Back to News
Article

Deterministic by Default, Probabilistic by Design.

AI should run every pipeline. But not every task inside that pipeline should be generated by AI. The difference between these two statements is where most implementations go wrong.

March 2026 · 4 min read
01

AI Orchestrates. It Does Not Do Everything.

An AI agent can manage an entire financial reporting pipeline. It can pull data from multiple sources, identify discrepancies, generate summaries, and route exceptions to the right people. That is orchestration. It is what AI is good at.

But when it comes time to calculate the actual numbers on an invoice, the AI should not be doing arithmetic. It should be calling a deterministic tool that does. A function that takes inputs and produces an exact, verifiable, reproducible output. The same way a calculator works. No interpretation. No probability. Just the correct answer.

This distinction is fundamental. The AI decides what needs to happen and in what order. Deterministic tools execute the tasks where precision is required. The pipeline is intelligent. The calculations inside it are exact.

02

Hallucination Is a Configuration Problem

The conversation around AI hallucination has become detached from reality. Every article about enterprise AI mentions it. Every sceptic points to it as proof that AI cannot be trusted for serious work. Most of these complaints come from people who have used AI without configuring it properly for their use case.

An LLM will hallucinate if you ask it to do something it should not be doing. Ask it to calculate tax on a cross-border transaction and it will give you a confident, wrong number. That is not a flaw in the technology. That is a flaw in the implementation. Tax calculation is a deterministic problem. There are rules. There are rates. There is a correct answer. The LLM should never have been asked to generate it. It should have been given a tool that computes it.

When an AI system hallucinates in production, the problem is almost always architecture, not the model.

Properly configured AI systems do not hallucinate on tasks that matter. They use tools for calculations, database lookups for facts, and APIs for real-time data. The LLM handles what it is good at: understanding context, making decisions about workflow, generating natural language where appropriate. Everything else is delegated to deterministic functions that return exact results.

03

How to Structure It

The practical implementation is straightforward. The AI agent sits at the top of the pipeline as the orchestrator. It receives a task, breaks it into steps, and for each step decides whether to generate a response or call a tool. Financial calculations, data validation, compliance checks, unit conversions, date logic: these are all tool calls. The agent knows what to delegate because it has been given a clear set of tools with defined inputs and outputs.

Content generation, summarisation, classification, intent detection, natural language responses: these are what the LLM handles directly. The boundary is simple. If the task has one correct answer, use a tool. If the task requires judgement or language, use the model.

This is how modern agentic systems are built. The AI is not a single block that does everything. It is a coordinator with access to specialised tools. The quality of the system depends on how well those tools are defined and how clearly the boundaries between generation and execution are drawn.

04

The Result Is Trust

When an organisation builds AI this way, the trust question answers itself. The numbers are correct because they come from deterministic functions. The workflow is intelligent because it is managed by AI. The outputs are auditable because every tool call is logged with its inputs and results. When a regulator or a CFO asks why a specific number appears in a report, you can trace it to an exact calculation, not a probability.

Organisations that struggle with AI trust have usually built systems where the LLM does too much. They ask it to reason, calculate, look up data, and generate text all in one pass. Then they are surprised when the output is unreliable. The fix is not better models. The fix is better architecture. Let AI do what AI does well. Let deterministic tools handle the rest.

Build systems you can
trust.

Speak with our team →