MACH X Amsterdam 2026: Three Hackathons, One Build Lab, and No Revenue Slides

MACH X Amsterdam ran from 28 to 30 September 2026 at Felix Meritis, the eighteenth-century society house on the Keizersgracht that was built for science, art and new ideas. More than 90 member companies were in the building, which the Alliance described from the stage as the most it has ever had in one place. We were there as delivery, not as a sponsor or a member. Fidget Labs co-founder and CTO Kerrigan Baron serves as Delivery Lead for the MACH AI Exchange under contract with the MACH Alliance Tech Office, and the AI Exchange hackathons opened the Technology Track on Tuesday morning.

The Alliance opened Member Day with the number it campaigned on all week. In its research across 600 enterprise decision-makers in seven markets, 78 percent of organizations with scaled composable foundations report clear AI ROI. Among those still in early planning, it is 13 percent. The framing on the slide was that this is an architecture gap, not a model gap, and three days of talks kept arriving at the same conclusion from different directions.

Three hackathons, three different kinds of proof

The MACH AI Exchange runs multi-vendor hackathons against a real brand's real problem. Each one pairs the brand's own team with member vendors and integrators, and the brief is the same every time: build something that works on the brand's stack, then show it to a room that will notice if it does not. Three of this round's builds were presented back to back.

Your Golf Travel: the human versus the agent, in public

Oliver Gunning, Chief Marketing Officer of Your Golf Travel, opened the track by calling himself the least technical presenter of the day and then ran the most confident demo format of the event. He read real customer emails aloud without looking at the screen, gave his own answer, and asked the room to vote on whether his reply or the agent's was better. We ran the live demos from the lectern. Two builds, both put to an audience vote against the person whose job they touch.

• The Greenkeeper reads the sales inbox, works out whether a message is a new enquiry, a returning customer or a complaint, and drafts a reply for a golf travel expert to review, edit or regenerate. A human approves every send.
• Jack is a conversational golf travel advisor, named for Jack Nicklaus, who was not taking calls. It is live on the YGT website and answers from what golfers actually booked rather than from brochure copy. Ocula Technologies built the semantic layer underneath it, and it was the part Oliver singled out from the stage.

The engineering lesson we took home from the Greenkeeper is the one we keep relearning. The rules that kept failing in the prompt held once they moved into code. A deterministic layer runs before any model call, skipping auto-replies, mailing lists and internal senders, and a price check in code caught demo drafts that quoted prices nobody had retrieved. The model writes the prose. The rules are code. That is the same argument we made in Deterministic vs. AI, and it held up in front of a full Concert Hall.

easyJet holidays: mock the data, keep the schema real

Alex Black, Head of Digital Product at easyJet holidays, presented with Adam Hake of Foster Made. The brief was two real customer problems for city breaks: where is my hotel, actually, and what do other people think of it. The build pulled 11,000 hotels across 189 destinations into one Amplitude data model, indexed geography and hotels in Algolia, and used Claude to parse search intent with a deterministic rules fallback behind it.

Two things made this the most transferable session of the three. It defined success before anyone wrote code, at three levels: whether it scales, whether customers book faster than the current average of up to 11 visits, and whether city break conversion moves. And it hit the wall every multi-company AI project hits. Six partner companies cannot be handed a brand's customer data. The team mocked the data where it had to and kept the structures representative of the real thing, so the architecture and the enrichment pipeline were exercised faithfully while no customer record left the building. If you are planning a multi-party proof of concept, design for that on day one.

Dr. Martens: agent-ready in six weeks, because the hard part was already done

The Dr. Martens and Conscia team set out to make drmartens.com buyable by an agent, ready for the day platforms open agentic shopping. A shopping assistant reaches the store through MCP tools, each one wired to a single orchestration flow in the existing backend-for-frontend, so the agent can search, size, read a product and add to cart through the same APIs the website uses.

Read it next to Peter Goggin's talk in the Commercial and Change track, titled "The Monolith Is Not Dead", and the story is clearer. Dr. Martens did not replace its commerce platform. It stopped making it do everything, moved content, orchestration and other responsibilities out to specialist services, and kept the core. That is why the agentic layer took weeks. Being agent-ready turned out to be mostly a composability project with a smaller AI project on top.

The Agent Build Lab: go break it

The most substantive thing the Alliance showed all week came from its Agent Enterprise Architecture working group, co-chaired by Vercel and Contentstack and run entirely in the open. Rather than standardize agent interoperability on paper, the group built two agents and ran them against each other: a pricing agent with standing authority over 50 SKUs, and a buying agent negotiating with four supplier agents from four different companies at once. It worked on open standards, with nobody owning the middle.

Then they published what went wrong, which matters more than what went right.

• A silent credential failure. The supplier with the best prices never received a message, because its credentials failed, and nothing alerted.
• A race that looked like a negotiation. Parallel negotiations turned into a race where the fastest supplier won rather than the cheapest, again with no alert.
• Lost lifecycle control. An agent that had been halted came back running after a restart.

None of those throws an error. Every one reports success. The conclusion from the stage was that interoperability is no longer the hard part. Everything around the agent is. On Tuesday the working group turned the work into a hands-on lab in the Dome Hall, with facilitators from Contentstack, Vercel, Outshift by Cisco, AWS and Orium, and an invitation from the Alliance's president, Jason Cottrell, that was worth the trip on its own: go break it. We have written before about the trust gap after an agent works. This was the same lesson at the protocol level.

What the rest of the program agreed on

Across more than a dozen brand talks and panels, four things came up often enough to count as findings rather than opinions.

• Nobody published agentic revenue. Not one brand put a conversion or revenue number on an agentic result. Dr. Martens noted that no conversion baseline was published for its build, and PhotoSì told the room it did not have results to show yet. That is not a failure. It is an honest market at an early stage, and it means agentic work should be justified on throughput, optionality and risk until the numbers exist.
• The semantic layer is the asset and it has no standard. At least six speakers independently said the model is a rental and the context underneath it is what you own. Hobbii's Tomas Antvorskov Krag put it best: the model stays swappable only while the semantics are yours, not a vendor's.
• A new procurement question. PhotoSì described wiring Claude over MCP to the tools they already run, and the question it left them asking every vendor: "Can an agent use the capabilities we need, with the permissions and controls we require?" Expect that line in RFPs within a year.
• Brownfield is the normal case. Virgin Media O2, Dr. Martens and Solventum all built on top of heritage systems rather than replatforming first. Virgin Media O2 said it plainly: greenfield is a fantasy, and real business happens in the brownfield. The skill is encapsulating the core, not replacing it.

The community part

The hackathon teams are the part that does not fit on a slide. The YGT build alone brought together people from Your Golf Travel, Ocula Technologies, Valtech, Iterable and Fidget Labs. The easyJet holidays team drew on Foster Made, Algolia, Amplitude, OneMagnify and Reply. Competitors in other rooms, one team in this one. In Alex Black's words from the stage, "It's honestly blown my mind how open and generous with their time everyone has been." That matches our experience of running them.

Jason Cottrell's single ask of the room on Tuesday morning was to leave Amsterdam with three introductions you did not have that morning. It is a better measure of an event than attendance. The next MACH X is 12 to 14 April 2027.

If your program is somewhere between the 13 percent and the 78 percent, that gap is what our Architecture Review is built to diagnose, including whether an agent can act through your stack under your permissions. If AI is already live and nobody can say what it returns, that is the AI Impact Audit.