Sales Engineer for distributors
RFQ in, quote out, with every value traced back to the customer's own words.
Paste the enquiry, attach it or connect the inbox it lands in. Kabaido reads it into typed lines, matches each line against your own catalogue with a fit score you can inspect field by field, and hands you a draft quote. Nothing is invented: a value the request never gave comes back as a question, not a guess.
The job as it stands
The quote takes twenty minutes. Getting to it takes the morning.
A distributor's inbox is the bottleneck, and it is a typing bottleneck rather than a thinking one.
Fourteen enquiries before ten o'clock. One is a paragraph with three part numbers buried in it, one is a spreadsheet with columns nobody agreed on, one is a scanned schedule with a handwritten quantity in the margin and one is the fourth message in a thread where the specification changed twice. None of them is a line sheet. Every one of them has to become one before anybody can price anything.
So somebody retypes. Part reference into the search box, hope the customer's notation matches yours, work out whether that figure is a diameter or a length, decide whether the grade they wrote is the grade you stock, notice that nobody said how many. Then the same thing on the next thirteen. The skill in the job is knowing what to offer and what to charge. The morning goes on getting the request into a shape where that skill can start.
The value is in the answer. The cost is in the typing, and typing is the part a machine should be doing.
That is what this is for. It reads the request you were actually sent, sets out every line with each value tied to the text it came from, searches your catalogue and scores what it finds, and asks about the things the request does not say. What you offer and what you charge stay yours.
What comes off the enquiry
The values you would have retyped
Each one is captured with the slice of the customer's own text it came from. A value the request never gave stays empty with a question against it rather than being filled in.
Part and drawing references
A part number, a manufacturer reference, a drawing number and its revision, all carried word for word. A reference is identity, so it is never tidied, expanded or corrected on your behalf.
The technical specification
Diameters, lengths, thicknesses, threads, grades, coatings, tolerance classes and fit classes, each one against the pack key it belongs to, with the customer's original wording held beside the normalised figure.
Units, normalised, original kept
Millimetres unless the value is inherently something else, microns kept as microns and never quietly turned into millimetres, and an imperial size converted exactly with the wording it came from still visible.
Quantities and the call-off
How many, in what unit, and the schedule where the enquiry gives one. A quantity nobody stated stays unstated all the way to the quote rather than becoming one on the way.
The commercial frame
Required-by dates, delivery terms, a reference to an existing agreement and the customer's own contact details from the signature, each read as context rather than as a request line.
The application, where it is stated
The material being cut, the operation, the machine. Stated application does not become a specification: it feeds a separate recommendation with the engineering reason written out, which you accept or ignore.
Can you quote the following please, needed by 26th June: 10 off 10mm 4FL carbide endmill AlTiN coated, LOC 30 min
RFQ-2026-aero.pdf- L110×10 mm 4 flute carbide endmill, AlTiN, LOC 30 mm min
- L25×6 mm 2 flute ball nose endmill, uncoated
- L320×6.35 mm (1/4 in) 3 flute square endmill, TiAlN
- L44×ER32 collet 6 mm
Every value cited. Missing details become questions, never guesses.
Matched, priced and quoted nextA fit of 91 is a number you can take apart: field by field, which value matched, which was close, which the catalogue never stated.
Matching
The number is arithmetic, not an opinion
A line with a part reference in it takes the short path: a direct lookup, and an exact hit or a similarity above 0.92 locks the product and skips the search entirely. That path costs no model call at all, which is why a distributor's simplest enquiries are also the cheapest ones.
Everything else goes to search. Candidates come back on two channels: full text over the title, part number, category and every typed attribute value, and a similarity pass on the part number and the title, so a different punctuation style or a mistyped character still finds the row. Hard attributes become filters, and the top twenty five are then scored in plain TypeScript against your catalogue schema. Each field the request stated contributes its weight times how well the candidate satisfies it: an exact value scores 1, a value inside the tolerance band scores 0.85, a documented equivalent scores 0.6, a field your catalogue leaves blank scores 0.35 and a genuine mismatch scores nothing. A hard field that misses disqualifies the candidate outright. Where no tolerance is configured for a numeric field the band is five per cent.
The fit is the weighted total on a scale of 100, and every term is on the card. You can see that the diameter was exact, the length was one size up and within the overshoot allowance, the coating was a documented equivalent rather than the one asked for and the flute count is simply not populated in your data. No part of that number comes from a model.
What the score decides
Four outcomes, and only one of them is automatic
The thresholds are fixed and the same for everyone, so a match never quietly gets more confident on a busy day.
- High match: chosen for you
- A fit of 86 or better with at least 8 clear of the runner up. The product is selected and priced into the draft. Four things veto it however high the fit: a substituted dimension, a candidate that contradicts something the platform assumed, a match carried mostly by assumptions, or an application rule that says this candidate is wrong for the stated job. Each one drops the line to review for a human to confirm.
- Review: the top three, side by side
- A fit of 60 to 85, or a close second. You get the leading three candidates with their field breakdowns next to each other and you choose. Your choice is written back to the line, so the quote carries what you picked rather than what the score preferred.
- Closest: the nearest thing you stock
- A fit of 40 to 59. Below the review floor, but there is still a nearest viable item and hiding it helps nobody. It is surfaced with its substitutions and assumptions written on it, and it is never auto-priced into a total. If it reaches a quote it reaches it as an unpriced line for someone to confirm.
- No match: still a line
- Below 40, or nothing in your catalogue is in the right area at all. The line does not fail. It arrives typed, unit-normalised, cited and unpriced, waiting for your figure, which is also exactly what happens on a free account that has imported nothing yet.
Codes and designations
Six places a plausible guess would cost you money
A quote is a commitment. At each of these the honest answer is to carry the customer's characters or to ask, so that is what happens.
It decodes a code, it does not extend one
An ISO 1832 insert code or an ISO 5608 holder code is its own source: every character is in what the customer wrote, so reading CNMG120408 as an 80 degree rhombic negative insert with a 0.8mm corner radius is citing, not inventing. Only the positions actually written are decoded, and the full code is kept verbatim alongside them.
It never fills in the positions nobody wrote
A grade, a chipbreaker, a coating or an ISO material group that the code does not carry stays absent. A round insert gets no corner radius decoded from a code that cannot state one. Absent means asked about, never inferred.
It never converts a grade on your behalf
The designation the customer wrote is the designation that travels. Where a standard cross-reference is documented, EN24 and 817M40, 304 and 1.4301, 316 and 1.4401, it is offered as an equivalent for you to accept, scored below an exact hit and shown as an alternative rather than swapped in underneath you.
It never cross-references a carbide grade
No vendor grade is mapped to another vendor's grade, to a coating or to an ISO group beyond its own leading letter. There is no authorised table for it, so there is no feature for it. An unrecognised grade is carried word for word and asked about.
It never turns a call-out into a number
A fit class, a tolerance class or a general tolerance call-out is captured as the literal string the drawing used. It qualifies a nominal dimension and never merges into it, and it is never interpreted into a figure the customer did not state.
It never invents a price or a quantity
No rate, no unit price and no quantity is produced out of nothing. An unpriced line waiting for your figure is an honest line. A plausible one is a liability with your name on it.
When the enquiry is thin
It asks. It does not fill the gap in.
The rule the whole platform runs on is cite or abstain, and on a quote that rule is worth money.
Every value carries the span of source text it came from, and a value that is not in the source is not produced. A thin enquiry therefore produces a thin line with specific questions against it rather than a confident line that is wrong. A missing attribute renders as missing, in its own colour, with nothing behind it.
The questions sit against the line they belong to and they are the ones you would have asked. Which grade. Is that a cut length or an overall length. How many. Is there a drawing, because the message mentions one and nothing came with it. They come back as a list you can send straight on, so the chase email more or less writes itself.
Where the platform does put a value forward it says so plainly. A parameter the stated application settles is drawn as a recommendation with the engineering reason and the words it reasoned from. A parameter it has had to assume so the line can be matched at all is drawn as an assumption with its basis, capped at three per line, never covering a price, a quantity, a date or a part reference, and always reversible in one tap. An assumed value can never score as high as a stated one, and a candidate that contradicts an assumption is sent to review rather than being disqualified, because a hard gate is earned by what the customer said and not by what we supposed.
Cite or abstain
The quote
A draft you finish, not a document you retype
The resolved lines land on a draft quote in the order they arrived. A matched product carries your catalogue price and any customer-specific pricing that applies. An unmatched, closest-band or unselected review line lands as a manual line carrying the description it was read from, ready for your figure. Every line keeps its link back to the request it came from, so anyone can see where it started.
A line the customer never gave a quantity for is the one case worth spelling out. It commits at one, unpriced, with the words for the missing quantity on the description itself and the placeholder recorded as an assumption, so nobody can send it without setting both. It is deliberately awkward, because the alternative is a plausible total that nobody checked.
From enquiry to order
Where the reading sits in the day
- In the inbox, if you want it there
- Connect a Gmail or Microsoft 365 inbox and new enquiries are read as they arrive, checked every five minutes, with the reply drafted in your branding on the original thread. Review mode is the default, so nothing goes out until you say so, and mail that arrived before you connected is never touched. How the email agent works
- The purchase order that comes back
- A PO is read on the same path and becomes a draft order rather than a quote. It carries no agreed price, whatever a template may imply, so every figure on the committed order is yours at the moment it was raised and each line records that basis rather than pretending the customer set it.
- The customer, the history and the paperwork in one place
- An accepted quote becomes an order without anything being re-keyed, and the account, the previous quotes and the documents sit alongside it rather than in three different folders. See the Sales Engineer
The connected inbox
Reading the mail without answering the wrong mail
An agent in a real sales inbox is only safe if it is careful about what it refuses to touch.
Classification is heuristics first and a model only where the heuristics genuinely cannot decide, which keeps the ordinary day deterministic and cheap. Before any of that, the loop guards run: your own sent mail, bounces and out-of-office replies are never answered, because answering an autoresponder is how mail loops start.
Then your own filters. Include and exclude lists on senders, domains and subjects decide what the agent is allowed to look at in the first place, and list mail and newsletters are recorded but never processed. What is left is scored as an enquiry, a purchase order or a technical question, and each one routes to the mode that suits it.
Connecting is an organisation admin action with full provider consent, and review mode is the default: every draft goes to an address you choose until you decide it can reply directly. Nothing that arrived before you connected is ever read.
Being straight about it
Three things this is not
Worth knowing before you send anything, because a landing page that leaves them out is one you find out about later.
It is not a price list and it does not know what your market rate is. There is no built-in library of prices, no benchmark and no suggestion of what a line ought to be worth. It prices from your catalogue, your customer pricing and your discounts, and where none of those apply it hands you the line unpriced.
It is not CAD and it does not read geometry. A DXF, a DWG, a STEP file or a photograph of a drawing is not interpreted as a model: nothing is measured or priced from one. It reads requests, including a PDF whose pages are scans, and an image that arrives is reported back as not read rather than quietly skipped.
It is not a replacement for your judgement about what to offer. Every band above the automatic one puts a person in the loop on purpose, and the automatic one is deliberately narrow. A tool that matched everything confidently would be quicker and considerably more expensive.
Against the inbox and the search box
Where it helps, and where it does not
A well-run inbox and a good search box are a real answer. This is what changes and, just as usefully, what does not.
| Task | Inbox and search box | Kabaido |
|---|---|---|
| Reading the enquiry | You | Read for you |
| Retyping the lines | By hand, every time | Set out from the source |
| Checking a value against the customer's words | Scroll back through the thread | Every value shows its source text |
| Finding the product | One search box, one line at a time | Every line searched and scored |
| Knowing why this product | In the estimator's head | Fit score, field by field |
| Spotting what the enquiry does not say | If you notice | Listed as questions per line |
| Designations and standard codes | Retyped, sometimes wrong | Carried verbatim, decoded only where written |
| Your prices and your discounts | Yours | Yours |
| Deciding what to offer | You | You |
| Producing the quote document | Typed again | Generated, branded, with a portal link |
| The revised enquiry next week | Start again | Read again against the same lines |
| Reading geometry out of a CAD file | Your CAD | Your CAD |
Getting started
Test it on an enquiry you have already quoted
No card at signup, and the free plan is not a trial.
Sign up and open a new request
The request screen is the first thing you see. Nothing has to be imported or configured before it will read something.
Paste one you know the answer to
An enquiry you quoted last month, in the state it arrived. Compare the lines it sets out against the ones you typed.
Import your catalogue and run it again
A CSV with your own headings. Now the same enquiry comes back matched and priced, and the fit scores tell you where your data is thin.
On the call we run one of your own enquiries through it with you, against your own catalogue, and you keep whatever comes out of it.
Questions
What distributors ask first
- Paste the text of the enquiry into a new request, or attach it as a PDF up to 4MB, including a PDF whose pages are scans. Connect a Gmail or Microsoft 365 inbox and the wider set of things customers really send arrives with it: spreadsheets, CSV files, documents and forwarded messages, each parsed into the same lines. Images are not read, and you are told so rather than left with a silent gap. A CAD file is not read as a model either: nothing is measured or priced from one, so send the written request alongside it.
- Better data gives better matches, and the fit breakdown tells you exactly where yours is thin: a field your catalogue leaves blank scores well below one the customer's value hits, and it shows on the card as unknown rather than as a miss. You can start with a description and a price and add attributes as you go. Nothing is blocked while the data is incomplete, because an unmatched line still arrives typed and cited.
- Because the number is arithmetic you can take apart. Every field the request stated appears on the card with what was asked, what the candidate carries and which state it scored: exact, within tolerance, a documented equivalent, unknown on your side or a miss. A match only selects itself at 86 or better with a clear lead of 8, and a substitution or an assumption vetoes even that and sends the line to review.
- It still commits. The line arrives typed, unit-normalised and cited, carrying the description it was read from and the quantity that was stated, and it lands on the quote as a manual line waiting for your figure. That is also what a brand new account sees before anything is imported, which is why the free plan is usable in the first minute.
- You get a short line and a specific question rather than a full line that is wrong. A value with nothing behind it in the source is shown as missing, and the questions sit against the line they belong to so they read as a list you can send back to the customer as it stands. A quantity nobody stated stays unstated: if such a line reaches a quote it arrives unpriced and says on its face that no quantity was given.
- Yes, on the same reading path, and it routes to a draft order rather than a quote. One thing is worth being clear about: an inbound purchase order does not hand us an agreed price, and the extraction never captures one. Every figure on the committed order is our own price at the moment it was raised, and each line records that basis so nothing downstream can call it agreed.
- No. Pasting an enquiry into the app is the whole flow and it needs no connection at all. Connecting an inbox is what lets the reading happen without anyone copying anything, and an organisation admin does it with full provider consent. Review mode is the default, so every draft goes to an address you choose until you trust it to reply directly, and mail that arrived before you connected is never touched.
- No. Your commercial data is not used to train models and the AI provider is configured for zero retention, so what you send is used to answer you and nothing else. Every row is isolated by database-enforced row level security, hosting is in the UK, and the detail is set out on the security page along with the data processing agreement.
- The Free plan carries the whole reading flow with one seat and no card at signup, and includes 25 credits a month. A credit is one resolved request line: a matched catalogue line is one credit, a structured line with no match is one, a servicing line is two, and each scanned page beyond the first is one more. Scanned pages are also capped per request: two on Free and twenty on a paid plan. Manual work costs nothing: searching, pricing a line, editing a quote and sending it are never metered. Starter is £149 a month and lifts the seat and product limits.
Send us the enquiry you are quoting today
Paste it in as it arrived and watch the lines come back typed, cited and scored against your own catalogue. Free, no card, and it works before you import a single product.