AI assistant

You photograph a label or a container. The assistant proposes candidates from our database with a confidence against each, flags the likely hazard class and gives a safe next step.

Reading a faded label is one thing. Knowing what is inside the bottle is another, and no photograph can do that. So the assistant has firm limits. If its confidence is below the threshold, or if the item resembles a peroxide former, picric acid, sodium metal or a pre-1990 container, it recommends a chemical store assessment and blocks self-serve quoting of that line.

The advice to staff stays the same as it would be without any software. Do not open an unknown container to smell it or look inside. Take the photograph with the bottle where it stands.

During an import, the assistant works through free-text names, suggests a match, a storage class and an EWC code, and lists what it could not place. You check its suggestions before they are saved.

A suggested EWC code is a starting point. Working out a waste code properly can depend on what the waste is, where it came from and, for many entries, an assessment of its hazardous properties. Our waste code finder explains why.

You can ask questions in everyday language over your organisation's own data: what expires before July, what is stored with the oxidisers that should not be, or which items have been unopened for three years and could form a disposal list. Answers point to your own rows, so you can check them. The assistant reads what the register holds and cannot see what the register does not know.

The assistant can draft hazardous substances register entries and storage risk notes from hazard data. Each one is marked as a draft for the competent person to read, correct and sign. A draft knows the substance. It does not know your room, your ventilation or who is using the chemical. Those are the parts your competent person adds. More on this is on the alerts and COSHH page.

Using expiry dates, stock check flags and periods of non-use, the assistant proposes what could go on your next collection and shows an estimated collection value before you submit anything. You remove what you want to keep. The estimate is not a quote, and a priced quote only comes through the normal disposal request.

There is a sixth use, in our own office and not in your account. It reads an itemised request, predicts the lab pack count and a likely price band from job history, and pre-fills the price request we prepare. Our staff check the result before anything goes further.

The model is DeepSeek V4 Flash, reached through Hugging Face. We do not use any other paid model API, and we do not use OpenAI's API. If the Hugging Face route is unavailable the assistant returns an error and the rest of the register carries on working. It will not quietly switch to a different provider.

Calls send chemical data and label photographs, with personal data stripped. Try to photograph the bottle alone, so that a name on a nearby form or a face in the background does not appear in the image. The Hugging Face provider may process data outside the UK, so it is named as a sub-processor in our data processing terms.

Language models can be confidently wrong. They may misread a formula, merge two similar names or state a storage rule that does not apply to your dilution. That is why outputs are labelled, logged and held back until confirmed. We would rather you treat the assistant as a quick and tireless junior who needs its work checked. If an answer looks wrong to you, trust your own judgement, your risk assessment and your supplier's safety data sheet.

The assistant is included in the School plan and above.