An AI agent applied to quotes reads incoming documents, extracts their line items, checks the amounts and writes the result into your business software. The same setup runs in reverse. It then produces a first price estimate from your history of past deals. Tandem deploys these chains at printers, industrial joinery manufacturers and engineering firms, on volumes ranging from a few dozen to several thousand documents a month.
In the FPV Industries case, a quote arrives as a PDF and comes out structured for the ERP in thirty seconds, at an average cost of around fifteen cents per document. This article describes how the chain is built, what accuracy to expect, where to place human validation and what processing really costs.
What is an AI agent applied to quotes?
An AI agent applied to quotes is a program that receives a commercial document, understands it and acts on it without human intervention on standard cases. It differs from a general assistant on one decisive point. An assistant answers a question typed into a chat window. A quote agent triggers itself when the document lands, applies your business rules and writes into a third-party tool.
The word agent is misleading here. In the vast majority of delivered projects, the component that handles quotes is a deterministic sequence of steps. The language model is called at one point only, where text needs interpreting. That distinction has very concrete consequences for reliability, and it comes back further down.
Reading a quote or producing one, the two directions of the flow
Two uses live under the same label, and they demand neither the same work nor the same guarantees.
- The inbound direction. Supplier quotes, price requests or purchase orders arrive by email, through a form or as a photo taken on site. The agent reads them, pulls out the line items and feeds the business software.
- The outbound direction. The agent starts from a customer requirement, finds comparable past deals in the history, and proposes a first estimate that the estimator corrects. The gain is on preparation time, never on the pricing decision.
The inbound direction is the more mature of the two, because the truth there is verifiable line by line. The Sprint case shows the scale. This industrial printer was receiving thousands of quotes a month as PDFs, all keyed in by hand. Six document families now run through the same extraction chain, which feeds twelve sites.
How the chain is built, step by step
Processing an incoming quote takes five steps. Each one has its own technology and its own failure mode, which makes it testable in isolation.
- The trigger. A dedicated mailbox, a form or a shared folder signals that the document has arrived. It is the most mundane step and the one that breaks most often, because it depends on third-party access.
- Reading. An optical recognition engine turns the page into positioned text. A quote is a table before it is text, so position matters as much as characters.
- Extraction. A language model pulls out the expected fields, references, quantities, unit prices and terms. This step replaces the old per-supplier templates, which broke at the first layout change.
- Checking. Business rules match every line against the catalogue, recalculate the totals and flag discrepancies. At FPV Industries, that layer amounts to close to a thousand lines of rules.
- Writing. A structured file goes into the ERP or the accounting software, with the original document attached. Keeping the original is what makes an audit possible six months later.
On tooling, Tandem orchestrates these chains with n8n, which can be hosted on your own servers, an optical recognition engine for reading and a small fast model for extraction. I detailed this setup in my Optimia newsletter and I keep coming back to the choice of orchestrator. To understand the tool itself, start with our n8n guide.
That sheet comes from my n8n one-pager, published in February 2026. The integration count printed on it has already moved since, which is the rule with this kind of document.
Why optical recognition alone falls short
Reading the characters of a quote and understanding what it says are two different problems. Optical recognition faithfully returns what is printed. It has no way of knowing which value is the right one when several contradict each other on the page.
The clearest example comes from a test I published in my post on image extraction in April 2024. On a plain receipt, the amount printed in large characters differed from the amount actually due, because a discount applied further down. A reading engine on its own copied the large figure. The language model understood the nuance and kept the correct total.
That is exactly what plays out on a real quote. A line discount, a struck-out option, a deposit already paid, a unit expressed per square metre rather than per piece. Each of those reads without difficulty and gets understood wrong. The interpretation layer is therefore the value of the system, and reading is only its entry point.
Workflow or autonomous agent, which one holds up in production
The short answer fits in one sentence. For quote processing, a predictable sequence of steps almost always beats an agent left free to decide.
I set out that position on LinkedIn in October 2025, in my post on AI agents. Nearly all the use cases Tandem meets in companies are solved by a workflow, and that is good news. An autonomous agent is hard to test, since it would need a dataset covering every possible path. It is also hard to monitor, because two identical runs can diverge.
On a quote flow, the sequence is known in advance and repeats identically. Autonomy brings nothing and costs reliability. Artificial intelligence therefore steps in at a single point, interpreting the document, and the rest of the chain stays ordinary code.
That boundary deserves a detour, and our article on agents and workflows covers it case by case.
What accuracy to expect, and where to place human validation
No extraction chain reaches a hundred per cent on heterogeneous documents. The useful question lies elsewhere. It lies in what the chain does with the documents it doubts.
Every document leaves extraction with a confidence score, computed line by line. Confident quotes go straight into the ERP. Doubtful quotes enter a validation queue where the reviewer sees the original document and the proposal side by side. Unknown documents go back to manual handling, and the case then enriches the rules.
| Processing step | Automatable | Stays human |
|---|---|---|
| Reading the document | Yes, even on a poor scan | Illegible handwritten document |
| Extracting line items | Yes on mapped families | A new document family |
| Catalogue matching | Yes once the rules are written | Reference missing from the catalogue |
| Recalculating totals | Yes, every time | Nothing |
| Pricing decision | No | Always, it is a commercial judgement |
How much does processing a quote with an AI agent cost?
Two costs need separating, and confusing them distorts every profitability calculation. There is the cost of building the chain, paid once, and the cost of processing a document, paid on every run.
On our engagements, processing a simple document stays in the order of a few cents. A quote that needs catalogue matching costs more, around fifteen cents at FPV Industries, because it mobilises more reading and more checks. Against the human time spent keying data in, the ratio leaves no room for comparison.
The real budget sits elsewhere. It lies in mapping the business rules, in the number of document families to cover and in the quality of the reference data. A company without a clean catalogue will first pay to clean up its catalogue, and that spend has nothing to do with AI.
What electronic invoicing changes for your documents
A regulatory deadline has just reshuffled the deck. According to the French finance ministry, the business-to-business electronic invoicing reform came into force on 1 September 2026. Every company must now be able to receive its invoices in electronic format, whatever its size. The obligation to issue currently covers large companies and mid-caps, with the smallest having until September 2027. More than ten million economic actors are concerned.
The reform covers the invoice and leaves the quote aside. Its effect on automation projects is nonetheless direct. Invoices now arrive structured from an approved platform, so extracting them loses much of its point. Quotes, meanwhile, remain free-form PDFs sent by email, and they mechanically become the bulk of the remaining reading work.
Business software vendors understood this before many companies did. In January 2026, Cegid and EBP announced their new management platform, EBP Open Line Plus. It combines compliance with the reform, automatic extraction of key invoice data and electronic signature of quotes. EBP claims more than 130,000 companies equipped with the previous generation of its solution.
These native functions cover the standard case, and they cover it well. They stop where trade-specific rules begin, matching against a catalogue of configurable products or writing into an internal system. That is exactly the zone custom chains occupy.
The mistakes that sink a quote automation project
The failures observed on this kind of project repeat themselves, and none of them comes from the language model.
- Starting without mapping the document families. A chain that works on one supplier does not work on the next. Counting the families before writing code avoids a nasty surprise in month three.
- Aiming at the hardest document first. The right pilot is the most frequent and most regular family. The one that hurts most waits its turn. At Sprint, the sequence started on a pilot family before extending to all six.
- Removing all human validation on day one. The validation queue also measures the real quality of the chain. Emptying it too early amounts to flying blind.
- Forgetting the write into the target system. Many projects stop at a nice structured file that nobody picks up. Integration into the management system is half the work, and often the worst estimated half.
- Confusing time saved with jobs removed. Teams accept the chain when it gets them out of data entry. They sabotage it when it arrives as a threat, and a sabotaged project returns nothing.
These flaws show up across every trade with heavy document volumes. Our engagements in construction give a sector-by-sector view.
Should you automate your quote processing?
Yes, as soon as three conditions are met. A monthly volume counted at least in dozens of documents. Identifiable and relatively stable quote families. And clean reference data, a catalogue or a history, to anchor the checks on.
If one of the three is missing, start there. An up-to-date catalogue is worth more than an AI agent sitting on approximate references. Conversely, a company that ticks all three boxes gets a first measurable result within weeks, and the processing cost becomes so low that it drops out of the conversation.
Tandem has supported more than 50 companies on optimising their processes with AI, across around 100 engagements. Document processing comes up in almost all of them. If you want to see what the chain would do on your own quotes, our Autopilot offer starts precisely from a sample of your documents.
