Private banking: Creand rebuilds Merkaat with Claude in the delivery and inside the product

Context

Creand runs Merkaat, its digital wealth management product, as a standalone platform separated from the online banking channel. Its customer base had shifted: most Merkaat users are already customers of the bank, and the separation forced constant assisted support during onboarding. The bank itself identified that operational cost as the ceiling on its ability to grow the service.

Meanwhile, inside its online banking, Creand had already digitised and standardised pieces — unassisted digital onboarding, digital signature, document pack generation, KYC remediation — that the standalone product could not reuse. The same work twice, and a worse experience.

The project was a rebuild from zero: replace the standalone platform with a fully integrated journey inside online banking — contracting and day-to-day operation, web plus iOS and Android — across two regulatory jurisdictions with genuinely different operating rules, integrated with the banking core and with the MiFID II suitability test end to end.

And with an uncomfortable reference overhead: the previous Merkaat already existed. There was no need to estimate what building this costs, because the bank had already paid for it once.

This is regulated construction. The suitability test, the document pack, the signature matrix and the per-jurisdiction eligibility rules are not features you can approximate.

Collaboration

Claude, Anthropic’s AI, is on both sides of this project: in how it is built, and in what is built.

In how it is built

Thinkia brought its AI-SDLC methodology packaged as a Claude Code plugin, installed in the delivery team’s environment. It is not a coding assistant bolted onto a normal project: the methodology itself runs as governed workflows inside Claude Code. A command loads the workflow graph, resolves and validates its inputs against the project tree, compiles a context pack and executes each step through a version-pinned skill with a declared output contract and health checks.

Some 43 governed workflows are available as commands, backed by around 306 atomic version-pinned skills. This project’s sessions ran on Claude Opus 5 and Claude Fable 5, alongside MCP servers for design and document work. Six people on the delivery team work with Claude Code: it is not one person’s tool, it is how the team delivers.

Findings are graded. A blocker halts the run and nothing persists. An open question stops and asks a person: this project runs supervised. Warnings remain as named considerations that travel with the artefact. Nothing is written to its final path until its gate passes, and the gate is enforced by a Claude Code hook, not by team discipline: a run whose gate has no fresh health report stops hard. Every run leaves an immutable record. Agents produce; humans sign.

Business validation runs on the same rails: the business annotates the capability documents, feedback is applied to the underlying model and not just to the document surface, the document is regenerated, and coverage is checked adjustment by adjustment.

In what is built

Merkaat is not a system built with Claude: it is a system that runs on Claude. At account opening — the exact point where the bank identified its growth ceiling — Claude extracts the documentation the customer submits, processes it, triages the information and understands the content provided. That is what lets an existing customer contract without assisted support: the work of reading, classifying and understanding what arrives stops being a queue of people.

And it is subject to the same governance as the rest of the product. In a MiFID II build across two jurisdictions, a model that reads customer onboarding documentation is not a loose feature: it is a component with traceability, limits and defined human accountability.

Outcome

Scope and schedule — the result of replanning and commercial close by the account team, not a measurement: the same scope, replanned from 12 months to 5, at 70% less budget. And the qualifier that has to be said: because code generation had not started when those figures were closed, both are the product of replanning rather than of a finished delivery.

Verifiable in the method, not in the customer — a Thinkia measurement over the project artefacts, describing how the delivery platform works rather than a Creand result: around 43 governed workflows executed as Claude Code commands, backed by some 306 atomic version-pinned skills. Every artefact closes in a gate signed by a named person, and no gate has ever been signed by an agent. The gate is not a written norm: a Claude Code hook stops any run without fresh evidence.

What changes for Creand: the digital investment product leaves a standalone platform and enters the online banking channel, so existing customers contract and operate without assisted support — the operational cost the bank itself had identified as its ceiling for growing the service. The business validates on navigable prototypes before anything is built, instead of reading documents and discovering the gap at acceptance testing. The design language was reconciled against Creand’s real online banking product rather than inferred from screenshots, so what was prototyped translates one to one and there is no second design pass during build. And every requirement is traceable to the artefact that implements it, so a regulatory change is scoped rather than estimated — in a MiFID II product across two jurisdictions, that is the difference between a controlled change and a re-audit.

Status, stated plainly: Merkaat is in production, and the project continues on a 2026 horizon. The schedule and budget figures are planning figures, not those of a closed delivery, and this page does not present them otherwise.

Case metrics correspond to specific projects and do not constitute a promise of results; each context produces its own range.