Mantic AI Beats Human Forecasters: What Businesses Should Test

What changed on 18 September?

Confirmed fact: Reuters reported on 18 September 2026 that Mantic had raised $25 million in seed funding. Radical Ventures led the round, with backing from Microsoft's M12, Thinking Machines Lab, Balderton Capital and other investors.

Mantic was founded in 2024 by Toby Shevlane, a former Google DeepMind research scientist, and Ben Day. The company specialises frontier AI models from other laboratories for forecasting, tests the resulting system against historical events, grades its performance and uses those results to improve it.

The timing follows the summer 2026 Metaculus Cup, which closed in early September after 58 forecasting questions. Reuters reports that Mantic finished ahead of every human contestant and all but one other bot, called laertes. Metaculus describes itself as a forecasting platform focused on questions of global importance.

• $25 million seed round announced on 18 September 2026
• London-based company founded in 2024
• System specialises existing frontier AI models for forecasting
• Outperformed every human entrant in the summer Metaculus Cup
• Finished behind one competing bot, according to the published results

What the result proves—and what it does not

Confirmed fact: the competition required forecasters to assign probabilities to questions whose answers would later become known. Mantic was more accurate than the human entrants across that particular set of questions and time period. Reuters says the strong scores achieved by Mantic and other AI entrants marked the first time technology dominated the competition.

One example involved Colombia's presidential election. Mantic reportedly gave eventual winner Abelardo De La Espriella about a 40% chance early in the tournament, while the consensus estimate was roughly 30%. In another question, it avoided following an incorrect majority view about the chart performance of a Shakira song.

Our analysis: this supports a useful claim—that a carefully designed AI forecasting system can beat skilled humans in a defined contest. It does not prove that Mantic will be right about every market, customer or business decision. Contest performance can change, rare events remain difficult and a probability is not a promise.

Why this matters to a small business

Most small-business decisions are made under uncertainty. Will a promotion hit its target? Will a supplier deliver on time? Will a new product sell through its first batch? Owners often answer with instinct, a confident yes or a worried no, then forget what they originally believed.

Our analysis: forecasting turns that vague judgement into something testable. Instead of saying 'the launch should go well', write 'there is a 65% chance we sell 50 units by 31 October'. The precise question exposes assumptions, gives the team a review date and makes it possible to learn whether its judgement improves.

AI can help gather evidence, propose alternative scenarios and challenge the story the owner already wants to believe. The human still decides which evidence is trustworthy, how much risk is acceptable and what action follows from the probability.

Five practical uses for creators and small firms

Start with questions that are measurable, reversible and close enough to resolve within weeks. The aim is not to predict distant world events. It is to improve the quality of everyday decisions and create a record that can be reviewed.

An online shop could estimate the chance that a product sells 60% of its stock within 30 days before placing a larger order. A creator could forecast whether a video will exceed the channel's median 48-hour view count before choosing how much production time to spend. A consultant could estimate the chance that a proposal closes by a stated date and plan follow-up accordingly.

The same method can be used for project deadlines, customer renewals and campaign targets. Keep each question specific and avoid pretending that a neat percentage removes real uncertainty.

• Stock demand before a repeat order
• Campaign results before increasing ad spend
• Content performance before committing extra production time
• Project delivery risk before promising a deadline
• Customer renewal likelihood before planning capacity

The opportunity: turn forecasting discipline into a service

Our analysis: freelancers and consultants could offer a simple decision-review service to businesses that already collect data but rarely learn from it. The deliverable is not a magical prediction. It is a short list of measurable questions, an evidence log, three scenarios and a monthly review showing which assumptions were right or wrong.

A marketer might add probability ranges to campaign plans. An operations consultant could build a supplier-risk register. A virtual assistant could maintain a decision log and remind the owner when forecasts are due for review. The value comes from clearer thinking and consistent follow-through, not from claiming access to Mantic's private technology.

Mantic has not published public self-service pricing or general access details. Reuters says some companies and government agencies have integrated its AI, but the customers were not named. Do not sell Mantic access or imply an official partnership unless one actually exists.

The risks: false precision, weak evidence and high-stakes misuse

A percentage can look scientific even when it is built on poor information. AI may use outdated data, misunderstand the question, copy a herd view or invent a persuasive explanation. Historical testing can also reward patterns that do not survive a changing market.

Confidential business data should not be pasted into an AI service unless the account, contract and privacy controls are appropriate. Predictions about people can introduce unfairness, and forecasts involving credit, employment, health or legal rights need qualified oversight and stronger controls.

Reuters reports particular interest from hedge funds and trading firms, but a competition result is not investment advice or evidence of guaranteed returns. Do not use a chatbot's probability as the sole basis for trading, borrowing or risking money you cannot afford to lose.

One practical action: build a 30-minute forecast card

Choose one real decision due within the next 30 days. Write a binary question with a clear deadline and a result that an independent person could verify. Record the current base rate if you have one, then list three pieces of evidence for success and three against it.

Ask your preferred AI tool to challenge the wording, identify missing evidence and describe optimistic, expected and difficult scenarios. Tell it not to make the decision. Set your own probability and decide in advance what action different ranges will trigger—for example, increase the order only if the estimate rises above 70% after one week of sales data.

Revisit the card weekly, record why the probability changed and score the final result. Repeat this for ten decisions. The useful measure is not whether one prediction was lucky; it is whether the process reduces surprises, catches weak assumptions and improves choices over time.

• Write one exact yes-or-no question and resolution date
• Record a base rate and evidence on both sides
• Use AI to challenge assumptions, not issue the verdict
• Choose an action threshold before seeing the result
• Update the probability and score what actually happened