Order intake at Peter Fasano
It does the reading. You do the deciding.
Peter Fasano is a home-furnishings textile house. Orders arrive as email, typed, scanned, sometimes handwritten, and every one of them was entered into the ERP by hand, across modules that do not talk to each other. What got built is a reading layer between the inbox and the ERP. Every hour it reads the mail, extracts each order, checks and corrects it against the live catalog, and holds it for a person with every field traced to the paperwork it came from. It enters nothing into the ERP and decides nothing. It flags, never blocks.
What happened
It was live fast
First commit to a working demo on the client’s real orders: two days. The review board was already running at its own address during that first call.
The first flag caught a real rule
During the demo, a line priced under list disagreed with the catalog. Not an error: one showroom deducts its commission before invoicing, so entering the discounted price as written risks paying that commission twice. The system knew none of that. It just declined to stay quiet about a number that disagreed, which is the whole design. Flag, never block.
It says what it cannot see
Intake keys on attachments, so an ask typed into the body of an email was structurally invisible. Instead of pretending otherwise, a watcher was built for exactly that blind spot, and it found the asks people had been catching by hand. The gap was real, it was being covered by vigilance, and now the system names it instead of passing in silence. The same rule runs at line level: a price the catalog cannot check raises “No list price” rather than saying nothing.
It ran seven days with nobody watching
In August the laptop that publishes the review board travelled for a week, lid closed. The cloud half of the pipeline ran the whole week without a missed cycle. The board froze and said so: its freshness stamp told the stale truth all week, because the one thing a review surface must never do is look fresher than it is. On wake, a catch-up cycle pulled 32 missed messages through in six minutes.
The move to the cloud waited until the checks agreed
The laptop and cloud pipelines run side by side, and a comparator grades them against each other. Before cutover it caught a one-line disagreement, and checking rather than assuming found a defect on both sides: each had skipped a price check silently when the catalog had no price to check against. Both now say so out loud. The cutover was postponed on the record until the checks agreed, and the parity was re-verified across the full live queue in September 2026: 268 of 268 extractions resolve identically through both pipelines, and 135 of 135 report pages render identically from the same data.
It noticed how she works, and offered instead of changing
The operations lead clears her queue in two passes: the orders that passed every check first, then back for the flagged ones. The board noticed that pattern, deterministically, no model involved, and after three sessions offered it back as a saved sort with her name on it. Renamable, refusable, and only her default if she keeps it. The board’s own default stays chronological.
The posture, in one list
- It flags, never blocks. A discount can be legitimate. The flag says so, and the order moves on.
- Every automated correction is logged with before, after, and a reason.
- It says when it could not check.
- The freshness stamp tells the truth, including when the truth is stale.
- It stays quiet unless a person is needed.
- It observes, then offers. Nothing it learns becomes a default without a yes.
- Nothing trains on your data. Access is read-only until you say otherwise.
Where it stands
A live queue, checked hourly, unattended, at about five cents an order. Everything stays human-reviewed by agreement while trust builds. Writing back into the ERP comes later, with the same provenance, once the reading has been right for long enough.
Same six moves, pointed at your corner of the business.
Tell me how work reaches your team and I will tell you what a reading layer would catch. If that sounds familiar, let’s talk.