Just in time for the 64th sessions of the Subsidiary Bodies (SB64) under the UN Framework Convention on Climate Change (UNFCCC), we launched NegotiateCOP 2.0. SB64 ran from June 8 to 18, 2026 at the World Conference Center Bonn, bringing together negotiators, technical experts, and observers for the most substantive preparation round before COP31 in Turkey later this year.
For those less familiar with the UNFCCC calendar: the Subsidiary Bodies sessions operate quite differently from the COP itself. No pavilions, far fewer side events, a smaller crowd, and a noticeably more technical atmosphere. The focus shifts from the broad political side events of COP toward concrete negotiating work with preparations and thematic alignment with the incoming presidency’s priorities. That tone was exactly the right context to test a tool built for people in the negotiating room rather than those watching from the sidelines.
June Climate Meetings Conference Hall
© Lara Murillo (https://www.flickr.com/photos/unfccc/55344442946/)
Consolidating our climate negotiation tool ahead of SB64
After launching NegotiateCOP at COP30 in Belém last November – a tool that itself drew on the architecture and early lessons from NegotiateAI, our prototype for the UN Plastics Treaty negotiations – we spent the following months assessing and evaluating the feedback we received. The enthusiastic kind and the critical kind.
The result was a release with one guiding principle of “consolidation over new features”. That also meant that we temporarily took the Portal Chat offline. That feature let users query party submissions through a conversational “chat” interface. We did not take it down out of disinterest, but we decided for an approach that we would rather offer less that works reliably than more that half-works, and the Portal Chat deserves a proper rebuild rather than a patch. It’s probably coming back in version 3.0 for COP31, with a RAG architecture that does it justice.
What we focused on instead:
- A much larger dataset. Biggest achievement for version 2.0. The previous prototype only covered COP29 and COP30 documents. We’ve now integrated the full dataset available on the official UNFCCC Submission Portal, covering not just all COPs but all Subsidiary Bodies including CMAs, SBIs, and beyond. That’s a substantial jump in scope and brings the tool much closer to something genuinely comprehensive.
- Filters that actually work. All filters can now be combined freely and reliably, making it significantly faster to narrow down and analyse specific submissions.
- More metadata. Upload dates and mandate information are now included for each document. That sounds minor until you’re trying to understand whether something was submitted before or after a key negotiating milestone.
- A smoother workflow between features. The transition from the Submission Browser to Position Comparison is now fluid: once you’ve selected a submission, you can move straight into comparative analysis without losing your bearings.
Try NegotiateCOP 2.0 yourself: https://negotiatecop.org/
Side Events and Bilateral Conversations at SB64
Beyond introducing the tool directly to delegates, youth activists, NGOs, and other interested parties through bilateral conversations, we also organised and participated in some events worth recapping.
On June 11, we co-hosted an in-person side event with Climate Policy Radar (CPR) to an audience of around 35 participants. We ran live demos of both tools, walked through the underlying architecture and design decisions, and had a real discussion rather than just a pitch session. We’re genuinely glad to have shared a stage with CPR. They’re one of the most experienced organisations working on AI-powered climate policy infrastructure, and our two tools complemented each other naturally. The one-sentence distinction take away for differentiation: Climate Policy Radar offers a semantically queryable database of climate policies from across the world at national and subnational levels, covering the implementation side of things downstream from negotiations, whereas NegotiateCOP focuses exclusively on the negotiation process itself and the documents it produces. Different layers, different use cases, and a sensible pairing.
On June 16, we joined a side event hosted by CPR, which felt like a full-circle moment after co-hosting with them the week before. From the contacts made during SB64, we’re already in early conversations about planning a high-level side event at COP31 in Turkey, with the goal of presenting a more fully capable version of the tool to a broader audience once the remaining roadmap items have been worked through.
Some personal remarks
Something I keep recognizing over and over again: compared to 2-3 years ago, we’re definetly not alone anymore, but still in a minority in building domain-specific AI tooling for UN negotiation processes, which continues to surprise me.
Especially if we try to be honest about the current situation. Negotiators are already using AI to an extend that might surpirse you, and not in some niche experimental way. Many use it routinely for summarising submissions, translating, comparing positions, identifying overlapping language, and tracking how a party’s stance has evolved. I have seen the most interesting workflows, which include 1. automated transcription via recording on a phone, 2. going into translation software and 3. being handed into an AI-based summarization automation – all in one pipeline visualized on a Laptop screen. That’s just the reality of working life in 2026, especially in information-heavy negotiation processes, language barriers and so on. But most of this is happening through general-purpose tools without domain knowledge, without any understanding of UNFCCC document logic, and without architectural constraints. But that matters more than it might seem at first.
When a negotiator uploads a submission to a generic AI provider and asks for an analysis or comparison, several things can quietly go wrong. The model may confuse document types that carry different weight and function in the negotiation context. A submission talking about deforestation is maybe correctly “linked” to thematic neighbours like REDD+ or TFFF, but on other topics the link might be not so clear. It may flatten the precision of UN language, where individual words carry years of negotiated meaning: the difference between “shall” and “should”, or between “common but differentiated responsibilities and respective capabilities” reproduced faithfully versus paraphrased into something that sounds similar but lands differently at the negotiating table. Generic models also lack provenance and version awareness, meaning a model may quietly reason over an older draft of a document without flagging it.
There’s a reproducibility problem too, with the same prompt run twice against the same documents yields different results, which is fine in many contexts and genuinely problematic in this one, where delegations need to be able to retrace their analytical steps. And finally there’s the data question. Uploading sensitive negotiating positions to commercial APIs raises legitimate questions about where that information ends up. On a less dramatic note, PDF uploads to general-purpose tools frequently hit token limits, fragmenting the analysis in ways that are easy to miss and easy to act on incorrectly.
NegotiateCOP is definitively not a silver bullet!
We know that, and we’d rather say it openly. It is a free-to-use tool, without personal data tracking, with an architecture designed to keep the LLM’s general world knowledge separate from the party-submitted documents it’s reasoning over. It’s built around the specific logic of the UNFCCC negotiation process. There are still gaps, and honest feedback is always welcome.
But it’s worth being explicit about who we’re building for. Large parties and major negotiating blocs come to the table with legal teams, analytical units, and dedicated staff covering individual agenda items in depth. Small island developing states, least developed countries, and many negotiating groups fromt he Global South can often send only a handful of delegates responsible for the full range of an agenda that spans dozens of items simultaneously. The gap in analytical capacity between those two realities is significant, and it doesn’t close through negotiating skill alone. A tool that meaningfully reduces the time spent striving through hundreds of pages of submissions (even partially) has different stakes for a three-person delegation than it does for a team of thirty. That’s the context that shapes our design decisions, and that’s what we keep in mind when prioritising what to build next.
The goal remains what it has always been: not to replace negotiators, but to let the AI handle the “volume work” so that the people in the room can spend more of their time on the substantive thinking that actually requires them.
Against this background, we are strongly committed to further supporting our project partners at the BMZ Data Lab, the BMUKN Data Lab, and the Innovation Data Lab of the Federal Foreign Office. As part of the whole-of-government approach, we are continuing to work on delivering a robust, improved, and user-friendly Version 3.0 of “NegotiateCOP” for COP31. If you are interested in participating, please feel free to contact us. If you are a negotiator or know anyone for whom this tool might be of interest or helpful, please do not hesitate to reach out as well.
Thumbnail-Image: UN Climate Change | Lara Murillo (https://www.flickr.com/photos/unfccc/55341551117/)
