Comparing supplier quotations with AI means building a single standard structure from multiple commercial offers, identifying differences, and drafting questions before procurement. The model helps extract information from documents, but the final table must preserve sources, unknown terms, and calculation rules. The lowest amount in a PDF does not necessarily mean the best offer.
This article is intended for procurement leaders, business owners, and employees who select equipment, materials, or services. We are looking at commercial comparisons for an internal decision. Special procedures in regulated procurement, legal contract review, and tax implications require separate verification for the specific transaction.
Key takeaways: first define the same item being compared, then verify mandatory requirements, and only after that calculate the total cost. AI should show what data is missing, not fill gaps with plausible assumptions.
Prepared on September 8, 2026 using AI and technical source review. All numerical examples below are illustrative, are not AIrassvet commercial offers, and do not describe completed purchases.
Contents
- What to compare
- Preparing documents
- Step-by-step process
- Calculation example
- Request template
- Reviewing the result
- Ongoing process
- Economics
- Frequently asked questions
Why a simple price table gives an incomplete picture
Three suppliers can respond to the same RFQ in different ways. One includes shipping, another lists it separately, and a third offers pickup. One supplies a ready-to-use package, another offers a basic configuration without a required accessory. If you compare only the final totals, differences in scope will remain hidden.
The problem gets worse when offers come in different formats. In Excel, the details are visible; in PDF, payment terms are moved to the last page; and in an email, the account manager changes the delivery date. The comparison must account for the entire agreed package while preserving dates and versions.
AI is useful as a reading and structuring assistant. It can collect terms into one view and highlight items that need review. However, the decision on supplier selection requires confirmation of facts and the buyer's priorities. An automated ranking without rules creates the illusion of objectivity.
| Group | What to compare | Typical source of error |
|---|---|---|
| Subject | Model, configuration, quantity, specifications | An equivalent is mistaken for an identical item |
| Price | Unit, currency, tax presentation, discounts | Package price is compared with unit price |
| Logistics | Delivery, unloading, place of handoff | Costs are missing from the total |
| Timeline | Production, shipment, delivery, commissioning | Ship date is treated as the ready date |
| Payment | Advance payment, milestones, deferred payment | Same price with different cash-flow impact |
| Operations | Service, consumables, training | Purchase price replaces total cost |
What documents and rules are needed up front
Prepare the underlying need: specifications, quantity, delivery location, desired date, and required characteristics. If the request is worded vaguely, a good summary will not make the offers comparable. First define exactly what the company is buying.
Attach supporting documents to each quotation: clarification email, specifications, warranty description, and current delivery terms. Assign an ID to the package and record the version and date. If two files conflict, the system should show a "source conflict" status.
Do not mix mandatory requirements with preferences. Operating voltage, compatibility with an existing line, or the latest start date may be eligibility constraints. Interface convenience and extended warranty can be assessed after eligibility. A high score on price should not offset an unmet mandatory requirement.
Define the analysis horizon in advance. For a one-time service, contract cost may be enough; for equipment, consumables and service over the selected period matter. The horizon must be the same for all offers and stated explicitly in the report.
Also agree in advance on how to treat tax presentation in pricing. Do not turn the absence of a note into an assumption about the rate. A finance professional should determine a comparable basis using the company's rules. In the training calculations below, all amounts have already been brought to one basis without additional tax conversions.
How to organize AI-based quotation comparison
Step 1. Extract text, tables, and structure
Keep the original files and extract the information together with source coordinates: document, page, table, row. Simply merging all text into one paragraph can lose the link between a number and a column heading or note.
Document analysis tools can work with structure. For example, layout model Microsoft Document Intelligence extracts text, tables, and document elements. This confirms the technical capability, but the quality on your quotations should be measured separately.
If the document is a scan, OCR is added — optical character recognition. Check small notes, tables broken across pages, and merged cells. Details on setting up this workflow are in the document recognition guide.
Step 2. Fill in a card for each offer
The card should include the supplier, quotation ID, date, validity period, currency, line items, price, payment terms, delivery, warranty, and exclusions. Each field needs a status and a source. Use "not specified" for missing values and "needs clarification" for conflicts.
Do not identify the supplier company from the logo or file name alone. Match the document details with the correspondence. If a file was sent by an intermediary, that does not always mean the intermediary will be a party to the deal. Keep the ambiguity until it is verified.
Step 3. Match the line items
Different names can describe the same item, and nearly identical names can mean different models. Use part numbers, specifications, and the package contents for matching. Mark acceptable equivalents separately: they cannot be quietly inserted into the original specification line.
When exchanging data with 1C, it is useful to use already confirmed product matches. The 1C-EDO directory describes the item matching mechanism. In quotation comparison, the same principle helps avoid repeated manual matching, but the terms of each specific offer still have to be checked again.
Step 4. Normalize values to a common basis
Normalization includes units of measure, completeness, currency, the selected service period, and required additional costs. Record each conversion with a formula. If the packaging factor is unknown, the system should not calculate a per-unit price.
Do the arithmetic programmatically or in a verifiable spreadsheet. The model can propose the calculation structure, but the final result must be reproducible from the input values. Set rounding explicitly: otherwise, the sum of rounded line items may differ from the supplier's total.
Step 5. Verify eligibility and prepare the decision
First, build a matrix of mandatory requirements with the statuses “confirmed,” “not met,” and “unknown.” Then compare the acceptable options based on the agreed preferences. A lack of information does not mean the requirement is met, and it does not mean the supplier has refused to meet it.
The output should be a summary table, a list of discrepancies, questions for suppliers, and a brief explanation of the choice under the given conditions. If the data is insufficient, a useful result is a list of clarifications after which a decision will become possible.
Training example: three equipment proposals
A company is selecting an equipment package for a production area. To launch operations, it needs a base unit, a required module, delivery, and commissioning. Suppliers A, B, and V offered different configurations. All amounts in the example are illustrative, stated in rubles, and already adjusted to a common comparison basis.
| Article | Supplier A | Supplier B | Supplier V |
|---|---|---|---|
| Base unit | 900,000 | 980,000 | 850,000 |
| Required module | 80,000 | Included | Not specified |
| Delivery | 40,000 | Included | 70,000 |
| Commissioning | 60,000 | Included | 90,000 |
| Confirmed total | 1,080,000 | 980,000 | 1,010,000 without the module |
Comparing only the base unit would make Supplier V look cheapest. Once the configuration is normalized, it becomes clear that the confirmed portion of its price is already higher than Supplier B’s offer, while the cost of the required module is unknown. It is not correct to state V’s total as 1,010,000 rubles without qualification.
Now add the timeline. A claims readiness for use in 20 business days, B in 35, and V lists only shipment in 15. Assume the company needs startup no later than 25 business days. B is cheaper, but it does not meet the schedule constraint; V cannot yet be evaluated on startup date.
There is no universal winner here. If the timeline is mandatory, A is the only option in our example with confirmed compliance with that condition. If the timeline can be changed, a cost comparison with B becomes possible. The decision should reflect business priorities, not hide them inside a ranking.
Add three years of service: hypothetically 60,000 rubles per year for A and 30,000 for B. All else being equal, the total purchase price plus this service would be 1,260,000 and 1,070,000 rubles respectively. The difference is 190,000 rubles. This is still not the full total cost of ownership: electricity, downtime, and consumables are not specified and therefore are not included.
This calculation method makes the constraints visible. It is possible to review delayed start, the cost of waiting, and the option of temporary rental separately. But these costs should be added only after the underlying data is received, not just to complete a table.
How to compare payment terms, warranty, and risk
The same total price can come with full prepayment or payment after delivery. This affects cash flow and execution risk. For an initial report, it is enough to show the payment schedule and the amount of money transferred before the result is received.
If the finance team compares discounted payment costs, it must define the rate and the dates. The model does not choose the cost of capital on its own. A long deferral should also not be treated automatically as an advantage if it comes with restrictions, penalties, or a higher price.
Compare warranty terms by substance: when coverage starts, what is covered, exclusions, shipping for repair, response time, and whether replacement equipment is available. The phrase “two-year warranty” without these details does not give a complete picture of the consequences of a failure.
For risk, use observable facts. “Price lock period not specified” is a verifiable note. “The supplier is unreliable” is a conclusion that requires other data. Separate document analysis from counterparty due diligence and legal review.
A technical inconsistency should also not automatically be called a contract breach. A correct formulation would be: “Specification lists model X, while the letter lists X-A; confirmation of the scope is required.” It helps resolve uncertainty without replacing it with an accusation.
Ready-made prompt for an initial review
Below is a model task template. It can be used with documents approved for processing after they are labeled A, B, and V. The template itself does not guarantee error-free output: the result must be checked against the source documents.
Compare the attached commercial proposals for purchasing the same scope. First, create a separate card for each proposal, then a summary table. For each material item, indicate the document, page, or line. If a value is missing, write “not specified.” If the sources conflict, keep both values and mark the conflict. Do not substitute market prices, timelines, or specifications from memory.
Identify the mandatory requirements from the attached specification. Separate exact compliance from proposed alternatives. Do not treat missing information as compliance with the requirement. Convert prices to a single unit only when the conversion factor is known, and show the formula. Do not combine prices with and without tax unless there is a specified comparison rule.
Prepare four outputs: a scope table, a commercial terms table, a list of incomparable or unknown data, and questions for each supplier. Do the final ranking only for options that passed the mandatory requirements, and only using the weights provided. If no weights are given, show advantages and limitations without a single overall ranking.
After the first response, ask it to check one specific thing: whether notes, excluded work, and additional costs were lost. Then verify the arithmetic separately. Asking the same model to “check yourself” is useful as an extra step, but it does not count as independent verification.
How to verify an AI result
It is convenient to break verification into extraction, matching, and calculation. First, make sure the number was read correctly. Then confirm it belongs to the right item and unit. Finally, check that the formula uses the correct fields. One overall quality score hides the source of the error.
For scans, preserve recognition quality information. Microsoft explains the difference between accuracy and confidence: the confidence score must be interpreted in the context of the data. In a workflow, it is more useful to flag the questionable field than to show the user an average confidence score for the whole document.
Prepare a test set with known traps: missing delivery, different currencies, packaging units, a version without a mandatory option, two versions of the same proposal, shipment date instead of delivery date. Also include a document that does not contain the answer to the question. Correctly recognizing that something is unknown should count as a successful result.
For each error, record the consequences. A missing comma in a description and an incorrect total by hundreds of thousands of rubles should not carry the same weight. Evaluate the share of correctly extracted critical fields, the number of false statements, the completeness of identifying mandatory costs, and the time spent on human review.
Do not ask the system vendor to evaluate quality only on its own demo files. Use your holdout set and a pre-agreed reference standard. Include in testing documents created after setup to check whether the solution generalizes.
How to integrate comparison into the procurement process
For infrequent purchases, a workspace with document upload and a reviewable table may be enough. For a recurring flow, you need a connection to the purchase request, catalog, RFQ archive, and approval system. The integration should preserve which set of documents formed the basis of the decision.
Products built for this scenario already exist: for example, Komanda.ai describes comparing supplier quotes and terms. This is an example of a functional direction, not independent proof of quality or a recommendation for a specific service.
In your own system, separate the roles. The initiator defines the need, the buyer reviews the proposals, the technical specialist confirms compliance, and finance verifies the calculation base. Purchase approval should be an explicit action by an authorized employee.
If the process is built in n8n, human-in-the-loop tool review can be used to approve individual actions. For procurement, this might mean creating a clarification request or sending an approved summary. The right to send a message should not automatically grant the right to place an order.
After you receive an updated quote, create a new comparison version. Do not delete the old decision: it explains the approval history. Show what changed — price, configuration, delivery time, or warranty. Approval of the previous version does not silently carry over to the new terms.
The adjacent, broader process is covered in the article on AI for procurement and tenders. Preparing your own sales documents is covered in a separate article on AI-powered commercial proposals.
How to calculate the financial impact
Let's use a training scenario: 40 comparisons per month, 90 minutes of manual preparation each, and an hourly rate of 1,000 rubles. Labor input equals 60 hours, and the monetary equivalent is 60,000 rubles. If, after automation, one comparison takes 30 minutes of review, 20 hours of work remain.
Add an assumed 8,000 rubles in document processing costs and 6,000 rubles in support costs per month. The recurring cost of the new process would be 34,000 rubles. The potential difference is 26,000 rubles. With assumed investments of 180,000 rubles, the simple payback period is about 6.9 months.
This calculation reflects labor only. Savings on the purchase price must be proved separately: a cheaper option could also have been found by an employee without the system. A prevented error can be described as a detected discrepancy, but you cannot automatically count the entire procurement budget as "saved money."
Check the low-volume scenario. If a comparison is needed only a few times per quarter, a full integration may be excessive. In that case, a standard supplier request form and a review assistant are more valuable than a complex platform with ongoing costs.
Frequently Asked Questions
Can AI choose the best supplier?
It can rank options according to set rules after the facts have been checked. But the company defines priorities, acceptable risks, and mandatory requirements. If key terms are unknown, the correct result is a list of clarifications, not a confident winner.
Can you compare offers in different currencies?
Yes, if the exchange rate source, date, and conversion rule are defined. Keep the original amounts and show the conversion separately. Currency clauses and the possibility of price changes should be treated as separate terms.
What should you do if delivery is not specified in the quote?
Mark it as "not specified" and prepare a question. You cannot assume it is free or insert an estimated cost without a separate scenario calculation. Confirmed amounts and assumptions should be kept in different columns.
How do you compare services if contractors use different approaches?
Start with expected outcomes, scope of work, acceptance criteria, and responsibility boundaries. Comparing hours without the same result is often misleading. Put additional work and exceptions in a separate section.
Is it safe to upload a quote to an external service?
It depends on company rules, document content, and processing terms. Check what data is transmitted, where it is stored, who has access, and how it is deleted. Confidential procurement information should not be uploaded to an unapproved tool.
Do you still need Excel after implementing AI?
It can remain a convenient format for review and approval. The important thing is that values preserve their sources and formulas, and that the export matches the decision version. The file format matters less than the reproducibility of the calculation.
How AIrassvet helps automate quote comparisons
AIrassvet focuses on LLM integration and process automation. For procurement workflows, this includes preparing and extracting data from documents, building a reviewable summary, and connecting approvals to operational systems.
The first step is to choose one procurement category and collect several completed comparisons with known outcomes. This will help align the required fields and test whether automation actually helps the buyer in practice. Discuss the project.
Where to start with comparison in a company
Create a single need template and ask suppliers to disclose the same terms. Then test AI on a small set of documents where the differences are known. Evaluate completeness, calculation accuracy, and review time. A useful system makes the decision more transparent: it shows confirmed facts, constraints, and the questions that truly affect the choice.