Agents that do the work, files to prove it

Back-office automation is not a new subject in Morocco — RPA has had both its successes and its disappointments here. AI agents change the picture on one precise point: they process unstructured information, which makes up most real-world files. An identity document scanned askew, a handwritten accident report, a supporting document in Arabic followed by a contract in French: where RPA stops, the agent reads, understands, extracts and carries on with the process.

Our approach comes down to one rule: the agent proposes, the organisation decides. Every agent deployed by Hunter BI operates within an explicit perimeter — the actions it can perform on its own, those it prepares for validation, those forbidden to it. This framing is not a rhetorical precaution: it is a direct requirement of the Moroccan regulatory context, where Law 09-08 governs automated individual decisions and where sector supervisors expect a complete audit trail. A refusal of credit or of a claim remains a human decision; the agent has merely prepared the case for it — faster, more completely and more consistently than manual handling.

In banking, three opportunities dominate. KYC and remediation: gathering documents, checking completeness, detecting inconsistencies, preparing the file for the compliance officer — a massive workload that anti-money-laundering requirements keep making heavier. Complaints: qualification, linking to the customer file, drafting a response, escalation by criticality. Credit: pre-processing of files, verification of income supporting documents, a summary for the committee — the analyst keeps the decision, the agent removes the data entry.

In insurance, under ACAPS supervision, the main opportunity is the claims chain: opening the file when the declaration arrives, reading the accident report and the documents, checking the guarantees against the contract, a preliminary estimate, routing to the case handler or the expert. To this are added underwriting — analysis of questionnaires and medical documents in compliance with the sensitive-data regime of Law 09-08 — and documentary fraud detection, where the agent flags the inconsistencies a rushed handler no longer sees. On these processes, the gains observed are measured in days of processing time saved and in case-handler hours reallocated to complex cases.

The agent demo is built in three days; the production agent, however, must survive the awkward cases, the outages of the systems it queries and the internal audit. Our typical architecture comprises four layers. System access: MCP connectors to the core banking system, the document-management system, the claims tool — governed, reusable, with rights specific to the agent, never those of a human user. Business rules: explicitly encoded, versioned, editable by your teams without rewriting the agent. Guardrails: a scope of action, confidence thresholds below which the agent hands over, detection of sensitive data. Logging: every read, every system call, every proposal is traced — this is what makes the agent auditable by your internal control and defensible before a supervisor.

This level of demand explains our position on platforms: we build on OpenAI or Anthropic models depending on the use case, in the cloud or on sovereign infrastructure depending on the sensitivity of the data — both paths are detailed on our dedicated pages.

The first agent is the most important: it establishes — or ruins — the organisation's confidence in the approach. We therefore choose it against three criteria: enough volume for the gain to be visible, a documented process so the rules can be encoded, and controlled risk so any residual error can be caught. Triaging complaints or checking KYC completeness often tick all three boxes; the automatic refusal of a claim, never.

The pilot runs six to eight weeks, with indicators defined before launch: the share of files handled without human rework, the average lead time, the error rate compared with manual handling — because manual handling has an error rate too, and it must be measured to compare honestly. Only then comes the expansion: new processes, new scopes of action for the existing agents, shared connectors. The Moroccan organisations that succeed with their agents are the ones that treat them like new recruits: a trial period, measured objectives, a gradual increase in responsibility.