AI in Procurement: Comparing Quotes, Questioning Framework Contracts

Tuesday, late August, shortly after nine. In the procurement department of a mid-sized manufacturer, three documents are ready: the quote from Bidder A, the quote from Bidder B and the running framework contract TP-118 with two amendments. The subject is shipping cartons, about 120,000 units a year, under tender AS-26-02. The company, the documents and every figure in this article are a fictional example.

The task sounds simple: which quote is better, and does the running contract have to be terminated? In practice that means laying prices from three tables side by side, understanding tiers, finding terms and notice periods, tracing which amendment changed what, and working out how long each quote is valid. Done carefully, that takes half a day. And at the end there is the nagging feeling of having missed a footnote somewhere.

AI suits this work well, but not in the way people first expect. The obvious expectation is that the assistant names the best quote. The more useful expectation is that the assistant puts the terms into a common structure, backs every figure with its source location and calculates what it is told to calculate. Procurement assesses, negotiates and decides.

Our thesis in one sentence: the value of AI in procurement is not naming the cheapest quote, but making terms comparable, backing every statement with evidence and giving procurement the time for the actual decision. This article covers the typical tasks, a workflow, the traps hidden in PDFs, an example calculation, the logic of deadlines and the line between what an assistant delivers and what stays with procurement.

What Costs Procurement Time: Five Questions That Keep Coming Back

Whether it is packaging, spare parts, transport services or professional services, the questions procurement asks of quotes and contracts repeat. Five topics account for most of the reading.

None of this is complicated. It is tedious, because the answers are scattered and every document is built differently. Bidder A states the tier in a paragraph, Bidder B in a table on page 4, and the framework contract in an annex that an amendment later overwrote. Anyone who wants to compare has to translate first: bring every statement into a common structure in which like things sit next to each other.

Looking Up and Comparing Are Two Tasks

It helps to keep two tasks apart, because they fail in different ways. Looking up concerns one document and one question: what notice period does the framework contract state? The risk lies in completeness, for example an amendment that changes the answer. Comparing puts the same question to several documents. The risk lies in comparability: two prices of 0.88 euros and 0.85 euros look comparable and are not if the tier refers to different quantities.

Both can be handled with the same mechanism. The control questions differ, which is why we show both in one running example.

The Workflow: From a Stack of Documents to a Cited Comparison

A workable process has five steps. The assistant takes the first four, procurement keeps the fifth.

Enlarge

1. 1.Ingest: Quotes, framework contract and all amendments go into one shared collection. Whether that collection is complete is the first check, and it belongs to procurement.
2. 2.Extract: The assistant reads text, tables and footnotes. Scans additionally need text recognition.
3. 3.Ask: The same question goes to every document, per criterion, for example “Which notice period applies?”
4. 4.Evidence: Every answer names its source location. Without evidence, an answer is only a claim.
5. 5.Decide: Procurement checks, assesses, negotiates and decides.

Under each step the graphic shows a typical source of error: a missing amendment, a misread table, a question with too much room, an answer without a source location, and a result adopted unchecked. The order is no accident. Errors from steps 1 and 2 look like confident answers in step 4. An amendment that was never ingested goes unnoticed, because nothing points to it.

Permissions Belong in the Workflow

Quotes are confidential. Bidder A's price is none of Bidder B's business, and even inside the company not everyone sees everything. An assistant should therefore draw only on documents the person asking could open in the folder themselves. For connected sources such as Google Drive, TheroAI adopts the permissions of the source system, as described on the connectors page. In practical terms for procurement: the folder for tender AS-26-02 has a small, named group, and exactly that group sees the answers.

The Example: Two Bidders, One Framework Contract

Bidder A is the current supplier and serves the company under framework contract TP-118. Bidder B is a new provider. Both have submitted a quote for tender AS-26-02. The five documents at a glance:

Which price and which deadline apply today can only be answered by reading the main contract and both amendments together. That is exactly where a question across several documents earns its keep. The comparison the assistant builds from them looks like this:

Enlarge

Four things stand out at first glance that are not stated in any single document:

• The prices are not directly comparable. Bidder A tiers by annual quantity, Bidder B by quantity per call-off. The same number means something different.
• The unit price is not the total price. Bidder B delivers ex works and charges 180 euros per delivery. Bidder A delivers free to the door.
• The commitment differs. Bidder B binds for 36 months with no ordinary termination. Bidder A binds for 24 months and renews afterwards only if nobody terminates.
• The quotes expire at different times. Quote B lapses on 15.10.2026, quote A on 31.10.2026.

The value of such a comparison lies less in speed than in making the differences visible before anyone sets prices side by side. Anyone who calculates first and discovers afterwards that the reference quantities were different has done the calculation for nothing.

Where PDFs Mislead: Tables, Footnotes, Amendments

A PDF looks like a document and behaves like a picture of a document. For people that is convenient, for any automatic extraction it is a source of errors. The graphic shows a fictional price sheet page from quote B with three places where a comparison can tip over.

Enlarge

Tables: The Reference Is Not in the Number

The values 0.90 euros and 0.85 euros sit in the table, while the reference quantity sits in the column heading: “quantity per call-off.” If a tool reads only the numbers, the result is a tier with no reference. It looks complete and is useless. Good extraction reads the cell together with its row and column headings and returns both.

That is why an answer like “B: 0.85 euros” is not enough. A usable one reads: “B: 0.85 euros per unit from 25,000 units per call-off, price sheet, page 4.” Tables that run across a page break also sometimes lose their header row. The figures on the next page then have no heading at all.

Footnotes and Annexes: The Small Print Is Part of the Price

The freight charge of 180 euros per delivery sits in a footnote in quote B. It is not included in the 0.85 euros. Anyone who compares only the table compares Bidder A free to the door with Bidder B ex works. Annexes work the same way: general terms and conditions, price lists and service descriptions are often separate files that a quote makes part of the contract without repeating them. They still belong in the collection.

Amendments: The Main Contract Is Rarely the Whole Contract

Framework contract TP-118 consists of a main contract and two amendments. Amendment 1 raised the price from 0.91 euros to 0.94 euros, and amendment 2 extended the notice period from three to four months. A question put to the main contract alone therefore yields two answers that are wrong even though they appear word for word in the document. The assistant has to do more than read, it has to sort: which version applies today, and which clause has which amendment superseded?

Validity: A Quote Has an Expiry Date

Quote B is valid until 15.10.2026, quote A until 31.10.2026. This is often found in the header or the last paragraph, rarely in the table. A comparison that does not carry the expiry date along may recommend a quote that has already lapsed by the time the decision is made.

Five Checks for Every Figure From a PDF

• Reference quantity: per unit, per 100 units, per pallet, per call-off or per year. Which one is in the heading?
• Footnotes and asterisks: Which extra costs or conditions hang on the figure?
• Unit and tax: net or gross, euros or cents, units or thousands?
• Page breaks: Does the table still have its header row on the next page?
• Scans: Does the file have a text layer, or was it read by text recognition? For scans it pays to look at the original, because digits such as 0 and 8 or 1 and 7 can be confused.

An assistant does not replace these checks, but it can ask them every single time. A checklist that people gladly skip on a Friday afternoon can be built into a workflow as a fixed question.

The Tier Decides: An Example Calculation

Back to the example. The question is: what does the annual quantity cost with the two bidders, and how strongly does the result depend on how we order? For the calculation we use round assumptions. It is an example calculation and not a forecast, and all values are invented.

The tiers from the quotes work like this: Bidder A charges 0.95 euros per unit up to 100,000 units of annual quantity, and 0.88 euros per unit for all units once the annual quantity exceeds that. Bidder B charges, per call-off, 0.90 euros up to 24,999 units and 0.85 euros from 25,000 units, plus 180 euros freight per delivery. The running contract costs a flat 0.94 euros per unit after amendment 1.

Enlarge

Three findings follow from it:

The ranking flips. With four large call-offs, Bidder B is 2,880 euros cheaper than Bidder A. With twelve small call-offs, Bidder B is 4,560 euros more expensive. Which quote is cheaper depends on an assumption that appears in neither PDF: how do we really order?

Bidder A depends on the annual quantity. The lower price applies only if the annual quantity exceeds 100,000 units. If demand falls to 90,000 units, 0.95 euros applies, which is 7 cents more per unit. If the lower price applies retroactively to all units, it often takes effect in practice only at year end as a rebate, and until then the money flows at the higher price. How quote A would handle this has to be stated in the quote and belongs in a follow-up question.

Both quotes are below the running contract, but by different amounts. Bidder A saves 7,200 euros in both scenarios against 112,800 euros. Bidder B saves 10,080 euros in the first scenario and 2,640 euros in the second, and binds for three years in return.

An assistant runs these variants in seconds if it is given the assumptions. The assumption itself, how call-offs really happen, is known to planning and procurement. So do not ask the assistant to guess it, give it. And have the calculation steps shown, as in the table above: language models can miscalculate, and a calculation path can be rechecked in a minute.

Deadlines: What Applies, Since When and Until When

The most expensive piece of information in this example is not a price but a date. The main contract says: the contract runs until 31.12.2026 and renews by 12 months unless notice is given three months before. That makes 30.09.2026 the last day. Amendment 2, in force since 01.04.2026, sets four months. The last day is therefore 31.08.2026, a month earlier.

Enlarge

Anyone who questions only the main contract on this Tuesday believes there is a month left, and is wrong. Anyone who questions the complete collection sees that the deadline expires in a few days.

Thought experiment: The assistant answers “By when can I terminate?” from the main contract only. The answer is cleanly worded, names a clause as its source, and is still wrong, because it does not know amendment 2. Nothing about it looks suspicious.

This is a gap that no model closes by itself, only the order of the documents does. Three habits follow:

1. 1.Check completeness before asking. Are all amendments in the collection, and are they named as such? A folder “Contracts new” with files like “Scan_0417” helps neither an assistant nor a person.
2. 2.Ask for the position today. Not “Which notice period does the contract state?” but “Which notice period applies today, and which amendments changed the clause?”
3. 3.Have deadlines returned as a timeline. A list of dates invites skimming, while an axis with expiry dates shows what comes first.

Whether a notice is on time also depends on how the contract regulates receipt and the calculation of periods. This is not legal advice, so clarify the individual case with your legal counsel. The assistant supplies the date and the source location, not the legal review.

And then one question remains that no document answers: terminate or not? Terminating by 31.08.2026 creates room to negotiate, but carries the risk of standing without a contract in the autumn if none of the new quotes holds up. Not terminating renews the contract for 12 months at 0.94 euros per unit. That is a weighing of risk, relationship and negotiating goal. It has no source location.

Questioning the Framework Contract: Questions That Hold Up

The quality of an answer depends on the quality of the question. Good questions to a framework contract follow a pattern: they ask for the position today, they allow the answer “not in the document,” and they ask for the source location.

The last row is the most important. An answer with evidence can be checked without reading the whole document: you open the location, read the sentence and know whether the answer is right. In TheroAI the answer carries numbered references, and the “Quellen” panel lists the documents behind them. The following capture comes from the German interface and shows a different fictional example, an acceptance report. The principle is the same as in procurement.

Enlarge

Anyone who has the same tasks again and again does not have to word the questions anew each time. TheroAI's prompt library contains templates such as “Lieferantenvergleich” (supplier comparison) and “Ausschreibung zusammenfassen” (summarize a tender) that your team can take as a starting point. And if you need the comparison as a document, for example for the award memo, you can have the assistant create it in the canvas and keep working on it there.

How this looks in your industry is shown on our pages for logistics and supply chain, where framework contracts and tenders also appear as use cases.

Where the Line Is: What Stays With Procurement

So far we have covered what an assistant can deliver. Just as important is what it should not deliver. The following table separates the work by responsibility.

Three points deserve a paragraph of their own.

Assessment is more than paper. How reliably has Bidder A delivered in recent years, how does he react to disruptions, how good is the quality of the cartons? None of that is in any quote. An assistant can provide documents on it, such as complaints from the mailbox, but the weighting is procurement judgment.

Confidentiality has rules. One bidder's prices must not be shown to the other, and an assistant must not pass them on through the back door. That is why the permissions question belongs at the start, not at the end. For public contracting authorities, procurement regulations add binding rules on the form and course of a comparison. This is not legal advice, so clarify the individual case with your legal counsel.

Nothing goes out without approval. A comparison leads to letters: a notice of termination, a rejection to Bidder B, a query to Bidder A. Such drafts should run in a workflow with an approval step. A named person sees the draft, approves or rejects it, and only then does anything go out. In TheroAI such a step can be set up with a deadline, for example 72 hours, optionally with the four-eyes principle. How this looks is described on the workflows page.

Our Position

We consider AI in procurement worthwhile when three conditions are met. First, every statement has a source location that opens in seconds. Second, the collection is complete, including all amendments, and the permissions are right. Third, the decision stays with a person who signs for it.

Then AI changes procurement at one point, and the right one: the reading shrinks, and attention for negotiation and assessment grows. The biggest gain in our example was neither the price difference of 2,880 euros nor the one of 4,560 euros. It was a date: 31.08.2026, which became visible only by reading the main contract and amendment 2 together.

We do not think much of letting an assistant crown the “best” quote. A ranking suggests an objectivity that prices, risks and relationships do not provide. A cited comparison that shows where the quotes differ and what the result depends on is more honest and more useful.

What You Can Take Away

If you start with this topic in your procurement department, these questions and steps help:

• Is the collection complete for every framework contract, with all amendments, and are the files named so that an amendment can be recognized as one?
• Do you ask for the position today rather than for the main contract? Does every answer name document and location?
• For every tier, have you clarified what it refers to: unit, call-off, year or order?
• Do freight, packaging and surcharges appear in the comparison, even if they only sit in a footnote?
• For every running framework contract, do you know the date of the last possible notice, and is it in a list that somebody maintains?
• Who approves letters going out, and how long does that person have to do so?

And an exercise for this week: take a running framework contract with at least one amendment and ask the question “Which notice period applies today, and which amendments changed the clause?” Check the source location in both documents. If the answer is right, you know how much trust you can place in the workflow. If not, you know where to start.

If you would like to see how questions across several documents, source locations, permissions and approvals fit together in one assistant, we are happy to show you with sample data, without pressure and without obligation: View the demo.