The centre of the keynote at AI Tech Summit Skopje 2026 was a single question, asked of the same agent in two different ways. I want to tell it properly, because it is the clearest illustration of the demo-to-production gap I can produce on a stage.
The question
A legal-research agent for a professional-services firm — accountants, auditors, tax advisers. It has to answer:
“Under Swedish law, can an auditor be held liable for aiding accounting fraud — bokföringsbrott — by issuing a clean audit opinion despite known deficiencies in the bookkeeping? Cite the actual case law.”
On its own, that is a public legal question. An agent that looks it up and returns a real citation is a beautiful demo. And then a real user adds one sentence:
“…in the engagement we’re looking at right now.”
That sentence is the whole talk. The moment it exists, the question stops being a legal query and becomes client-confidential financial data. The question the model answers now carries facts about a client engagement the firm is legally obliged to keep. The demo and the platform are the same agent answering the same question — and one of them is unshippable for a reason that has nothing to do with the agent’s quality.
Build 1 — what works in an afternoon
One file. One API call. It fits on a slide:
client.beta.messages.create(
model="claude-opus-5",
mcp_servers=[{"type": "url", "url": "https://ratiolex.se/mcp", "name": "ratiolex"}],
tools=[{"type": "mcp_toolset", ...}],
messages=[{"role": "user", "content": question}],
)No local MCP client, no tool-execution loop, no server to host. It really is this short — and it really does work. And it’s unshippable. The question goes to a model this firm does not run, over a network they do not control, carrying facts about a client engagement they are legally obliged to keep. Nothing here is broken. It’s just not allowed.
Build 2 — ConfidentialMind
ConfidentialMind: a private model plus RAG plus MCP, inside the firm’s own boundary. The agent can read a client file and a law library, but the client file never leaves the building. When it runs the keynote question, here is the receipt:
- A registered platform tool —
ratiolex, listed in External MCP Servers, owned and scoped. - It looked. It didn’t remember. The
sok_rattsfalltool call firing, with the citation the tool returned. - The cited answer: NJA 2009 s. 3, Högsta domstolen, 2009-01-13 — 11 kap. 5 § and 23 kap. 4 § brottsbalken (1962:700),
https://ratiolex.se/nja/2009/s-3/. Retrieved live from the tool. Read the case before you rely on the summary.
The answer isn’t the achievement. Getting it without client data leaving the building — that’s the achievement.
Six things had to be true
For that answer to exist on the firm’s own infrastructure, six platform decisions had to hold — the same six from the keynote:
- Containers & Kubernetes — the model is a containerised service the platform scheduled, not a process somebody started by hand.
- GPU orchestration — it ran on the firm’s own GPUs, sized and scheduled by the platform.
- Observability — every model call and every tool call is a traced, logged request. The tool call you watched is an audit record.
- Cost control — the reasoning model is small, because tools do the remembering. Cheap models plus good tools beats an expensive model guessing.
- Security — the client documents were indexed inside the boundary. The question never left it.
- Governance — the MCP server is a registry entry, not a line in somebody’s script: named, owned, scoped, revocable.
The agent took ten minutes. The platform took months. The agent was always going to be the easy part.
A second system, the same six questions
A different problem, the same kind of buyer, the same wall: an invoice exception workflow. AI reads the invoice, flags an anomaly, waits for a human to approve. Flawless when the click comes in ten seconds. Stretch it to three days and the pending item exists only as execution state — nobody can query “what’s waiting on me,” and editing the workflow orphans everything in flight.
The fix is three workflows over one shared table:
- Exception Watch (every 4 hours) — extracts, compares to vendor history. Clean → approved. Anomaly → proposed + alert.
- Review Decision (webhook) — fires only on the human’s click, seconds or days later. Nothing waits in between.
- Settlement Sequence (every 15 minutes) — syncs approved rows to the ledger, escalates anything still waiting as its due date nears.
All three read and write one table. Settlement only claims rows still marked APPROVED, and the ledger post is idempotent on the invoice ID — they agree on the truth because the database arbitrates, not because they coordinate. The two runs on stage:
- Run 1 · clean invoice, +0.1% — clears automatically. Nothing for a human to do.
- Run 2 · anomaly, +47% — flagged, paused, routed to a person. (Threshold: >15% above the 12-month vendor mean routes to a human. Synthetic, illustrative.)
The platform’s job isn’t to remove the human. It’s to know exactly when to ask for one.
The takeaway
Everyone has an AI demo. Almost nobody has an AI platform. The gap is where platform engineering happens — containers, GPUs, observability, security, governance, cost. A demo quietly gets away with one execution path, one owner, no persisted state, nobody paged, one idle card, one retry loop. Six ways it breaks; six decisions that fix them. You don’t need an AI factory to have an AI platform. You need six decisions and a database.
The companion lab builds both versions in 90 minutes; the full keynote recap has the sessions I photographed.