The software company that created FT Mercati Go to ftmercati.com

Guide · Commodities and data

AI agent for purchasing: what it is, what it does

What an AI agent does in the purchasing department of a company that buys metals: comparisons, checks, deadlines and the price picture, with clear limits.

By FT Support Updated on 8-minute read

The short answer

An AI agent for purchasing is a software assistant that, when asked in plain language, reads ERP data (orders, contracts, consumption, supplier quotes) together with market data such as LME prices, Milan Indexes, premiums and exchange rates, and prepares the work that takes hours today: comparing quotes, checking clauses and invoices, deadlines, summaries for decision-making. It does not decide for the buyer: it prepares, checks and flags.

  • It brings together two sources: the company's internal data and market data.
  • It pays off most on quote comparison, checks on indexed invoices, deadlines and history.
  • It makes no price forecasts and recommends no hedges: it puts the facts in order.
  • It reads a lot and writes little: orders and requests start as drafts and go through a confirmation.
  • Start with a single task, measuring the time saved.

The purchasing department’s work, seen through the data

People who buy metals and raw materials spend much of the day putting together information that sits in different places, and this is the work where artificial intelligence (AI) pays off most: the LME price on a screen, the premium in the supplier’s latest email, the material requirements in the ERP (the company’s management software), the clause in a PDF signed two years ago, the budget in an Excel sheet.

The final decision, whether to buy now or wait, this supplier or that one, takes experience and stays human. But the work that comes before it is largely searching, comparing and checking: exactly what an AI agent does well.

What it can do, in practice

1. The morning picture. “How do things stand today?” The agent summarises prices and exchange rates for the metals that matter to the company, compares them with the budget and with the prices of the latest purchases, and lists the contracts and premiums about to expire. Five lines, with the real numbers, instead of half an hour spent moving between screens.

2. Comparing quotes. Three suppliers reply in three formats: one in dollars with a premium over the LME price, one in euros at a fixed price, one with different delivery and payment terms. The agent reads the quotes, including from PDFs and emails, and puts them on a comparable basis: same unit, same exchange rate, same delivery terms, with the cost of deferred payment made explicit.

3. Checking indexed invoices. For every supplier invoice whose price is linked to the metal, the system recalculates the price expected under the clause and the agent flags the discrepancies, explaining where they come from: a different period, a different exchange rate, a premium that has not been updated. It is often the task that pays for itself fastest. The mechanics are in the guide on indexation clauses.

4. Deadlines and requirements. Contracts expiring in the coming months, quantities still to be collected, requirements not covered by a contract, premiums to renegotiate. One question, one list.

5. History, one question away. “How much have we paid for the aluminium premium over the last eight quarters, and to whom?” “How has our average purchase price for copper compared with the LME average?” Answers that today take an afternoon of data extraction.

6. Draft documents. Requests for quotation, summaries for senior management, draft orders: the agent prepares them, a person rereads them and sends them.

7. Alerts linked to your contracts. Price alerts already exist on data platforms. The difference is linking them to the company’s situation: not “copper has crossed a threshold”, but “copper has crossed the threshold beyond which the November order goes over budget”.

What it must not do

  • It does not decide purchases or hedges. It puts the facts in order; the choice belongs to the person responsible.
  • It makes no price forecasts and gives no financial recommendations. It can show what has happened and how things stand today, not promise what will happen.
  • It sends no orders or commitments to suppliers without explicit confirmation.
  • It does not make up numbers. Every figure comes from a source, with its date; if a figure is missing, it says so.

What data it needs

Internal data, read-only: supplier records, orders and contracts with their clauses, consumption and stock, purchase invoices, budget.

Market data, from an authorised source: LME prices with history and monthly averages, premiums, exchange rates, Milan Indexes, scrap prices. Data from the exchanges is subject to licence: it has to be taken from an authorised distributor and used as the contract allows. How to bring it into company systems is explained in the guide on how to connect LME prices to your ERP.

This is the advantage of working with the people who distribute the data: FT Support is the company that develops FT Mercati, and knows that data field by field.

How to build it safely

The rules of any well-built agent apply:

  1. A limited set of tools: the agent has a closed list of functions, almost all of them read-only.
  2. The code does the sums: averages, conversions and formulas are deterministic functions; the model calls them and explains the result.
  3. Source and date on every number: no figures without an origin.
  4. Writing is an exception: drafts yes; nothing is sent without confirmation.
  5. Same permissions as the users: anyone who cannot see a supplier’s prices cannot see them through the agent either.
  6. A log of requests, data used and answers.

It is the same design as the CRM with AI agents we use every day.

Where to start

  1. Choose one task, not “the purchasing department”. Checking indexed invoices and the morning picture are the two most common starting points.
  2. Measure today: how many hours a week, how many errors found late.
  3. Check the data: where the contracts are, how the clauses are written, which market data source the company already has.
  4. Run a pilot of a few weeks with a buyer who really uses it.
  5. Measure again, and decide whether to extend it.

The full path is in the guide How to bring AI into your company. If you want to think it through for your own case, let’s talk.

Sources

  1. FT Mercati, Guide to the LME price
  2. FT Mercati, Guide to LME premiums
  3. FT Mercati, Guide to the Milan Indexes

Checked on 17 September 2026. Rules and deadlines change: always check the official source before you decide.

Frequently asked questions

The short answers

What is an AI agent for the purchasing department?

It is a software assistant based on a language model that carries out purchasing tasks on request in natural language: it retrieves data from the ERP and from contracts, cross-checks it against market prices, compares quotes, checks invoices and deadlines, and prepares draft documents. It works with the tools and permissions the company assigns to it.

Can an AI agent decide when to buy metal?

No, and it must not. It can put in front of the buyer everything needed to decide: material requirements, open contracts, prices, premiums, exchange rates and variances against budget. The decision to buy or to hedge stays with the person responsible. A well-designed agent makes neither price forecasts nor financial recommendations.

What data does an agent for the purchasing department need?

Two families of data. Internal data: supplier records, orders, contracts and clauses, consumption, stock, invoices. And market data from an authorised source: LME prices, Milan Indexes, premiums and exchange rates, with history and averages. Read access is enough; you start from the sources tied to the first task chosen.

How do you stop the agent getting a number wrong?

By having the numbers calculated by code and not by the model. The agent understands the request, calls functions that read the data and do the sums deterministically, and then presents the result, stating the source and date of each value. That way every figure can be verified.

Which activity is it best to start with?

A frequent, measurable task. The best candidates are checking supplier invoices against indexation clauses, comparing quotes, and the morning summary with prices, exchange rates and expiring contracts. You measure the time spent today and compare it after a few weeks.

Do you have a process you would like to see working better?

Tell us about it in a first conversation, with no obligation: together we work out whether and where AI brings a measurable benefit, and which step to start with.