Nordglade / Public laboratory

Working experiments

Working demonstrations of AI handling real business tasks. Each experiment shows what information it used, what it checked, what it found, and what still needed a person.

Completed

The system answered, and the record shows what each answer rested on.

Stopped

The records needed to support an answer were not there, so no answer was given.

Human review

The case was passed to a person, with the reason it could not be settled.

Filter by outcome

Experiment 06 Completed 2026-09-04 · 2 min 7 s

Deciding which customer messages a chatbot may answer

It answers from the company's own documents, so the real question is which messages should get that far. Six were checked: two were answered with named sources, one was sent back to be reworded, and three were refused.

Evidence viewInvented customer messages

What it used: The customer's message, and nothing else. The fourteen source documents stay shut until a decision opens them.
What it checked: Whether each message was allowed through, using a written list of phrases first and a language model for anything the list did not recognise.
What it could not support: Two attempts to redirect the service, one question about another subject, and one too vague to answer. None of the four cost a search.

Open experiment →

Experiment 05 Human review 2026-08-31 · 37 s

Finding the three records that needed attention

The system checked 212 reconciled records, filtered out 209, and passed three exceptions to a person for review.

Evidence viewFictional company data

What it used: Sixteen supplier declarations, a 212-line emissions file reconciled overnight, and the filing calendar.
What it checked: Which three records need a person before Friday's return, and how long each has been waiting.
What it could not support: The 41 tonnes between a converted declaration and the company's own record. The conversion is exact; the gap is not.

Open experiment →

Experiment 04 Completed 2026-08-24 · 29 s

Checking overdue invoices against written rules

Six overdue invoices were checked against two company limits. The system showed which rule applied and did not guess the missing contract details.

Evidence viewFictional company data

What it used: Six overdue invoices, two limits the company drew, and Finland's published late-payment rate.
What it checked: Every invoice against both limits, with one written line per invoice naming which one decided it.
What it could not support: Any verdict at all until a person confirmed the contract type, and none on the two a person had already marked.

Open experiment →

Experiment 03 Completed 2026-08-10 · 50 s

Answering from company records, with checks attached

The system answered a question from stored records, verified two figures, and returned the checks and limitations behind its answer.

Evidence viewFictional company data

What it used: The company's stored records, read by the one search tool the record names.
What it checked: Grounding passed, two figures verified, sourcing returned n/a.
What it could not support: That the answer is correct, or that a record gets written every time.

Open experiment →

Experiment 02 Stopped 2026-07-29 · 1 min 58 s

Stopping a project that lacked the required data

An audit found that the proposed predictive-maintenance system had no maintenance history to learn from, so the project was stopped.

Evidence viewFictional company data

What it used: Public-record research, a kickoff brief, and four interviews.
What it checked: Each estimate of hours saved against how ready the company was for it.
What it could not support: Predictive maintenance now: no maintenance history exists for a model to learn from.

Open experiment →

Experiment 01 Human review 2026-07-21 · 65 s

Escalating a decision with missing evidence

The system traced three figures to company records but could not determine whether the 40-tonne reporting threshold applied.

Evidence viewFictional company data

What it used: Nordvind Energy's stored company information, searched with one tool.
What it checked: Three figures traced back to the retrieved records.
What it could not support: Whether 40 tonnes triggers CBAM reporting: the threshold was not in the records.

Open experiment →