Work samples: what the AI Implementation Audit actually produces

I count two kinds of rule in every workflow.

A hard rule is enforced by a system. The TMS will not let you tender a load to a carrier with no insurance certificate on file. You cannot skip it, so it does not decay.

A soft rule is a step a person has to remember. Nothing blocks you, nothing errors, and skipping it produces no visible signal at the moment you skip it. "Check the customer's accessorial history before you quote" is a soft rule. So is "read the whole thread before you write the dispute note."

The model: assume each soft rule gets followed about 90% of the time when the person is under time pressure. That is a generous assumption, and it is an assumption, not a measurement of your team. Nobody here is ignoring rules. What decays is the probability that a single run of the workflow follows all of them at once, and it decays exponentially with rule count. Six soft rules gives you 0.9⁶≈ 53.1% fully clean runs. Thirty gives you 0.9³⁰≈ 4.2%.

Two honesty notes before the table.

A run that is not "fully clean" is not automatically a bad quote or a wrong invoice. Most of the time the missed step did not matter. What the clean-run rate measures is how much of the output's accuracy is resting on somebody's memory instead of on the process. That is the exposure, & it is the exposure AI multiplies, because AI fills a gap confidently instead of leaving it blank.

The one number I would argue about is the quote desk's per-rule assumption. I dropped it from 0.90 to 0.85 there because every step on that desk competes with a customer on hold. That is a judgment call from four interviews, not a measurement. If you think 0.90 is right for your estimators, use 0.90 and the workflow still lands at the bottom of this table.

This is the single soft rule at Meridian with the highest dollar consequence and the lowest observability. When an estimator skips it, the quote goes out looking completely normal. The consequence arrives four to six weeks later as an invoice dispute, on a different person's desk, filed as a billing problem.

I traced it. Of the 61 invoice exceptions in the sample week, 9 came back to an accessorial charge that was already in that customer's own history and was not in the quote. Average value $455, so about $4,095 in one week. I am not going to multiply that by 52 and hand you a scary annual number from a single week of data; take it as a rate worth measuring rather than a finding. What matters is the shape: the money leaves at the quote desk and the pain lands in billing, so the two teams have never seen it as the same problem.

The AI angle makes it worse in a specific way. The CRM add-on now auto-fills the customer's "usual lanes" panel. Estimators told me they read that panel instead of opening the account history, which is exactly what it is for. But the add-on summarizes shipping patterns, not billing exceptions. It is confidently answering a question nobody asked it.

A biller opens a disputed invoice, Copilot summarizes the email thread, & the biller writes the dispute note from the summary. In the six weeks since licenses landed, this has become the default on all four billing desks. Nobody decided it. It just happened, because it saves about eleven minutes per exception and 61 exceptions a week is 42.7 hours of work across four people.

The problem is not summary quality. The summaries I read were good. The problem is that a summary is a lossy compression, which is fine for orienting yourself and dangerous for attribution. A dispute note is pure attribution: who agreed to what, on which date, in which message. When the compression drops the one sentence where the customer's own dispatcher approved the detention, the note reads perfectly and the argument is unwinnable.

There is no error state for this. The tool never says it left something out.

The chain is: carrier gives a verbal update on a check call, transcription tool writes the note, Copilot drafts the customer email from the note, rep sends. The ETA in the customer's inbox has now traveled through two AI tools and never touched the TMS field that the rest of the company plans against.

I spot-checked 12 of the 406 status emails from the sample week. Three carried an ETA that did not match the TMS at the moment of send. Twelve is a small sample and I will not claim a rate from it, so treat it as a reason to look rather than a measurement. Two of the three were off by under two hours and nobody would have cared. The third was a next-day delivery described as same-day.

This is the failure mode I would flag even if the other two were clean, because it is the one where the tool works exactly as designed and the customer gets the wrong answer anyway.

Eight steps, five of them enforced by the TMS. You cannot tender a load to a carrier without a DOT number, a current insurance certificate, a W-9, a signed carrier packet, & banking details on file. The system blocks it. Those five steps cannot decay under pressure because pressure has nothing to push on.

That leaves three soft rules, which is why this workflow scores where it does. Somebody built those gates years ago, probably after getting burned, and the gates are still doing their job. Copilot drafting the welcome email touches none of it.

Two soft rules, a locked template, and six people who read it every Monday morning. If the numbers are wrong, three of them will say so before lunch. Copilot writes the narrative paragraph on top of a TMS export it cannot alter.

This is the profile of a workflow that is safe to point AI at: short remembered chain, fixed structure, & fast visible correction when it goes wrong. Loud failures are safe failures.

Compare the two lists and the rule writes itself. The workflows that hold are the ones where a system holds the facts and a person supplies the judgment. The ones that collapse are the ones where a person holds the facts in their head, and now an AI tool has offered to hold them instead.

Carrier onboarding has five gates and three memories. Quote intake has zero gates and eleven memories. That difference, not tool choice, is what separates them.

Three things follow from this page, and all three are in the week 2 plan with sequencing and effort estimates.

First, Meridian's AI problem is not a tool problem. The three tools that were purchased are reasonable tools. Two of the five workflows are already fine with them. Swapping vendors would change nothing on this page.

Second, the highest-return work available to you costs no money. Four of the eleven soft rules at the quote desk correspond to fields that already exist in your TMS quote form and are not marked required. Making them required moves that workflow from 0.85¹¹ = 16.7% to 0.85⁷ = 32.1%, which takes your predicted clean quotes from roughly 53 a week to roughly 102. That is one afternoon of admin work and no spend.

Third, the ordering matters more than the list. Doing the invoice-exception fix first would be a mistake, because a meaningful share of those exceptions are created upstream at the quote desk. Fix the quote form, watch the exception count for 30 days, then size the billing work against what is left.

The plan lands at the end of week 2 with all three sequenced, plus the one-pager for whoever has to approve it.