94% Accurate, 40% Automated: The Gap That Eats Your ROI

The demo posted the order. Tuesday didn't.

The pilot report says 94% extraction accuracy. Everyone signs off.

Six weeks later the same two people are still in the shared inbox at 8am. Not retyping — reviewing. Every order lands in a queue with the data pre-filled, and someone clicks approve. Faster than before, sure. But the headcount didn't move, and the headcount was the business case.

Nothing broke. The system does what it was sold to do. The problem is that accuracy was never the number that paid for it.

The number that pays: zero-touch rate

Zero-touch rate is the share of orders that go from arrival to ERP record with no human involved. Not "a human glanced at it." Nobody.

Accuracy and zero-touch rate are different questions:

• Accuracy asks: when the system extracts a line, is the line right?
• Zero-touch rate asks: how often was the system confident enough to act alone?

A pipeline can be right about everything it commits to and still send 60% of orders to review, because it isn't sure. High accuracy with a low zero-touch rate is a system that has learned to hedge. Your savings live entirely in the orders nobody opens.

Run the arithmetic. Take 120 orders a day at 6 minutes each of manual entry — 12 hours, roughly two full-time people. Now automate with review on every order at 45 seconds a look:

• 40% zero-touch: 72 orders reviewed, 54 minutes a day. You saved most of the labour but you still need somebody in that seat every morning, which means you still pay for the seat.
• 85% zero-touch: 18 orders reviewed, 14 minutes a day. That is an exception process, not a job. It absorbs into someone's existing role.

Same accuracy in both cases. The difference between a line item and a salary is the routing decision, not the extraction.

Why pilots land at 40%

Three reasons, in the order they usually bite.

The confidence threshold was set during the honeymoon. Nobody wants a wrong order posted in week one, so the threshold goes conservative and then never gets revisited. It quietly becomes the ceiling on your ROI. Reviewing it monthly against actual correction data is fifteen minutes of work that most teams never schedule.

Catalog matching carries the whole load. "The usual," "the small mozzarella," a customer's internal part number that matches nothing in your system — a model reading the order text alone has to guess, and correctly refuses to. The same model with that customer's last twenty orders in front of it resolves the shorthand and commits. Most stalled deployments are not model problems. They are context problems.

Two customers generate most of the exceptions. Order formats follow a power law. In a stalled pipeline the review queue is usually dominated by a handful of accounts with genuinely awful order habits — a photo of a handwritten fax, a spreadsheet with merged cells. Fixing those specific accounts moves the aggregate number more than any amount of general tuning.

What good looks like

In our food distribution case study, order entry went from about 6 minutes per order to under 30 seconds of human attention on the roughly 15% that needed review. That 15% is the figure to interrogate, not the 30 seconds. It means 85 of every 100 orders were posted by the pipeline alone.

Ask any vendor — us included — for that split before you ask about accuracy:

• What share of orders posted with zero human involvement in month one? In month six?
• What is the confidence threshold, who owns it, and how often is it reviewed?
• When a low-confidence order goes to review, does the reviewer correct a field or re-enter the order?
• Which accounts generate the exceptions, and what is the plan for those specific accounts?

A vendor who can only answer the accuracy question is describing a model. You are buying a process.

Measure it from day one

Zero-touch rate is trivial to instrument and almost nobody does it. Every posted order gets a flag: human touched, or not. That single boolean, plotted weekly, tells you whether the deployment is improving or has quietly plateaued at the threshold somebody picked in week one and forgot about.

Plot it. If the line is flat after month two, the problem is the threshold or the catalog context — not the AI.

The full mechanics of capture, extraction, validation and ERP writing are covered in AI order entry automation for distributors, and channel-specific handling in WhatsApp order automation for wholesale. If you want the pipeline itself, see AI order processing.

Accuracy tells you whether the system is right. Zero-touch rate tells you whether it was worth buying.

Want the zero-touch estimate for your own order mix? Book a free consultation — or see what AI automation costs first.