shows up in your P&L and AI that shows up in your QBR deck is not model quality, not data quality, not budget. It is placement. And almost every logistics AI deployment in the last three years got the placement wrong.
The Theory: Daniel Kahneman and the two systems
In Thinking, Fast and Slow, Daniel Kahneman established the most practically useful model in modern behavioural science: System 1 and System 2 thinking. System 2 is slow, deliberate, and analytical; it handles novel problems requiring conscious reasoning. System 1 is fast, automatic, and pattern-based; it handles familiar problems through trained recognition, without conscious deliberation. Most human expertise is the process of converting System 2 decisions into System 1 responses through repetition: what once required careful thought becomes automatic.
The practical implication for management: System 2 is expensive. It requires time, attention, and cognitive bandwidth. The goal of operational design is to identify which decisions can be safely converted to System 1 and build the systems that make that conversion possible.
The Modern Adaptation
Kahneman described System 1 and System 2 as properties of human cognition. But the distinction maps directly onto how AI can be deployed in an operational workflow, with one critical difference in consequence.
AI deployed beside the process functions as an augmented System 2: it analyses faster, surfaces patterns the human would have missed, and produces better-informed decisions. But the human is still in the loop. The decision still requires a person, a review step, and the time both consume. The latency stays. The headcount stays. The cost stays.
AI deployed inside the process functions as a trained System 1: the decision point is removed entirely, replaced by an automated resolution rule that fires without human review when the pattern is recognised. The latency goes. The headcount goes. The working capital recovers. The distinction is not about AI sophistication. The same model, the same data, the same training - placed differently - produces entirely different financial outcomes. One shows up in your QBR. The other shows up in your CFO's working capital position.
The Global Proof: Amazon Logistics vs. the industry
Amazon Logistics processes billing across millions of daily transactions spanning fulfilment, last-mile, and third-party carrier payments. Their AI is not deployed beside the billing process - it is the billing process. When a rate exception occurs, an AI classification layer identifies the exception type, matches it to a resolution rule, and closes the billing cycle without a human review step. The outcome: Amazon's fulfilment billing cycle time averages under 24 hours from shipment close to invoice trigger. The industry median for comparable multi-modal operations sits at 6–12 days. That gap is not a function of Amazon's model quality, it is a function of where the AI sits in the workflow.
The mid-market 3PLs who deployed AI in the same period largely built reporting layers: exception dashboards, DSO analytics, billing status monitors. All valuable. None of them removed a single human from the billing path. DSO analysis is faster. DSO itself is unchanged.
The instinctive executive objection: "Amazon built that at a scale and with an engineering budget that has no comparison in mid-market freight." The objection is about scale, not about architecture. The decision to place AI inside the workflow rather than beside it is not a budget decision, it is a design decision. A $300M freight forwarder processing 15,000 invoices a month has enough exception volume to train a classification model, enough resolution patterns to automate the most common categories, and enough billing cycle latency to make the working capital recovery material. The question is not whether you can afford to do it. It is whether you designed your AI deployment with the placement decision in mind at all. Most did not.
(To see what AI embedded inside a billing workflow produces in practice - including a 10% improvement in working capital availability in production - the Automated Collections module on our logistics practice page has the full outcome breakdown.)
The Reality: What we repositioned and what changed
We run billing operations for a major global 3PL group company, processing 2.5 million invoicing transactions per year across multiple modes and acquired entities.
When we first engaged, they already had AI. A reporting and analytics layer that classified exception types, surfaced billing delays by station and mode, and produced a daily DSO dashboard the finance team used every morning. The AI was good. The analysts trusted it. The dashboard was accurate. And the DSO had not moved in 18 months.
The problem was placement. Every insight the AI produced required a human to read it, decide what to do, and act. The exception was classified and then a person resolved it. The delay surfaced, then a person chased it. The AI had converted a slow System 2 process into a faster System 2 process. The human was still in every loop.
We repositioned the AI from an advisory layer to an execution layer. The same classification model that had been producing reports began triggering resolution workflows directly, routing each exception type to an automated action path matched to that category, without a human review step in between. We build the execution layer while your record layer stays the same. The TMS did not change. The ERP did not change. The AI moved from beside the process to inside it.
Within four months: 71% of exception billings resolved without human intervention. Invoice cycle time on exception shipments fell from an average of 8 days to under 18 hours. The DSO impact was immediate - the acquired divisions carrying 16 days of unbilled status came to under 6. The analytics layer that had been producing reports for 18 months without moving the P&L began moving the P&L within a quarter of repositioning.
(The invoice accuracy outcome of this repositioning - 95% touchless billing accuracy in production - is documented in the Invoice Reconciliation module on our logistics practice page.)
The CEO's Playbook: The placement audit
Before you commission another AI initiative, run a placement audit on the AI you already have. The question is not whether your AI is working. The question is where it sits in the workflow and whether it is removing decision points or informing them.
The Mandate: The AI Placement Test
For every AI tool currently deployed in your finance or billing operations, ask one question: "Does this AI remove a human from the decision path, or does it help a human make a better decision?" If the answer is the latter for every tool you have deployed, your AI investment is producing better System 2 decisions. It is not producing System 1 outcomes. And the difference is your entire working capital gap.
The Surgical Tool: The AI Placement ROI Formula
(Human decisions removed per month × Avg. resolution time
per decision × Analyst cost/hr)
+ (Days of billing cycle latency eliminated × Daily
revenue × Cost of capital)
= Monthly AI placement ROI
Advisory layer baseline (AI beside the process):
Human decisions removed: 0
Billing latency eliminated: 0
Monthly ROI: $0 in working capital
Monthly value: Faster analysis only
Embedded layer (AI inside the process):
Worked example — $300M freight book, 35% exception rate,
8-day avg. exception cycle, 12 analysts at $85K fully-
loaded:
Decisions automated: 15,000 invoices × 35% × 71% =
3,727/month
Time saved per decision: 45 min avg. → 3,727 × 0.75hr
× $41/hr = $115K/month
Billing latency eliminated: 8 days → 18 hrs = 6.7-day
improvement
Daily revenue: $300M ÷ 365 = $822K/day
Working capital recovered: $822K × 6.7 × 0.35 = $1.93M
Monthly financing at 6%: $1.93M × 0.5% = $9,650/month
Total monthly ROI: $115K + $9,650 = ~$125K/month
= $1.5M/year
Run your own numbers. The formula is the same. The only variable that changes the outcome is the placement decision - and that decision was made before the first line of code was written.
The Summary
Kahneman's System 1/2 framework is not a theory about AI. It is a theory about where cognitive work happens and what it costs. Applied to logistics AI deployment, it exposes a structural flaw in how most AI projects are scoped: they are designed to make human decision-making faster and better, not to remove human decision-making from the workflow entirely. The result is AI that is genuinely useful and completely invisible to the CFO.
The AI that moves your P&L is not more sophisticated than the AI sitting in your reporting layer. It is placed differently. It fires on its own. It does not wait for a review. It does not produce a dashboard. It produces a closed invoice and that is the only output your working capital can use.
Is your AI investment producing better analysts, or producing fewer decision points? Because your CFO can only measure one of those.
Next time in Part 4: You completed the acquisition. Integration is underway. Your CFO modelled the DSO impact. She was off by 60 days - not because the model was wrong, but because nobody priced in what happens when two billing cycles collide.