All case studies

Case study · Engineer · AI platforms & engineering

A governed business agent, in production at a Swiss services SME

How a Swiss services SME replaced back-office overload with a governed agent layer on top of its existing systems — architecture, controls, measured results.

Anonymised case — representative example from real Numezis engagements

The back office was absorbing most of the growth: every new client file added data entry, document filing, reminders and cross-checks across three entities. Process knowledge was concentrated in two key people, and invoices consistently went out late.

Management had received several vendor offers promising turnkey “AI automation”. None answered the concrete questions: where does our clients’ data travel, who approves an accounting entry before it exists, what happens when the model gets it wrong?

The decision on the table

Buy yet another line-of-business software suite — or build a governed agent layer on top of the systems already in place?
01

Sensitive fiduciary data

Financial and personal data of ~450 end clients, subject to the Swiss nFADP and professional secrecy. No data could be used to train third-party models.

02

No in-house IT team

Operations had to run without an engineer on site: simple procedures, explicit alerts, documented recovery playbooks.

03

SME budget

An investment the partners could defend, committed in tranches against proof milestones — not a multi-year programme signed blindly.

04

FR/DE bilingualism

Client documents and correspondence in both languages, with frequent switching within a single file.

What we built

We built an agent layer on top of the existing systems rather than replacing them. The agent reads context (incoming email, documents, accounting entries), prepares actions, and executes them only within explicit boundaries: role-based permissions, a human approval queue and an immutable audit log.

System components

Business agentQualifies incoming email and documents, prepares entries, reminders and draft replies for approval
MCP connectorsControlled access to the accounting software, DMS and mailbox — each tool with its own permissions
Document retrievalContext file by file, segregated by legal entity and by client
Approval queueEvery outbound action (entry, email, reminder) goes through human approval with configurable thresholds
Audit logEvery read, proposal and decision is logged — reviewable by the external auditor
Deterministic rulesVAT, due dates and charts of accounts remain verifiable code, never a model prediction
01

Scoping & baseline 3 weeks

Mapping of actual workflows, measurement of processing times and invoicing delays, data classification, selection of the first scope.

02

Foundations 6 weeks

Identities and permissions, read-only MCP connectors, evaluation environment with anonymised test files, audit log.

03

Supervised pilot 8 weeks

Agent in production on one entity, 100% of actions approved by a human, weekly error reviews and threshold tuning.

04

Rollout 6 weeks

Deployment across all three entities, switch to sample-based approval for low-risk actions, team training.

05

Hardening & handover 3 weeks

Operating procedures, alerting, security review, monthly governance handed over to management.

Measured results

The three-week baseline made it possible to measure the real gap — not a gut feeling. The figures below compare the 90 days before the pilot with the last 90 days of the engagement, on the same scope.

BeforeAfter
Processing time per incoming client file~45 min18–25 mindepending on complexity
Median invoice issuance time12 days3 days
Data-entry errors caught in reviewbaseline−70%on the pilot scope
Client reminders sent on time~60%>95%
Adoption among the staff involved9 / 12weekly usage at 90 days

Figures are rounded, measured on the pilot and then the rollout scope. Anonymised case: a representative example drawn from real Numezis engagements and product development work, not a nameable reference.

This section is part of our editorial standard: no case study without its lessons.

  1. 01

    Instrument the baseline from week one. We consolidated it during scoping, but six weeks of finer-grained reference data would have strengthened the final value measurement.

  2. 02

    Involve the external auditor from the foundations phase. Their review of the audit log during the pilot validated the approach — doing it earlier would have saved two iterations on the trace format.

  3. 03

    Build fewer connectors upfront. Four MCP connectors were delivered; two were enough for the pilot. The other two should have waited for proof of use.