Kalshi, the prediction market exchange that accepts wagers on elections, sports, award ceremonies and an expanding list of real-world events, has built an internal AI agent called Harrison to review the contract language that defines how those bets are settled.
The problem Harrison solves is specific and expensive. Every bet on Kalshi is governed by a contract that defines what counts as a win, what counts as a loss and what happens in edge cases. A contract that says "will X happen by the end of June" needs to specify which time zone, which definition of "happen" and what constitutes authoritative confirmation.
Get the wording wrong and the platform faces disputes, processing errors and the kind of operational disruption that erodes trust in a market built entirely on precision.
Why contract language is harder than it looks
Prediction markets price probabilities. The contracts are binary: an event either happens or it does not. But the real world is not binary. Elections get contested. Award ceremonies change categories. Sporting events get postponed, cancelled or overturned on appeal.
Every ambiguity in a contract is a potential dispute when millions of dollars are at stake. Co-founder Luana Lopes Lara said Harrison tackles some of the "thorniest contract-wording questions" the business faces, functioning as a quality-control layer that reviews language before contracts go live.
Kalshi routes millions of wagers per day through its platform. Each one depends on the contract being clear enough that settlement is automatic and uncontested.
Use case
Kalshi did not disclose technical details about Harrison's architecture or how it integrates into existing workflows. The framing was deliberately understated: an internal tool that reduces risk, not a product launch designed to impress investors.
That understatement is itself interesting. In a market where every company wraps its AI in superlatives, Kalshi built an agent that proofreads contracts and presented it as exactly that.
The use case is narrow, practical and tied directly to revenue protection. A single ambiguous contract that triggers a settlement dispute costs the platform credibility, legal fees and customer trust. An AI that catches the problem before the contract goes live pays for itself the first time it prevents a bad settlement.
Not every AI agent needs to replace a department, disrupt an industry or justify a trillion-dollar valuation. Some of them check whether the contract says "June 30 at 11:59pm ET" instead of "the end of June."
Harrison does that. It might be the most honest AI deployment of the year.
