Commerce Online · Singapore Practical AI for the businesses that already run this economy Singapore · Est. 1988
Commerce Online

The short version


What we touch
One system you already run, left as it is.
What we do not touch
No new schema, no migration, no re-training.
Where answers come from
Your live records, and nowhere else.
How it is judged
A pilot measured on a single number.
01The problem we solve

A better interface does not make the questions easier

Search boxes, filters and dashboards assume you already know where to look and what it is called. Three failures turn up in almost every system.

  • A misspelling ends the search. Type a supplier’s name slightly wrong and the screen returns nothing. The record exists; the search cannot reach it.
  • Some questions cross modules. ‘Everything one vendor has open’ spans four registers. Nobody answers that without reconciling by hand.
  • An empty result explains nothing. No rows and no reason. So the user phones the department that owns the system.
02What you get

An assistant that answers from your records — or says why it cannot

The value is not that it talks. It reaches data your interface hides, and it refuses before it guesses.

  • It resolves meaning, not spelling. A mistyped vendor name still finds the right vendor. A typo stops costing a phone call.
  • It answers across every module. One question, four registers, one answer — from whatever screen the user is on.
  • It refuses rather than guesses. If your records do not hold it, it says so and offers the nearest real thing.
  • Every refusal is a free gap analysis. The questions it cannot answer are the reports your business wanted and nobody built.
  • The model is a setting, not a commitment. Accuracy, cost and confidentiality are dials — including a model on your own hardware.
  • Nothing structural changes. No migration, no change to approvals or permissions. It can be switched off without a mark.
03How one question gets answered

Four registers, one pass, one answer

Nobody reconciles four registers by hand to answer one question. This is what happens instead.

  1. 1The question“What has Sim Heng got open with us?”Typed the way a manager would say it, from any screen.
  2. 2Resolve the meaningA typo still finds the vendorMatched against the real entities in your data, not against the spelling.
  3. 3Read every registerOne pass, four placesWithin what the signed-in user is already allowed to see.
    • Contracts
    • Options
    • Invoices
    • Deliveries
  4. 4One answerWith the records it came fromPlus a plain statement of any part it could not answer.

↑ The step most tools skip

When your records do not hold the answer, the chain stops at the last honest step and says so. It will not invent a figure.

04See it in action

The same system, before and after

Real screens from a procurement system with an assistant added. Read them in order — the third one is the one that matters.

A procurement register showing 0 of 25 records after a vendor name is mistyped.
Fig. 2A misspelling ends the search. The name is typed slightly wrong, so the register returns 0 of 25 records. The contract is there. The search cannot reach it.
The assistant resolving a mistyped vendor name and returning the matching contract.
Fig. 3The same misspelling, answered. The assistant resolves it to the real vendor, answers in a sentence, and shows the contract it drew the answer from.
The assistant declining a question about contract end dates and offering an alternative.
Fig. 4It declines, and says what it holds. Asked for something the records do not store, it names the gap instead of guessing, then offers the nearest real answer.
Why these are not a live client’s screens

The screens above come from a working development build with seeded test records. The product branding and the signed-in account have been replaced.

Before any client’s screens appear here we need their written consent, with the same anonymisation applied.

05How we engage

A pilot you can judge on one number

A pilot is not there to prove the technology works. It is there to find out whether it removes work.

  1. 01The questions people cannot answerWe collect the questions your staff currently pick up the phone to ask.
  2. 02Read-only connectionWe connect through existing permissions. Nothing is copied out or altered.
  3. 03Assistant in the systemIt appears inside the screens your team already uses, answering from live data.
  4. 04Supervised pilotOne team, real questions, a fixed period. Every refusal logged as a gap.
  5. 05Decide with evidenceKeep it, widen it, or switch it off. All three stay open.

Talk to us

Which question does your system refuse to answer?

Most owners can name one immediately — the question that means four screens, a spreadsheet export, or asking the one person who knows.

On the first call

Nothing is connected. We ask what the question is, which system holds the answer, and who currently gets phoned when it cannot be found.