AI for chemical quoting: what to automate and what to keep human
Where AI helps at a chemical distributor's quote desk, where it is risky, and ten questions to ask before you choose a quoting tool.
On this page
- What are the three jobs at a quote desk?
- Where is AI strong? Reading messy requests
- Where is AI risky in chemical quoting?
- Grade and concentration
- Units and density
- Prices and discounts
- Same as last time, and invented values
- What do European chemical distributors report?
- Ten questions to ask before you choose a quoting tool
- What are the options? A category-by-category view
- Where does Vuntus stand?
- When Vuntus is not the right fit
- Frequently asked questions
- Sources
Quoting at a chemical distributor is three jobs: reading the request, finding the product in your own catalog, and taking the price from your own price list. Reading messy text is a good fit for AI. Product and price decisions need a visible source for every value, a system that stops when something is unclear, and a person who approves.
This article is for owners and sales managers at chemical distributors who are weighing up AI for the quote desk. It covers where AI genuinely helps, where it is risky in chemistry, ten questions to put to any vendor, and an honest map of the options.
What are the three jobs at a quote desk?
| Job | What it involves | Where mistakes come from |
|---|---|---|
| Reading the request | Turning an email, PDF, spreadsheet or WhatsApp message into lines: product, quantity, unit, concentration | Free typing, abbreviations, several products in one sentence, “1.000” in a spreadsheet from a customer who uses a decimal comma |
| Matching the product | Finding the right code in your catalog: grade, concentration, pack | Same name with a different grade or concentration; the customer's own nicknames for your products |
| Pricing the line | Taking the price from your list for that customer, quantity and unit | Price per kg against a request in litres; expired prices; customer-specific prices; discounts |
The three jobs carry different risks. A reading mistake is usually visible: the line looks wrong next to the customer's text. A matching or pricing mistake can look perfectly plausible and still be wrong. That is why the bar for automation is higher for the second and third jobs.
Where is AI strong? Reading messy requests
Large language models are good with text that does not follow a template, which describes most enquiries a chemical distributor receives.
- Free typing. “Need caustic 50 pct x1 IBC + 4 drums IPA” has no columns, but it holds two lines with a product, a pack and a count each.
- Several products in one message, including a list pasted into the body of an email.
- Mixed spelling and abbreviations: caustic, IPA, H2O2, sulphuric and sulfuric.
- Tables that arrive as text, with columns flattened by copy and paste.
- Drafting the follow-up question when something is missing, in the customer's language.
In each case a mistake is cheap to catch if a person can see the original text next to what was read. That is the condition for automating the reading step: the source text stays visible, line by line.
Where is AI risky in chemical quoting?
The risk is not that a model reads badly. It is that it fills a gap with a confident, plausible answer. In chemical quoting, plausible and correct are often a few percentage points apart.
Grade and concentration
Hydrogen peroxide 35% and 50% are different products, with different safety data sheets and different prices. A tool that maps a 35% request to the 50% product in your catalog because it is the closest match has made a commercial decision and a safety decision without telling anyone. Grade works the same way: a technical grade and a pharmacopoeia grade of the same substance share a name and a CAS number. A CAS number identifies the substance, not its grade or concentration.
Units and density
A customer asks for 1,000 L and you price per kilogram. Converting needs the product's density, which depends on the product and on temperature. A tool that assumes 1 kg per litre would quote about 1,000 kg for 1,000 L of 50% sodium hydroxide solution; at about 1.52 g/cm³ (20 °C) the real mass is roughly 1,520 kg. Density should come from your SDS or technical data sheet and be approved once by a person, not taken from a model's memory. There are worked examples in litres to kilograms: density and pack sizes.
Prices and discounts
A price is a commercial commitment. It should come from a record you can point to: your list, for this customer, at this quantity break, valid today, in this unit and currency. A language model can produce a number that looks like a price, but it cannot be the source of one. Discounts belong to people with the authority to give them.
Same as last time, and invented values
“Citric acid, same as last time, 20 bags” needs an order history. If a tool cannot see that history, the honest output is a question, not a guess. Fluent invented detail is what people call an AI hallucination: output that reads well but is not supported by the source. At the quote desk the defence is simple. Every value carries its source, and a value without a source is not priced.
What do European chemical distributors report?
The best recent picture is the Pulse Check published in January 2026 by BCG and Fecc, the European association of chemical distributors. It surveyed 170 chemical distributors in ten European countries in summer 2025; 68 of them have fewer than 50 employees.
- 63% ranked digitalisation among their top three investment areas.
- 51% were already testing or using generative AI: 29% in initial internal testing and 22% in a pilot or live use. A further 23% were evaluating it.
- 83% listed artificial intelligence among the technologies most relevant to their company over the next two to three years.
These are survey answers about plans and adoption, not measured results, and the report does not say which quoting tasks AI handles well. What it does show is that the question in this article is being asked across European distribution, by small firms as well as large ones.
Ten questions to ask before you choose a quoting tool
These questions work for any option, from a general chat assistant to specialist software. Ask for a demonstration on your own requests, not on the vendor's examples.
- Does it search your own catalog? A tool that suggests products from general knowledge can propose something you do not stock.
- Does the price come only from your price list? Including customer-specific prices, quantity breaks and validity dates.
- Can it say “not found”? A tool that always returns a match will sometimes return the wrong one.
- Does it show the source of every value? For each line: where in the request the quantity came from, which catalog record was matched, which price record was used.
- How does it handle litres to kilograms? Does it guess a density, or does it ask you to approve a product-specific conversion?
- Does it stop on a concentration or grade conflict? A 35% request against a 50% product should stop the line, not pass silently.
- What exactly does it check about the SDS, and what does it not check? “SDS checked” can mean anything from “a file is attached” to a content review.
- Which AI model and sub-processors does it use, and where is your data processed? Get the answer in writing.
- Are approval and discount authority limited by role, and is every decision logged? You should be able to see who changed what, when and why.
- How will you measure accuracy on your own requests? A practical measure is the uncorrected line rate: lines that were right without any human correction, divided by all lines. See how to measure quote turnaround honestly.
What are the options? A category-by-category view
No vendor names here: products change quickly, and the fair comparison is always a test on your own requests. The categories, though, have stable strengths and weak spots.
| Option | What it does well | What to watch |
|---|---|---|
| Experienced team, spreadsheet and manual ERP entry | Chemistry knowledge sits with people who know your customers; handles odd requests well; no new software | Depends on a few people; slows down when the queue grows; hard to see where a value came from; hard to measure |
| Your ERP's own quote or order agent | Works where your prices and stock already live; little extra cost if you already hold the licence | Depends on your ERP edition and on how packs and concentrations are modelled; chemistry checks may not be configurable; test it on your hardest requests |
| General document reading or order-entry automation | Strong at extracting fields from varied documents at volume; established integrations | Built for generic order lines; grade, concentration basis, density and SDS rules are usually yours to configure |
| Pasting into a general chat assistant | Cheap or free; good at reading messy text and drafting replies | Does not know your catalog, your prices or which products have a current SDS; can invent plausible values; no audit trail |
| Chemistry-specific quote-desk software | Built around concentration, units, packs and documents | A young category: ask for evidence on your own requests and for a plain list of what it does not do |
There is no single right answer. A business with a few requests a day and a small catalog may be well served by an experienced person and a spreadsheet. A business already running an ERP that offers its own agent should test that first. Whatever you test, compare each output line with the quote you actually sent.
Where does Vuntus stand?
Vuntus is software that turns chemical distributors' incoming price requests into quote drafts built from your own catalog and price list, with every line traceable and approved by your team.
As of October 2026, Vuntus does not use an AI model to read requests or match products: it splits lines with rule-based software, searches your catalog, stops where it is unsure, and leaves every line for your team to approve.
Here is how it answers the ten questions as of October 2026. Vuntus is in an invite-only pilot.
- 1. Your own catalog
- Yes. Products are matched only against your catalog; the best candidate is listed first, but no line is confirmed until a person confirms it.
- 2. Prices from your list
- Yes. The list can hold customer-specific prices, quantity tiers and validity dates, and “Price source” shows which record was used. There is no currency conversion.
- 3. Says not found
- Yes. A line with no suitable product shows “No match”.
- 4. Source of every value
- Yes. “View source”, “Why this product?” and “Price source” open the evidence for quantity, product and price.
- 5. Litres to kilograms
- No guessing. The line is not priced until a manager or the owner approves a product-specific conversion with a reason, which is written to the audit trail, or the conversion is supplied in the catalog CSV.
- 6. Concentration and grade
- Concentration: a line whose percentage differs from the catalog cannot be confirmed, not even by a person. Grade: a grade difference does not stop a line at approval today, so it stays your team's check.
- 7. SDS
- A quote containing chemicals cannot be approved without a verified SDS link on the product. Vuntus does not read SDS content and does not check regulatory compliance.
- 8. Model and data
- No AI model is used to read requests or match products. Ask for hosting and sub-processor details in writing, as you would from any vendor.
- 9. Roles and logging
- Yes. Owner, Manager, Sales and Viewer roles; a discount cap for the Sales role; an optional rule that the person who prepared a quote cannot approve it; an append-only audit trail.
- 10. Measured accuracy
- Not yet. Vuntus has no accuracy figure measured on real customer requests. The honest test is a set of your own.
When Vuntus is not the right fit
- You handle a few requests a day and a few hundred products: a person with a spreadsheet, or a general chat tool used with care, may be enough.
- Your quote has to be created automatically inside your ERP: Vuntus does not connect to an ERP today. Catalogs come in as CSV, and quotes leave as PDF, HTML or CSV files.
- Most of your requests arrive as scanned PDFs or photos: Vuntus does not read lines from them, so the lines are added by hand (which files lines are read from).
- You want the software to email quotes to customers: Vuntus does not send them. Your team downloads the quote and sends it from its own email.
- You want to open an account and try it yourself straight away: as of October 2026 Vuntus is in an invitation-only pilot, with no public sign-up.
For the full list of what Vuntus does and does not do, see What Vuntus is (and isn't); for short answers, the FAQ. In the user guide, Receiving requests, Reviewing request lines and Creating, approving and exporting quotes show each step on screen. To see the request-to-quote flow, read how it works or the checks Vuntus runs on chemical lines. For the questions to settle on each line before pricing, see the chemical RFQ checklist. Sources last checked on 10 Oct 2026.
Frequently asked questions
Can ChatGPT prepare a chemical quote?
It can read a messy request and draft a reply, and that part can save effort. It does not know your catalog, your customer prices or which products have a current SDS unless you give it that data, and it can fill gaps with plausible values. Use it for reading and drafting; keep product, price and approval with your own records and people.
Should AI set the price?
No. The price should come from a record in your price list for that customer, quantity and unit. Software can help find the record, but the number itself should be traceable to it, and discounts should stay with people who have the authority to give them.
Which quoting steps does Vuntus handle with rules?
Splitting lines, searching your catalog, taking the price from your list and the concentration, unit and SDS checks are done by rule-based (deterministic) software; as of October 2026 no AI model is used in these steps. It stops where it is unsure, and every line goes through your team's approval.
How do I measure a quoting tool's accuracy?
Run your own recent requests through it, including difficult ones, and compare each output line with the quote you actually sent. Count the uncorrected line rate: lines that were right without any human correction, divided by all lines. Look at a few dozen lines at least before you trust a percentage.
Where does my data go when I use an AI tool?
It depends on the tool and on the model provider behind it. Ask for the list of sub-processors, where data is processed, how long it is kept and whether it is used to train models, and get the answers in writing. This is general information, not legal advice.