AI Procurement Assistant
By Ai Flowi · Published .
An AI procurement assistant can prepare a supplier evidence matrix from quotations and documents you supply. Define the purchasing criteria first, extract comparable terms with source references, and leave unknowns visible for a buyer to review. The comparison artifact informs the buyer; supplier selection, spending approval and order placement remain in the purchasing process.
example configuration, not a recorded test. This proposed assistant design is documented in the catalogue; not production-tested by Ai Flowi. It is a workflow you can configure, not a ready-made procurement product offered here.
1. Define the purchasing question before comparing suppliers
Collect the specification, requested quantity, unit of measure, delivery location and any mandatory purchasing requirements. Keep mandatory criteria separate from preferences. Ask the buyer to define the comparison rules rather than allowing a model to decide which commercial terms matter.
Give every quotation an identifier, supplier name, supplied date and version. Preserve its scope, validity wording, exclusions and attachments. A price without its unit or quantity basis is not a comparable price. Likewise, a delivery estimate without its triggering condition should stay as quoted wording until the buyer can clarify it.
For quotation review in AI Flowi Workflow, configure your provider account and required API key through model access (provider usage billed separately) and your Google account through Docs access, obtain permission for supplier material and appoint the reviewing buyer.
2. Load approved quotations and inspect their text
Load the quotations with Read File following the source-file instructions before inspecting their commercial terms.
Check the relationship between each line item and its price heading. A price per pack, per piece or per service visit describes a different basis, even when the item names match. Keep quantity breaks and minimum-order conditions attached to the affected row. Do not carry a unit from the first page across a later table unless the document explicitly makes it apply there.
Trace footnote markers back to their line items and record whether each exception applies to one row, a section or the whole quotation. A freight exclusion in a footer qualifies what an otherwise complete item price covers. Preserve the association in the evidence matrix rather than placing all footnotes in one unrelated notes cell. Also distinguish “not stated” from “excluded”; silence about installation is not evidence that installation is excluded or included.
If a scan is unreadable or a schedule is missing, record that source as incomplete. Never reconstruct a missing commercial term using a competing supplier’s quotation. A complete-looking matrix is less useful than a matrix that shows exactly which comparable evidence is absent.
Related: AI for logistics: prepare shipment exception evidence.
3. Extract terms into a source-linked matrix
Pass the inspected quotations and purchasing criteria into Structured Output through input_value. In system_prompt, require a quotation locator for every material term and prohibit transferring a missing term from another supplier’s offer. Separate the requested delivery date from the date actually promised in an offer; both belong in the matrix but carry different authority.
Prepare output_schema around supplier, source version, requirement, quoted value, unit, delivery phrase, exclusion, source locator and review status, using the extraction component reference for implementation details.
| Term | Preserve from the source | Leave unresolved when |
|---|---|---|
| Price basis | Quoted amount, currency, unit and quantity basis. | Any part of that basis is absent. |
| Delivery | The supplied timing phrase and its condition. | The start event or destination is unclear. |
| Included scope | The items or services expressly included. | An attachment defining scope is missing. |
| Exclusions | Quoted exclusions and additional-charge wording. | The document does not say whether an item is included. |
| Validity and review | The quotation’s own validity wording and buyer-entered decision. | The current availability has not been confirmed. |
Check the matrix across columns as well as down rows: an extracted amount may be accurate but belong to a different quantity tier or item. Verify that source locators identify the price basis and qualification, not just the supplier’s name.
4. Organise comparable evidence without hiding unknowns
Use the Table Operations contract for a buyer’s view, choosing operation "Select Columns" while retaining the unit, exclusion and source-locator columns.
Set columns_to_select only after the buyer identifies which criteria are mandatory. Removing the uncertainty column makes missing terms disappear visually without resolving them. If you later choose "Filter", inspect the excluded rows alongside the review view; an unknown requirement leaves the comparison unresolved until the buyer obtains the missing term, without establishing grounds to reject that quotation. Use "Sort" only within a genuinely comparable basis.
Do not sort unrelated price units into a single recommendation. Preserve supplied amounts without currency conversion or invented totals in this starting design. The buyer can request a like-for-like comparison once the commercial assumptions are established. A table operation changes presentation, not the validity of the underlying evidence.
5. Prepare a review narrative and Google document
Use Language Model guidance to describe unresolved comparisons, and Create Document instructions for the buyer’s review copy.
Supply the checked matrix as text in input_value and set system_message to distinguish supported differences from clarification requests. Ask the narrative to state why two terms cannot yet be compared, rather than ranking suppliers anyway. Set title to identify the purchasing request and quotation set, with initial_text containing the buyer’s accepted explanation; keep the matrix available beside the prose.
The form-to-document test’s 42/42 result with documented CR+LF conversion supports its own validated creation recipe and cannot establish this supplier matrix’s accuracy or the assistant’s execution.
6. How do I check mismatches before a buyer uses AI procurement output?
Test quotations with different units, omitted delivery conditions, an unreadable price table and a missing attachment. The matrix should show the problem and retain its source location. Review any narrative to ensure an unknown was not rewritten as a disadvantage or a supplier endorsement.
In an illustrative comparison, one quotation prices individual units and another prices packs without stating their contents. Record both price bases as supplied and leave the unit comparison unresolved. If the second also says delivery follows receipt of an order but gives no lead time, preserve that phrase without inventing an arrival date. The buyer’s next action is a clarification request for pack contents and delivery timing. Neither the extraction nor a sort operation can supply those missing commercial facts.
Keep purchasing approval, payment execution, ordering and live price commitments outside this comparison workflow. It has no purchasing-system transaction step. When evidence spans public sources, use research report writing; for source currency, use document auditing.
Plans: Go from USD8/month billed yearly (USD10 monthly), Plus from USD16 (USD20 monthly), Pro 100 from USD80 (USD100 monthly), Pro 200 from USD160 (USD200 monthly) and Pro 500 from USD400 (USD500 monthly), with 5,000 to 250,000 Flowi Tokens/month and top-ups; 1 token per node run; third-party API and model usage is not included.
The catalogue establishes the components, while use cases show existing starters. The buyer’s reviewed evidence matrix is the deliverable, with purchasing authority retained in your normal process.
Related: Purchase order approval workflow: a reviewed request checklist; AI for supply chain: connect supplied procurement, stock and delivery evidence.