We are a construction company that built its own execution apparatus.
Nexora is a German Generalunternehmer for energy and infrastructure projects — lightning protection to DIN EN 62305, electrical installation for retail rollouts, PV, heat pumps. Since April 2026 we have run an autonomous agent fleet that does not report on that work. It performs it.
This page exists because that is difficult to believe from a marketing site, and because the numbers below are measured rather than claimed.
The scale
Roughly 215 billion tokens a month: about 200 billion across ~20 engineering seats, plus another ~15 billion routed through OpenRouter to Fable, fast Opus 4.8 and open-source models. Seats are headcount, per-seat volume is read from a session store on 16 July 2026, and the OpenRouter figure converts our monthly bill at blended per-token rates — so every number here has a basis you can ask us to produce.
One seat, measured — Opus 4.8 as the primary model, Fable 5 for fast work, Haiku 4.5 and Sonnet 5 alongside:
And the construction fleet described below — a separate runtime, doing the physical-world work:
Which models, and where
Most companies write “AI-powered” and leave you to guess. Ours is a mixed estate and we would rather be specific about it.
Claude is how we work. Opus 4.8 is our primary model — 33,572 of one seat’s 45,303 API round-trips — with Fable 5 for fast work and Haiku 4.5 and Sonnet 5 alongside. Across ~20 seats that is roughly 200 billion tokens a month. Claude also ships inside client work: a financial dashboard for a tax practice, and the agentic tool-use loop that authors our own dashboard widgets.
We are a ten-person construction company in Saxony. That is the number worth sitting with — not the volume itself, but who is producing it.
A further ~15 billion tokens a month route through OpenRouter — Fable, fast Opus 4.8, and open-source models where they fit. Parts of the construction fleet run on other providers entirely, for historical rather than architectural reasons.
We spell the mix out because a page implying a single-vendor estate would be tidier than the truth, and because the interesting question is not which logo we use. It is what a ten-person construction company is doing that consumes this much inference.
What decides where a workload runs is rarely the benchmark. For a growing share of our clients it is data residency — explained below, because it is the constraint that shapes our roadmap more than any model comparison does.
What the agents actually do
Construction runs on facts that exist only in someone’s head, van, or voicemail. The interesting problem is not generating text. It is closing the gap between what happened on a site and what the project record believes happened.
An outbound voice agent that phones a site lead — every day
Nobody on a construction site types. So an agent calls the site lead each evening, asks what was done, where, for how long, and what material was consumed, then structures the answer into the project file. Bottom-up data capture in a trade where the data otherwise never gets recorded. This is the capability we are least able to describe without sounding like we are exaggerating, and the one we would most like you to test.
Project-file reconciliation against reality
Long-horizon runs that cross-reference supplier correspondence, voice recordings, invoices and schedules against a canonical project record, then reconcile the differences. Characteristic workload: multi-hour autonomous execution with dozens of tool calls against internal systems.
22 scheduled jobs that run whether or not anyone is watching
Daily briefings assembled from calendar, database and document stores; invoice ingestion into our accounting system twice a day; meeting transcripts pulled and filed; a weekly cost-documentation reminder. The fleet is not a chat window that someone opens. It is infrastructure with a cron table.
Cross-project intelligence
A retrieval layer over project documents, correspondence and site recordings, so that a question about one project can be answered with what was learned on another. Vector search over our own corpus rather than a general-purpose assistant.
The constraint that shapes what we build next
We build vertical agents for clients in electrical engineering, recycling management, e-commerce, consulting — and, increasingly, for law firms and tax practices. That last group is where the engineering problem becomes a legal one.
German Rechtsanwälte and Steuerberater are Berufsgeheimnisträger — bound by §203 StGB, which is criminal law rather than data protection. For them, where client data is processed is not a procurement preference. It is a professional-liability question with a custodial sentence attached to the wrong answer.
We want to be precise rather than promotional here, because the naive version of this argument is wrong and every German specialist lawyer knows it. “EU data residency therefore §203-safe” does not hold — a Gutachten commissioned by the German Federal Ministry of the Interior concluded that the decisive factor is not the physical storage location but control over the data. We do not make that claim and would not defend it.
The defensible claim is comparative. §43e BRAO requires protection comparable to domestic processing with special safeguards where it is not — mitigation, not immunity. On that standard, in-region processing with the model provider architecturally outside the data path is demonstrably better than the alternative. That is a claim we can substantiate; the absolute one is not.
The practical consequence: for these clients we deploy into infrastructure the client themselves controls, in-region, and we build and operate on top of it. It is why our roadmap runs through Amazon Bedrock in eu-central-1, and why EU residency — not benchmarks — determines which models we can put in front of a Kanzlei.
Why this page exists
We are a ten-person construction company in Saxony, not a consultancy with a thousand certified architects. We are unlikely to appear in anyone’s partner directory this year. What we have is an unusual deployment: agentic AI executing physical-world work with money, materials and deadlines attached, in a vertical that has almost no representation in anyone’s customer stories.
If you build models and want to know what breaks when agents leave the chat window and start phoning people, we are a useful conversation. We have been running this every day since April and we keep the receipts.