At the end of October I travelled to Gothenburg for ICSOB 2026, the 17th International Conference on Software Business, to present our paper TenderIntell: Designing and Evaluating Multi-Agent Support for Tendering. Three days of talks at Lindholmen, a lot of good hallway conversations, and one very memorable dinner among vintage Volvos. Here is a look back.
The conference
ICSOB 2026 ran from 26 to 28 October, hosted by Chalmers University of Technology together with ELLIIT and Software Center. This year's theme was Software Business in the Age of AI-Native Dominance and Digital Sovereignty, which could hardly have been a better fit for a paper about AI agents in public procurement.
Monday opened with a PhD Retreat in Chalmers' Jupiter Building, followed by a welcome reception at Greksson Deli och Café, reached by the free ferry from Lindholmen to Stenpiren. The main conference on Tuesday and Wednesday took place at the Lindholmen Conference Centre, with parallel tracks in the Pascal and Tesla rooms.
The two keynotes set the tone for the week:
- Ricardo Britto (Ericsson AB): How Much Is Intelligence Worth?
- Miroslaw Staron (Chalmers University): Crafting Quality Software: Design-Level Intelligence of Agentic AI Systems
Our paper: TenderIntell
We presented on Tuesday afternoon in Parallel Session 3, AI-Enabled Procurement and Public-Sector Digitalization, in Room Pascal, chaired by Sergio Rico. The paper is joint work by Muhammad Waseem, Abrar Morshed, Janne Harjamäki and Pekka Abrahamsson.
The problem. A tender proposal is judged from several angles at once. Is it legally sound? Does it meet the procurement requirements? Is it communicated clearly? That expertise usually sits with different people, and checking a proposal against all of it is slow. Asking a single LLM for a verdict hides those angles inside one opinion.
The key idea: a panel of AI reviewers. Building on work on LLM juries (Verga et al., 2024) and LLM-as-a-judge (Zheng et al., 2023), TenderIntell gives the proposal to three specialised reviewers: Legal, Procurement and Communications. Each one scores the proposal alone, without seeing what the others said. We then keep every score and report both the average and the spread.
The spread is the part I like most. When the reviewers agree, you can be fairly confident. When they disagree, that disagreement is itself useful: it points a human expert to the parts of the proposal that need a closer look.
Beyond our session
Agentic AI ran through much of the program. One talk I keep thinking about laid out how to evaluate agent behaviour as a whole system rather than one prompt at a time: define what success means (realistic cases, expected outcomes and boundaries), run repeated realistic trials while capturing both the execution trace and the actual end state, then assess the evidence with code checks and model grading that experts calibrate and review. Failures feed back into the next round: review, improve, rerun.
That loop maps closely onto what we are trying to do with TenderIntell. Our reviewers are agents too, and the hard question is always how far you can trust their judgement.
The best part of any conference is still the people. It was great to catch up with Krzysztof Wnuk and Slinger Roijackers between sessions, and to talk through procurement, agents and evaluation with researchers who are wrestling with the same questions.
Dinner at World of Volvo
On Tuesday evening the organizers took us by bus from Lindholmen to World of Volvo for the conference dinner and awards. Attendees got free entry to the exhibition.
Of course, the Tampere crew could not resist a group photo in front of the cars.
Dinner itself was upstairs in a warm, wood-beamed dining room overlooking the water, with long tables mixing people from different groups and countries. It was exactly the kind of evening where the best conversations of a conference happen.
Once the conference wrapped up, I set off on a walk around the city. Gothenburg's old brick-and-plaster apartment blocks look especially good in late-October sun.
Takeaways and what's next
A few things I brought home from Gothenburg:
- Evaluation is the real question. Building agents is getting easier. Showing how reliably they behave, and when to trust them, is where the interesting research is.
- Disagreement is a feature. Keeping multiple perspectives and their spread, instead of collapsing everything into one score, resonated with people working on very different problems.
- Sovereignty matters for the public sector. For procurement especially, where data lives and which models run on it are business questions, not just technical ones.
Next, we are taking the same evidence-grounded, multi-agent approach to a new domain: resume assessment. More on that soon.
I also brought home some reading for the trip: Accelerating Digital Transformation: 10 Years of Software Center (Springer), edited by Jan Bosch, Jan Carlson, Helena Holmström Olsson, Kristian Sandahl and Miroslaw Staron.
Thank you to my co-authors Muhammad Waseem, Janne Harjamäki and Pekka Abrahamsson, to our colleagues at GPT-Lab, Tampere University, to session chair Sergio Rico, and to the organisers at Chalmers for a wonderful conference. Gothenburg, it was a pleasure. See you at ICSOB 2027!
