Product updates · 8 October 2026

Intelligent AI agents: 3× faster, nearly 4× fewer mistakes

An Intelligent AI agent answers one question for every company in a mapping, researching each one on the web: does it sell to businesses, is it family owned, has it bought another company. On a long list that’s thousands of answers, and each wrong one keeps a company you should drop or drops one you should keep. Some questions caught it out, and in the same ways: a word that looks like M&A but isn’t, a company that was bought taken for the buyer, a parent group’s deals put on its subsidiary.

Starting today, Intelligent agents run on a newer AI model. In our test it answered each company three times as fast and made 3 mistakes in 40 where the old model made 11. The copilot’s research helpers run on it too.

M&A Activity Signal

Researching

40/40 companies

Verbeek Roboticsverbeekrobotics.nlTrue

Bought Lindqvist Vision, a machine-vision firm, in June 2025.

Noordam Industrial ITnoordam-it.nlFalse

Its “acquisities” are new client contracts. No company bought.

Kessler & Raes Systemskesslerraes.beFalse

Itself acquired by Halvorsen Group in 2024. It has bought no one.

Nearly 4× fewer mistakes

We ran one of our own agents, a True or False question on whether a company has made acquisitions, over 40 companies with the old model and with the new one. We checked every answer with a stronger model searching the web, and by hand where the answers disagreed. The new model got 37 right and the old one 29: 3 mistakes instead of 11.

The new model missed 2 companies that had made acquisitions instead of 5, and called 1 company an acquirer that wasn’t instead of 6. Most of the old mistakes were misreadings: “acquisitie”, Dutch for winning a client, read as buying a company; a company that was taken over read as the buyer; a parent group’s deals put on its subsidiary.

Every answer still comes with the agent’s reasoning. Open a company from the list to read what the agent found.

Open a company to read the agent’s answer and why it gave it.

3× faster

The new model reaches an answer with fewer searches and page reads, about 6 per company instead of 15. In the same test a company took 18 seconds instead of 54, at the median, so a column of answers fills in about three times as fast.

Answers already in a column were written by the old model. To get new ones, open the mapping’s AI Agents page, press Restart on the agent, then Start Research.

Restart, then Start Research: the column is researched again.

The copilot’s research helpers, too

When the copilot sends a helper off to research in the background, for People Research, a Hiring Company or a Comp Benchmark, that helper runs on the new model as well. Nothing changes in how you ask.

Available now

Every agent of the Intelligent type and every research helper runs on the new model now, in every space. There’s nothing to switch on, and agents keep their questions and answer formats. Agents of the Default and External types don’t change.

Answers already in a column stay as the old model wrote them until the agent runs again, and companies researched from now on get the new model’s answers, so a column can hold both until you restart it. Restart deletes the agent’s answers first, and the column fills again as the research runs. The buttons show for people who can manage a mapping’s agents, and you can also ask the copilot to rerun an agent.

The Topliner team