Blog, Tech: 27.10.26AI and fiduciary work: where to automate, where to keep the human

When a business owner asks their fiduciary whether it uses AI, they are asking the wrong question. The right question is where AI is used, where it is not, and who stays in control of the decisions that commit their company. At Synergix, this line is not blurred. It is explicit, stable and verifiable. Here is how we drew it, and why the Swiss legal framework left us no choice.

In a fiduciary’s day-to-day work, some tasks must always come out exactly right. A bank reconciliation that finds a real discrepancy must find it every time, whatever the hour and whatever the month. A VAT obligation triggered at a threshold must be triggered exactly at that threshold. A debit-credit balance check must run without variation. These are tasks where the only right answer is strict reproducibility.

There are also tasks that must accept nuance. Reading a supplier invoice in a format never seen before. Classifying a service that could be an entertainment expense, a maintenance cost or a commission. Understanding an ambiguous bank transaction label. These are tasks where the only right answer is the ability to handle what was not anticipated.

These two families of tasks are not handled in the same way. The first calls for code, testable and predictable. The second calls for artificial intelligence, able to absorb variability. A serious pipeline does not choose between the two. It combines them. The whole question is where the line falls, and who draws it.

At Synergix, every decision that carries irreversible responsibility sits behind deterministic code. It is never entrusted to an AI model, however good its performance looks on paper.

Coordination between agents. When a document enters our pipeline, it is not an AI model that decides which other agent will process it. It is explicit logic, with clear rules, defined conditions and predictable transitions. AI decides on the content. The mechanics decide on the flow. This is the direct continuation of the approach we have been developing for a year on artificial intelligence in fiduciary work.

Accounting consistency checks. Debit-credit balance, document number sequences, compliance with the Swiss SME chart of accounts (KMU) adopted by the client, respect of the open tax period, consistency of amounts with applicable VAT thresholds. Each of these checks is code, not an agent’s judgement. The result is always the same on the same data.

Approval workflows. When an entry reaches a materiality threshold, it automatically switches to human validation. This switch is never an agent’s decision. It is a stable rule, known in advance and testable beforehand.

Audit trails and regulatory archiving. Every action is logged with a unique identifier, a timestamp, a context and a result. Art. 958f CO on the retention of accounting records is not complied with through good intentions. It is complied with through a system that, by design, cannot forget.

At Synergix, AI is only used where it provides something code cannot. We have identified four precise areas.

Data extraction from unstructured documents. A supplier invoice arrives in highly varied forms: native PDF, scan, photographed image, office file. No code can cleanly extract the invoice number, amount, date, VAT number and line items with sufficient reliability across all formats. This is the core work of the extraction models we have integrated into our document pipeline. Each field receives a confidence score, and it is this score, then processed by deterministic code, that decides whether the invoice continues automatically or goes to human validation.

Classification in ambiguous areas. Some entries cannot be filed by a simple rule. A service that could be an entertainment expense, a maintenance cost or a commission. A bank movement that could belong to several legitimate expense accounts. In these cases, an AI model proposes the most likely classification, with an associated confidence level. Code then decides what to do with this proposal: process it automatically if the threshold is reached, escalate it to an accountant otherwise.

Natural language dialogue with staff. An account manager who needs to query a supplier’s history, understand an atypical entry or find a similar precedent in a mandate cannot open ten databases and cross-reference the results by hand. This is where a natural language AI agent becomes useful, returning in seconds contextual information that would otherwise take hours.

Interpretation under uncertainty. In edge cases, such as a new supplier with a never-seen pattern, a document whose format falls outside the reference set or a confusing label, an AI agent formulates a reasonable hypothesis and documents its reasoning. This hypothesis is then validated or corrected by a human, and that correction enriches the system’s memory. The next day, a similar case will be handled better.

This precise split between deterministic code and AI is not specific to Synergix. It is emerging in every serious technical team deploying agents in production, under different names. The formulation that has been circulating in the ecosystem for a year fits in one sentence: agents propose, code decides. In practice, this means two things.

An AI agent can propose a decision, formulate a hypothesis or bring judgement to an area of uncertainty. It cannot execute an action that commits a third party, validate a closing, sign a tax return or open access to sensitive data. These actions are properties of the system, guaranteed by code, not verbal instructions addressed to the model.

The difference is fundamental. “You should not do X” is a behavioural instruction, which degrades under prompt injection or at the end of a long session. “You cannot do X” is a technical property, guaranteed by the architecture. It is this distinction that makes an infrastructure governable in a profession built on responsibility.

This architectural choice is not merely an engineering preference. It follows directly from the legal framework in which a Swiss fiduciary operates.

Art. 398 CO imposes a duty of loyalty and diligence on the agent. This duty cannot be delegated to a probabilistic system. The final judgement on what protects the client’s interest always rests with an identifiable professional who bears responsibility.

Art. 754 CO sets out the liability of a company’s management. This liability cannot be borne by a model. It is borne by a natural person, who is held to account. A system that took decisions committing the company without a natural person being able to review them before execution would be legally problematic.

The Auditor Oversight Act (AOA) imposes methodological standards on auditors that include reproducibility and auditability. An audit based on untraceable probabilistic decisions would be incompatible with these standards.

The revised Federal Act on Data Protection (FADP) requires automated processing that produces legal effects or significantly affects a person to be auditable and open to challenge. An opaque AI decision, taken without human supervision and without traceability, does not meet this requirement.

This legal framework does not hold back the digitalisation of the fiduciary function, it steers it. We have discussed elsewhere how FinTech is influencing the fiduciary profession and why this development strengthens, rather than dilutes, professional responsibility.

These four frameworks converge on the same conclusion. A fiduciary infrastructure that took all its decisions through AI would be legally fragile. An infrastructure that used no AI at all would be operationally outdated. The right answer lies elsewhere: deterministic code for what commits, AI for what interprets, human judgement for decisions on the cases that matter.

This split has practical effects that go beyond compliance alone.

Full auditability. Every decision taken by our pipeline can be traced: which rule was applied, which data was considered, which agent was involved, which confidence score was returned, which person validated it. A tax audit, a statutory audit or an internal audit request can be served without retroactive reconstruction.

Clear responsibility. Every entry, every return and every significant decision carries the name of the account manager who validated it. This is not symbolic. The account manager has the power to reject an agent’s proposal, reformulate a classification or ask for a second review. Professional responsibility stays clearly assigned.

Reproducibility on stable flows. On long-running mandates, with recurring suppliers and known patterns, two successive closings on comparable data produce comparable results. Variability only comes in where it is useful.

Performance on variable flows. On non-standard documents, new suppliers and edge cases, AI takes over and absorbs complexity that code alone could not handle without multiplying ad hoc rules.

It is not the amount of AI that distinguishes a modern fiduciary infrastructure. It is the precision with which each component is placed where it belongs. At Synergix, less than 20% of our processes actually rely on AI. This is not a lack of ambition, it is a deliberate architectural position. The remaining 80% is deterministic code, explicit control and human judgement on the cases that commit.

This split is not fixed. It evolves as models improve and as the areas where AI is genuinely useful become clearer. But the governing principle does not change. Our agents propose. Our code and our teams decide. It is this hierarchy that makes our infrastructure governable and defensible. What protects our clients is not the sophistication of the model used, it is the discipline with which it has been put in its place.

This discipline is at the heart of our approach to the role of a fiduciary. It is also what allows us to tell our clients honestly where what technology can do ends and where what will always remain human judgement begins.