Few businesses lose money on the big decisions. The supplier choice, the price negotiation, the season plan — those get meetings, reviews and the best people in the room.

The money goes on the small things: a request for quotation that went out and was never chased, a test report that expired while the order was in production, a quote promised "for Friday" and forgotten by the next one. None of them is hard. Each one waits for someone to remember it, and a buying office working with dozens of suppliers cannot remember everything.

For your teams in buying, sourcing, QA/QC and logistics, the answer is a system that keeps the list: which supplier has not replied, which certificate is missing on which order, which reminder is due today. From the Field looks at supplier search, where the list of names is easy and the reply is hard. How It Gets Built follows a document file from the inbox to the shipment.

For you, the same backlog sits in your own inbox: the client who is waiting, the partner who wrote "end of the week" three weeks ago. That is why we built Leeloo, an AI assistant for Gmail and Outlook that tracks commitments in both directions, prepares follow-ups as drafts and sends nothing without your approval. It launches today.

In both places, the follow-up gets sent once something other than memory keeps the list.

Meet Leeloo: Say It, Leeloo Does It →

A sourcing services company asked us this month. Fair question: the platforms already sell cheap AI sourcing agents, such as Alibaba International's Accio Work, launched in March to run requests for quotation and negotiate with suppliers.

The catch is where those agents look. Each searches its own platform, where the top results are often sponsored listings, and where most of the rest are traders, because "manufacturer" is a status suppliers declare themselves. A shortlist built there is mostly a list of advertisers.

Our tool for this client sorts like an experienced China buyer. It searches four platforms at once, including 1688 in technical Chinese, and merges the shops that front the same factory. Sponsored listings are flagged. Each supplier is scored on factory signals: an address in an industrial zone rather than an office tower, workshop photos that match the product, one product family rather than ten, registry data where available. A name that differs between platform and website, or a price far outside the range seen elsewhere, costs points.

Then the reply. Factories ignore "Price? MOQ?". The tool writes each supplier a precise request — specification, quantity, target price, certifications, delivery terms — in English and Chinese, approved by the client's team before sending. Scoring is tuned by hand on the client's products until it keeps what a professional would.

The practical version: search one of your products and check the first ten suppliers' addresses, workshop photos and product ranges. Count how many are factories.

The situation. A buying office ships consumer products to European retailers. Every order needs its file: test reports, certificates, declarations of conformity, arriving by email, as scans or photos, sometimes in Chinese. Someone chases them into a shared folder. Nobody can see which references ship next month without a valid report, and the team usually finds out when the retailer asks for the file.

Strategy — the flow that slips. The costly failure is the document that was valid when it arrived and has expired by the time the goods leave. That is the flow to automate first.

Implementation — fixed rules. Each document is read on arrival: type, supplier, item reference, dates, results. The company's acceptance policy is then applied by fixed rules, not by the model: a test report at most 24 months old at the ship date, a certificate valid on that date, the exact item reference (a variant does not count), the applicant matching the supplier, a signed and dated declaration, any FAIL rejected.

Generation — the matrix and the email. One view of every reference on every open order against every required document: received, missing, rejected, expired or to check. Then one email per supplier listing what is missing, by order and reference, which someone reads and sends.

The second turn. When a supplier pushes the ship date back, the checks run again against the new date and flag the report that was fine last week. When the rules change — the EU packaging regulation has applied since 12 August 2026 — the new requirement becomes one more column.

What it does not do. No legal opinion on whether a document meets a regulation; compliance decisions stay with you. AI comments on standards or laws that the rules do not cover are shown apart, marked unverified. No email leaves without a person reading it.

Scope. Five days on your own files to establish what can be built: €1,500, deducted from the build if you go ahead. Build from €5,000, fixed price set after the assessment. Run, optional, from €600 a month.

See the supplier document example →

Most follow-ups are decided too late, when someone notices the silence. By then the sample date has already moved.

Decide it when you send. Every request to a supplier leaves with one line in your calendar or tracker, written as a plan: if there is no reply by Thursday noon, then I call or message on WeChat. A date and an action, chosen before you need them.

Psychologists call this an implementation intention, and it has been tested many times. The plan takes the follow-up out of memory: when Thursday noon comes, nobody has to remember it or decide what to do, only do it.

Leeloo runs these as missions: Say It, Leeloo Does It →

In a taxi to the airport, you remember the quote you promised, the supplier who never answered and the reply you owe. Leeloo, launching today, keeps the list…

4 p.m. in Ningbo. The Paris buyer asks where the corrected pre-production sample is. The answer sits in three email threads, two WeChat groups and a courier website…

Ask a purchasing team where its week goes: supplier documents, read, checked and typed into another file. Our new site shows which parts AI can take over…

A convenience store chain in Tokyo, a bank in Singapore: five AI lessons from Asia-Pacific, where companies experiment at scale and correct course fast…