Case 02 · Automation

Forty orders a week. Two hours of work.

A B2B AI service for the Dutch and Belgian funeral industry. Thirty paying clients, around two hundred funeral directors in the network, more than five hundred portraits delivered. AI-first operations across intake, billing, follow-up and marketing. Run by one person on a self-built agent orchestration platform.

Role
Founder, architect, sole operator
Duration
Live since 2024, AI-first operations since 2026
Team
One operator, six automated agents, freelance editors in the loop
Vertical
Funeral services, B2B AI-as-a-service
Live
memortium.nl
Flow diagram of the Memortium order pipeline: intake from email, phone and service bot on the left, Mono Dash orchestration in the centre connecting Notion, Google Drive, Telegram and Moneybird, a delivery email at the end, and a parallel track of six marketing agents.

What it is

Memortium provides memorial portraits for funeral professionals across the Netherlands and Belgium. A photo arrives, often the last one taken of someone who is no longer here, and within twenty-four hours we return a dignified portrait the family can keep. More than five hundred portraits delivered. The service lives at memortium.nl.

By early 2026 the service crossed thirty paying clients with thirty-five to forty-five orders per week. Around two hundred funeral directors have signed up or trialled the platform. A traditional setup at this volume would need at least four part-time roles. I have none of those. The rule from day one is simple: if a task takes more than two minutes of attention twice, I automate it with an AI-aware workflow.

Three failure modes, all engineered away

  1. Order intake fragmentation. A Telegram-driven intake agent reads order screenshots with vision, identifies the client in Notion, extracts the photo, places it in the right Drive folder, and routes the link to the editor. Hands off in under two minutes.
  2. Manual billing. A billing agent runs on the first of every month, pulls closed orders from Notion, generates invoices in Moneybird, sends them, and posts a summary to Telegram. Zero forgotten invoices in two years.
  3. Follow-up evaporation. A feedback agent runs three days after delivery. A re-engagement agent runs at thirty days for quiet clients. Both responsible for the steady stream of returning customers.

Agents, with the limits designed in

The six agents run on Mono Dash, an orchestration platform I built. Each agent owns one part of a flow and hands off cleanly. Each role was researched before being assigned, with the question "what can this agent credibly do, and where does it stop being honest." A re-engagement agent gets a small reasoning model because the work is repetitive and the failure cost is low. The billing agent gets a stricter approval gate because the failure cost is high. The intake agent gets vision plus an explicit fallback to me when the screenshot is ambiguous.

The point is that an agent fleet is a design problem before it is a model problem. Wrong role on the wrong model is the most common reason agent systems fail in production.

On the model side, Gemini 2.5 handles image generation. Claude Sonnet 4.6 handles reasoning and customer-facing email. Token-aware routing across the fleet keeps the cost in line. The Studio pipeline runs at €0.37 per order in the happy path.

Responsible AI, in code

The principle is identity is paramount, if in doubt do not deliver. Every generated portrait runs through ArcFace cosine similarity validation against the source. Anything below the threshold retries or falls back to a human editor. The system never ships a portrait it cannot stand behind. Designed human oversight in production, enforced by the architecture rather than a disclaimer.

Isometric view of the Memortium operating system: one operator at a desk, a Mono Dash orchestrator at the centre, six labelled agents radiating outward, and a customer dashboard.
One operator, six agents, the architecture is the team.

The human side

The hardest part of this system was never the pipeline. It was trust, on three fronts.

Funeral directors work under time pressure and emotional weight, and they hand over the last photo ever taken of someone. They do not adopt a tool, they adopt a promise. Onboarding is built around that: no portal to learn, no workflow to change. They send a photo the way they already send photos, and the system meets them there.

The freelance editors stayed in the loop by design. The validation gate does not replace their judgement, it routes the doubtful cases to them. Their work shifted from production to quality, which is the part they are best at.

And the families never see the system at all. That is the point. Every design choice upstream exists so that what arrives feels handmade, because the moment it feels automated, it fails.

Adoption here was not a rollout. It was designing the system around the people who were never going to change how they work, and making that the requirement instead of the obstacle.

Outcome

Two years in.

The pricing stays sharp because the operation is sharp. Customers come back because the service is constant and effortless on both sides. The unit economics work because the human work has been engineered out, not because the price was cut.

What this means for you

The same pattern transfers to any process-heavy organisation. Map the process, find the seams, replace each seam with an agent designed for that specific role, keep the human in the loop where judgement matters.

Why it stays

Memortium stays. I run it two hours a week, and that does not change when I take on a larger role.