For Cutting tool and metalworking distributors
Cutting tools: the technical call and the commercial one get made in the same twenty minutes.
A customer asks for a designation you do not sell, somebody spends twenty minutes reading across to what you stock, and then prices it — two different decisions made by one person at speed, and only one of them has anything checking it.
A working session on your numbers, not a slide deck. No account, no card, and nothing connected until you say so.
Four things that go wrong on a cutting tools book
Named rather than categorized. “Margin erosion” is a heading a reader cannot check against their own week; each of these is something that happened on a Tuesday, and if none of them is familiar then this is the wrong page and the rest of it will not improve.
Problem one
The twenty-minute line
A customer asks for something you do not stock, under a designation you do not sell. Somebody opens a conversion guide, reads across to an ISO application position, checks the edge length and the corner radius, and forms a view. It is skilled work and it is correct most of the time. It is also done one line at a time while the quote waits, and on a twelve-line enquiry it is a morning. And twenty minutes is the good case. The bad case is the enquiry that waits until the one person who does this reliably is free, by which time the customer has an answer from somebody else.
- A published cross-reference is a set of pairwise claims, not a lookup
- Two grades at one application position are comparable, not identical
- Substrate, coating and edge preparation still differ between brands
- Twenty minutes is the good case; the queue is the bad one
Problem two
One head, two decisions, one of them checked
The person who decided the part is comparable also decides what it is worth, in the same moment, with the same urgency. The technical judgement gets reviewed by whoever queries the part later. The commercial one is reviewed by nothing at all, because a substitution is recorded as an ordinary sale of an ordinary item. So the review that would have caught a pricing error is a technical review, and it asks a technical question. Somebody checks that the part fits; nobody checks what it earned.
- Your ERP cannot tell a substituted line from a stock line
- So the substituted half of the book has no margin figure of its own
- The two decisions have different reviewers and one of them is nobody
- A technical review asks a technical question
Problem three
Application knowledge sits with one or two people
There is a person who can do this reliably, and on a good day there are two. When they are on holiday the desk either guesses or the enquiry waits until they are back. Both answers cost the order, at different odds, and neither appears anywhere as a cost of anything. It is also why quoting capacity in this trade does not grow with headcount. Adding a salesperson adds somebody who can take the call, not somebody who can answer it.
- The queue behind that person is invisible in every system you own
- A guess made in their absence looks identical to their own work
- Nothing records which of the two answered a given line
- Quoting capacity does not grow when you add a salesperson
Problem four
The reasoning goes home at five
Three months on, a customer asks why they were quoted this part at this price. The chart that was consulted, the position that was matched and the judgement that closed the gap were never written down — they were in somebody's head, and that person is on another call or at another company. What you can reconstruct is the price. The same gap makes the work unteachable. A new hire can be shown the chart and cannot be shown the fifty judgements made around it, because none of them was written down at the time.
- The substitution decision leaves no record of its own
- Neither does the pricing decision that rode along with it
- A rebuilt catalog changes the answer and nothing says which edition answered
- And it cannot be taught, because it was never written down
What it costs is the way it accumulates
The loss is not a discount. It is that a substitution is a pricing event and gets handled as a fulfilment one.
How it compounds
A read-across lands on a part you stock and a number goes on it in the same breath — very often the price of the designation the customer named, or that price less something, because that is the figure in front of the person deciding. The part you actually supply has its own cost, and nothing in the moment compared the two. Repeat it for a year and the substituted lines are a slice of the book whose margin nobody has ever looked at separately, because nothing in the ERP marks them as a slice.
Why nothing catches it
A win looks like a win. The enquiry was answered the same day, the customer bought, the line shipped, and the only thing wrong with it is a number that was never checked against anything. That is precisely the shape of failure this platform is built around — a defensible answer with nothing behind it — and it is why the technical half of what PIE does refuses to guess as hard as the commercial half does.
No figure on this page is a claim about your book. There are no industry averages here and no recovered-margin percentages: the only numbers on the page are on the card further down, and they are sample figures re-derived by a test from the operands the card itself shows.
What a cutting tools desk does about it today
Not a list of mistakes. Every one of these is a sensible response to a real constraint, which is exactly why it has survived — and the last one is usually the most rational of them. A product that does not understand why the current answer is the current answer has nothing useful to replace it with.
A conversion chart, printed
A manufacturer's cross-reference sheet, usually a generation old, on a wall or in a drawer. It answers the technical question at the application position, which is the half it is for, and says nothing whatever about what either part costs you.
The one person who knows
An applications engineer, or an inside salesperson who has sold carbide for twenty years. Right nearly every time, and a single point of failure for both halves of the decision at once.
The price of the part they asked for
The commonest shortcut, and an understandable one: quote the substitute at the price of the designation the customer named, because that is the number already on the screen. It is correct whenever the two cost you the same, which is not often and is never checked.
A margin report, monthly
Correct numbers, after the fact, averaged over a book in which the lines you want to look at are not distinguishable from the lines you do not.
Where PIE intervenes, and at which moment
Every figure below is deterministic arithmetic over rows your ERP already wrote, stamped with the version of the margin policy that judged it. The AI on this platform reads those numbers and phrases them; it never produces one. What is not listed here does not apply to a cutting tools book, and the limits further down say which of those are permanent.
Your designations, read rather than matched
This is the one catalog the engine understands as more than text. ISO designations and grade systems decode into typed fields — shape, clearance, tolerance, edge length, thickness, corner radius, chipbreaker, grade — each carrying provenance, a confidence and the character span it was read from. An unrecognized token is captured verbatim rather than guessed into a field, and identical input produces byte-identical output.
Alternatives ranked on what was decoded, not on what looks similar
A shape or polarity mismatch excludes a candidate outright. Dimensions are graded by distance. Coating, chipbreaker and flute count nudge and never dominate. Where nothing discriminates between the candidates the engine abstains and says so, rather than returning the least-bad one — which is the answer your applications engineer would give, and the one an order-automation tool cannot.
A cross-reference stays a starting point
A score here is policy under your own equivalence bands: true of this quote, read per request, versioned, and never stored as a fact about the two products. A scored suggestion is never promoted to a confirmed identity, because an equivalence does not compose — two hops of a tolerance band put a 0.4 mm corner radius and a 0.8 mm one in the same class. Picking a different part is a substitution on one quote and stays one.
The floor, on the line you resolved
Once the line resolves, the price typed on it is checked against the policy you set, and a breach is held for a named approver with the rule that stopped it named. The substitution and the price are two records from here on, which is the whole point: one of them can be questioned without reopening the other.
Why you quoted what you quoted, last March
Every emitted field carries where it came from. Every computed row carries the version of the policy that judged it. And the catalog's own ruleset version is in the record too, which is what explains why the same text resolved differently before you rebuilt it — a rebuilt catalog decoding differently is a fact, not a discrepancy.
One quote line, start to finish
One line, from a competitor's designation arriving to the record that explains it next March. Sample figures on the card; the arithmetic under it is what the platform applies.
CNMG 120408
200 units for A repeat account in Ohio, quoted at $41.20 — routed to a sales manager, no individual owner. The platform holds it, not the salesperson.
| Quoted unit pricethis quote | $41.20 |
| Effective unit costbill | $38.08 |
| Margin flooryour margin policy | $44.80 |
| Recommendedthis customer’s own history | $48.70 |
| Margin at this pricecomputed | 7.6% |
The manager’s decision card, as the product draws it — with sample figures, because no customer’s numbers belong on a public page.
A designation you do not sell arrives
The enquiry is pasted as it came. The engine decodes the code into typed fields — shape, clearance, edge length, thickness, corner radius, grade — each with the span of characters it was read from, and captures verbatim anything it could not name.
Candidates are ranked, or refused
Your own decoded catalog is scored against the request. A shape or polarity mismatch drops a candidate; dimensions are graded by distance; coating and chipbreaker nudge. Every comparison's left operand is the request, never another candidate. If nothing discriminates, the engine says so and the line waits for a person.
A person picks, and the pick stays a pick
The applications engineer confirms one candidate for this quote. That is a substitution on one quote, recorded as one; it does not become an asserted identity between two products, and nothing downstream will resolve a later enquiry off it.
The price on the resolved line meets the floor
Cost, divided by one minus your margin floor, on the part you are actually supplying rather than the one the customer named. On the card the price asked for is below it, so the line routes to a manager with the rule that stopped it named. The desk never sees the cost.
Next March, the whole chain is one record
The decode and its spans, the candidates and their scores, the person who picked, the policy version that judged the price and the catalog ruleset version that produced the decode. The question is answerable without the person who answered it the first time.
Sample figures throughout. No customer’s numbers appear on a public page, every figure on the card is re-derived by a test from the same operands it shows, and the floor is cost divided by one minus the margin floor your policy sets — nothing rounded in your favor, because a floor rounded down is a floor that has moved.
The limits, before you find them in week three
Every item below is a thing this platform does not do on a cutting tools book today. They are on the page because a capability list is not what a distributor who has survived one implementation believes — and because the alternative is you discovering them after a pilot.
- Special pricing agreements and manufacturer rebates are not modelled. Margin is computed from the cost on the AP-invoice line and nothing is claimed back afterwards.
- The resolution engine carries no price and no stock. It deals in nomenclature only; availability, cost and list price are read from your books afterwards and keyed by the manufacturer part number it returned.
- Nothing resolves until you have uploaded a price list per manufacturer you sell and confirmed how each one is read. The corpus that ships with the engine is a seed for the first company, not a catalog of your book.
- Quotes are not imported from any ERP, so a win rate has no denominator until you start quoting here.
- An equivalence is never composed. PIE will not reason from one scored match to a second — every comparison is against the request, because two hops of a tolerance band put a 0.4 mm corner radius and a 0.8 mm one in the same class.
Why it can be trusted
The AI never computes a single number.
Your business data and your rules determine the number. AI explains it.
Same inputs, same answer, every time — with a paper trail.
- Deterministic calculations — every figure is arithmetic on your own records; turn AI off and every number still works
- Auditable decisions — each number names the policy that produced it, so a price you quoted last quarter still explains itself
- Your data — read from your own books, used for you alone, exportable and erasable on request; AI runs on your own account
- Your rules — you set the floors and thresholds, and PIE holds every quote to them
- Provenance — every row carries the connector, the connection and the id it came from, so nothing in PIE claims a source it did not have
What cutting tools distributors ask first
Which manufacturers' nomenclature does PIE decode?
Whichever ones you upload a price list for. Each list is analysed on its own to find the shapes of description it contains, a person confirms what the varying parts mean, and the catalog is built from that. There is no shipped list of supported brands and no default decoder — a decoder that guessed would be a wrong number with a real provenance stamp on it.
Will it pick a substitute for me?
It will rank candidates on the attributes it decoded and show which field contributed what. It will not decide. Where nothing discriminates it abstains, and a scored suggestion is never written down as an identity — that stays a person's call, which is also how your applications engineer would want it.
Can a salesperson see cost or margin?
No. The server omits those fields from the response rather than the screen hiding them, and a rule whose boundary is cost is withheld too. The desk gets the floor, the recommended price and this customer's own history — enough to negotiate, without the cost basis.
Does the AI read the enquiry?
The resolution does not. It is a deterministic parser with a versioned rule set, and identical input produces identical bytes — which is what makes a resolution auditable months later. A model may phrase what was found; it never decides what was found and never produces a number.
What happens when we rebuild the catalog?
Resolutions carry the catalog's ruleset version, so a line that resolved differently before the rebuild can say which edition answered it. That is deliberate: a rebuilt catalog decoding differently is exactly the fact that explains an old answer.
Which of the seven this trade runs
Prophet 21 runs most of this trade as well, though Epicor files it under industrial rather than naming it — with Business Central and Acumatica common in the mid-market and Zoho Books at the small end. The decoding is identical on all seven; what changes is what each book can tell you afterwards.
Your book already knows where the margin went.
30–45 minutes, your own numbers on the screen, and an honest answer about what PIE can and cannot see on a cutting tools book.