Faster and more precise AI in Cobrief

Over the past few weeks, we have put a lot of work into the AI in Cobrief. We looked at which model does which job, what the models get to read, and how we ask them for answers. The result is an agent that is faster and gets more done, and answers and extractions that hit the mark more often.

In short

The agent responds more than three times faster and gets nearly twice as much done with the same AI capacity.
Custom Assessment Criteria give more correct answers, in half the time.
Product Pick chooses the right product more often and fills in the right price.
Qualification requirements and deadlines are extracted even more precisely and in more detail.
The brief is shorter and does not repeat what you already see elsewhere on the page.

The agent: faster, and more for your money

The agent has a new language model. We tested it against the previous model on real conversations before switching. The agent now spends around 30 seconds on a task on average, compared with almost two minutes before.

It also uses far less computing power per task. For you, this means the agent gets nearly twice as much done before the AI capacity in your plan runs out. You can write more bid answers, go through more documents and ask more questions for the same price.

Custom Assessment Criteria: more correct answers in half the time

With Custom Assessment Criteria, you ask the same questions of every tender, for example whether you hold the required certifications or how long the contract runs. Cobrief reads the tender documents and answers.

More answers are fully correct and complete. We measured the answers against reference answers for 18 complete tenders.
Large tenders work much better. When the tender fits, the AI now reads the entire document package, not just excerpts. On large tenders with up to 190 documents, this gives 50% more fully correct answers.
Answers arrive in half the time.

Product Pick: the right product and price

Product Pick fills in the price sheet in the tender with products from your product register. We rebuilt it from the ground up, and it now also looks for cheaper products that are just as good.

The right product more often. The AI picks the right product in 84% of rows in our tests.
The right price. The price is correct in 98% of rows in our tests.
Empty row when nothing fits. If no product meets the requirements, the row is left empty instead of getting a product that does not fit.

Qualification requirements: one requirement per row, even more precise

Cobrief extracts the qualification requirements from the contract notice and the documents, so you can quickly see whether you can deliver. The extraction is now even more precise and detailed.

One requirement per row. The documentation sits with the requirement it belongs to, and each requirement appears once.
Clearer separation from contract terms. Contract terms, such as requirements for apprentices or wages, are kept separate from the qualification requirements.
Figures are quoted word for word. Thresholds, quantities and certificates appear exactly as the contracting authority wrote them, for example "no older than 6 months".
More requirements are captured, including those in separate requirement documents.
The link is even more precise. When you click a requirement, you land right where it appears in the document even more often.

Deadlines: precise dates, also from free text

Many deadlines only appear in the text of the contract notice, not in the date fields. This is especially common in notices from TED. Cobrief now reads the full text. In our tests, we get every submission deadline and the other supplier deadlines right in 97% of tenders. If the contracting authority moves a deadline with a new notice or a corrigendum, you get the new date.

A shorter brief

The brief is now under 100 words. It does not repeat the deadline, value and duration, which you can see elsewhere on the page. Instead, it uses the space on what sets the tender apart. Scope is given in the industry's own units, for example number of consultants or kilometres of road.

How we worked

For each feature, we built our own tests with real tenders and reference answers written from the full document package. Every change is measured against them before it ships. We keep some of the tenders aside while we tune, so we can see that the improvements also hold on tenders we have not adapted to.

Then we pick the model that gives the best result for each job, rather than one model for everything. When new models arrive, we can test them straight away.

The changes are already live in Cobrief. If you have questions or feedback, feel free to get in touch.