Why AI must not invent figures in procurement
A language model predicts text, so a figure it writes is plausible and not computed. A savings figure goes into a budget, a negotiation and a report to finance, and anyone must be able to recompute it from its inputs. The division that works is this. A model researches, reads and drafts. Deterministic code computes every figure and shows the calculation. What a model read from a document is checked by arithmetic and labelled until a person has reviewed it. Unverified inputs are labelled and kept out of totals, and a draft in which the model states a figure of its own is discarded.
A language model writes the number that fits the sentence. A savings figure has to be the number that follows from the data. In procurement the difference matters, because the figure goes into a budget, into a negotiation and into the report to finance. So the model must not produce it. Code computes it, and the model does the work around it.
Why a model's figure is not a computed figure
A language model generates text piece by piece and chooses what is likely to come next. Asked what six seats at 30.00 a month cost per year, it will usually write the correct product. Sometimes it will write a number that looks right and is wrong. Three properties rule it out as the source of a figure that is shown as fact.
- It is not repeatable. The same question can give a different answer tomorrow.
- It is not traceable. There is no formula and there are no inputs that someone could check.
- It sounds the same when it is wrong. The reader cannot tell a computed figure from a guessed one.
How to divide the work
The model is useful in procurement wherever language has to be read or written. The arithmetic belongs to code, and the responsibility to a person.
| Task | Model | Code | Person |
|---|---|---|---|
| Find a list price on a vendor's website | Researches it and names the source | Stores it as unverified | Verifies it |
| Read an invoice PDF into line items | Transcribes | Checks the arithmetic | Reviews what was read |
| Interpret an employee's reply | Labels it, for example "still needed" | Records the answer | Can see the original words |
| Compute a saving | Nothing | Computes it and shows the calculation | Reads the calculation |
| Write a message to a supplier | Writes the words | Fills in the figures | Approves the message |
| Compare offers | Reads an offer from a reply into a draft | Prices and ranks the offers | Confirms each offer and awards |
What a deterministic calculation looks like
A deterministic calculation has three visible parts: inputs with their sources, a formula, and a result. Two examples with invented figures:
- Licences: 6 inactive seats × 30.00 a month × 12 months = 2,160.00 a year. The seat count comes from the vendor's activity data, the price from the subscription.
- Purchasing: 2,000 units × (5.00 paid before − 4.60 billed now) = 800.00. Quantity and both prices come from invoices.
Anyone can redo both with a calculator. If the result is wrong, one of the inputs is wrong, and it is possible to see which. That is the property a generated figure lacks.
Three ways an invented figure gets in
- Through a draft. Asked to write to a supplier, a model readily adds "this would save you around…" or "other suppliers are offering…". The guard is to treat the figures of a message as variables that are filled in from the data, and to check every draft: if it states a figure that was not among the facts it was given, a fixed template takes its place.
- Through a document. A model that reads a scanned invoice can misread a digit. The guard is arithmetic. Quantity times unit price must give the line amount, the lines must add up to the net total, and net plus tax must give gross. An invented example: net 1,000.00 plus 19 % VAT is 1,190.00, so a gross amount read as 1,910.00 fails. What was read stays labelled until a person has reviewed it. The invoice audit checklist lists the checks.
- Through a benchmark. "Organisations of your size typically save…" is a figure about someone else. The guard is a rule: a saving nobody has quoted or billed is not counted.
Labels that travel with a figure
A figure is only as good as its weakest input, and the reader has to be told about it.
- Unverified. A price from research or from a catalogue, until a person has verified it. It is shown and not added to a total.
- Read by the model. An amount transcribed from a PDF or an offer read from a mail, until a person has reviewed or confirmed it.
- Whose figure it is. The saving of a finding, or the saving of the lowest complete offer in a comparison.
- Once or yearly. A refund for a duplicate invoice is recovered once. A lower unit price recurs. The two are never added together.
A label has to appear wherever the figure appears: in the list, in the report, in the export and in the mail that summarises the week. How the stages of a saving are kept apart is the subject of proving procurement savings.
Questions to ask about any AI tool in procurement
- Can I see the calculation behind every figure, with its inputs?
- Where does each input come from, and who has verified it?
- What happens when the model writes a number into a draft?
- Are unverified figures part of the total?
- Does the tool show a saving before any supplier has quoted?
- Does the same data give the same result tomorrow?
A tool that cannot answer the first question cannot answer the others either. For offers, the same discipline is described in comparing supplier offers.
How Cureman does it
In Cureman every figure shown to a user is computed by deterministic code and carries its calculation. A model researches prices, interprets replies, labels requests and drafts text, and it never produces a number that is displayed as fact. Seed and researched prices are unverified until a person verifies them and are labelled that way everywhere; a figure that rests on one is shown and not added up. An invoice read from a PDF is checked by arithmetic and labelled as read by the model until a person has reviewed it, and a finding that rests on it never gets high confidence. An offer read from a supplier's reply counts only after a person has confirmed it. The figures of a message to a supplier are variables filled in from the sourcing event, and a draft that adds a figure, a link or a claim about other offers is replaced by the template. A proposal of the buyer has its quantities, prices and amounts computed from invoices, and shows no saving, because nobody has quoted. The full set of rules is on the page about principles, and how it works describes the daily work they shape. The question of who approves what is covered in human approval for AI agents.
Questions and answers
Why can a language model not be trusted with the savings figure?
- A language model produces the text that is most likely to come next, and a likely number is not a computed one. It states a wrong sum in the same confident tone as a right one, and the same question can give a different figure the next time. A savings figure has to be the same every time and traceable to its inputs, which only a fixed calculation delivers.
What is a deterministic calculation?
- A calculation that gives the same result for the same inputs every time, with a formula anyone can read. Examples are seats times price per seat times months, or quantity times the difference between two unit prices. Each input names its source, such as a vendor's seat list or an invoice line.
May a model read invoices if it must not produce figures?
- Yes, as long as what it reads is treated as data and checked. The amounts a model transcribes from a PDF are tested by arithmetic, for example whether the lines add up to the net total, and they stay labelled as read by the model until a person has reviewed them. A structured e-invoice needs no model at all, because it is parsed.
What is wrong with benchmark percentages for savings?
- A benchmark describes other organisations. Put into your report, it becomes a figure although no supplier has quoted and nothing has been billed. A saving should be counted only when it follows from your own data, which means a price someone offered, a seat someone released or an amount on an invoice.
How does Cureman keep model-written figures out of messages?
- The figures of a message are variables that are filled in from the data when the message is rendered. A draft in which the model states a figure of its own, adds a link or makes a claim about other offers is replaced by the template, and the reason is kept. A person who edits the message may write figures, because that person answers for it.