How Fiskl AI Works

This guide explains how Fiskl's AI works, from specialised accounting models to the controls on every AI action. Fiskl is a global accounting platform built on specialised AI, financial intelligence, agentic orchestration and deterministic accounting execution.

Fiskl combines proprietary accounting decision models, retrieval-grounded financial context, interchangeable external language models and a deterministic accounting engine in one platform. Its AI classifies transactions, matches payments, reconciles accounts, reads documents and detects exceptions. It also coordinates complex financial workflows and carries out authorised accounting actions.

Fiskl's intelligence operates within the accounting system itself. Fiskl checks each decision against financial outcomes, traces it to its sources and uses it to improve future decisions.

What Makes Fiskl AI Different

Fiskl brings together four capabilities that financial software often keeps separate.

1. Specialised accounting intelligence

Fiskl develops its own AI models for accounting decisions. These models are trained on more than 500 million accountant-certified transactions. They handle high-frequency tasks such as classification, matching, reconciliation, document understanding and anomaly detection.

General-purpose language models add broader reasoning alongside Fiskl's specialised models and deterministic accounting logic.

2. Context-grounded reasoning

Financial decisions depend on the relationships between records. A payment may relate to an invoice, a client, a bank transaction, a currency conversion, a tax treatment and several ledger entries.

Fiskl assembles the relevant financial context and combines it with retrieved knowledge through retrieval-augmented generation (RAG). AI then reasons from your company's actual financial position, transaction history, supporting documents and applicable accounting guidance.

3. Controlled financial execution

Fiskl connects AI decisions directly to accounting capabilities. AI retrieves information, analyses transactions, prepares changes and initiates authorised actions. Permissions, accounting rules, approval requirements and deterministic validation decide what happens next.

Fiskl's deterministic engine performs every accounting calculation and enforces ledger integrity. AI reasoning connects to financial execution through these controls.

4. Recursive learning and continuous improvement

Fiskl compares AI decisions with later outcomes, corrections and exceptions. These signals improve transaction categorisation, decision accuracy, confidence calibration, knowledge retrieval and orchestration.

Fiskl separates immediate adaptation to your company's transaction history from controlled updates to its production models. Production model updates are evaluated and approved before release.

The Fiskl AI Architecture

Fiskl's architecture has seven connected infrastructure layers, with a continuous improvement system running across them.

A workflow moves between the layers as the task requires. It may pass through financial context, specialised models, external language models, accounting tools and validation several times before an action completes.

Layer 1. Financial Capabilities and Agent Tools

Purpose: Give AI structured access to real accounting operations.

This layer exposes Fiskl's accounting functions as defined, permission-controlled capabilities. Examples include:

• Classifying transactions and assigning accounting treatments
• Matching incoming payments to invoices
• Reconciling financial accounts
• Processing invoices, bills and receipts
• Retrieving financial statements and account balances
• Creating or updating permitted accounting records
• Preparing financial actions and workflow updates

For AI systems, these capabilities act as structured tools. An authorised agent invokes the right function through the platform, so no one has to carry out the steps by hand.

AI contribution: Tool calling and structured execution take AI from conversation into operational accounting workflows.

The Fiskl app, Fi and authorised external systems use these capabilities through supported interfaces, including the API and MCP.

Layer 2. Financial Context and Knowledge

Purpose: Ground AI in the financial entities, relationships and evidence each task needs.

Accounting data is relational. A single transaction can connect to several documents, ledger accounts, payments, counterparties, tax treatments and reporting outcomes.

Fiskl's financial context layer assembles these relationships into a structured view that models and agents use. It combines:

• Financial ontology: A consistent representation of accounting entities, concepts and their relationships. These include transactions, invoices, payments, ledger accounts, documents and tax treatments.
• Relationship and causal traversal: Following recorded relationships and action history to see how a transaction or correction affects other records and reported balances.
• Context retrieval: Selecting the financial history, account information, supporting evidence and current ledger state a task needs.
• Retrieval-augmented generation (RAG): Bringing authoritative internal and external knowledge into the reasoning, including tax guidance and accounting documentation.

Fiskl keeps jurisdiction-specific tax knowledge current through scheduled updates. See Tax Intelligence and Nexus.

For example, take a payment to a vendor. Its context can include the bill, the matching bank transaction and the vendor's history. It can also include the currency conversion, the tax treatment and the resulting journal entries.

AI contribution: Context grounding makes decisions more relevant and explainable, because model reasoning connects to authoritative financial data and retrieved knowledge.

The ledger remains the authoritative financial record. The financial context layer provides a structured view of its relationships.

Layer 3. Proprietary Accounting Decision Models

Purpose: Apply purpose-built AI to accounting decisions at transaction and workflow level.

Fiskl's specialised models are built for financial tasks. They include:

• Transaction classification: Identifying ledger accounts, tax treatments, merchants and transaction characteristics.
• Payment and invoice matching: Recognising relationships between bank movements, invoices, bills and settlements.
• Reconciliation: Identifying corresponding financial entries and resolving matching scenarios.
• Document understanding: Extracting, interpreting and validating accounting information from invoices, receipts and other documents.
• Anomaly detection: Identifying unusual patterns, inconsistencies and exceptions that need attention.
• Accounting decision support: Evaluating transaction context and recommending an accounting treatment.

These models are trained on more than 500 million accountant-certified transactions.

Each model output includes a confidence signal. Fiskl uses it to decide how a decision is presented, reviewed or processed.

For automatic transaction categorisation, Fiskl uses three confidence bands:

• Green: 80% confidence and above
• Amber: 60% to 79%
• Red: Below 60%

Assignments apply as transactions enter the ledger. The confidence bands show you which assignments to check or correct first. See How Auto-Categorisation Works.

AI contribution: Domain-specialised models make recurring financial decisions efficiently and consistently, so accounting-specific tasks rely less on general-purpose language models.

Layer 4. LLM Reasoning and Orchestration

Purpose: Coordinate specialised models, external language models, retrieved context and accounting tools to complete complex tasks.

Many financial tasks need more than a single prediction. A request such as "Explain why cash flow deteriorated this month and identify what we can do about it" can involve several operations:

1. Retrieve current and historical financial data.
2. Examine receivables, payables and cash movements.
3. Compare relevant periods and identify material changes.
4. Retrieve supporting transaction context and documents.
5. Use specialised models and reasoning to evaluate possible explanations.
6. Prepare recommendations or authorised follow-up actions.

Fiskl's orchestration layer coordinates these steps.

Fi, Fiskl's AI CFO, is a conversational and agentic interface in more than 200 languages. Fi interprets requests, plans workflows, retrieves information, invokes accounting capabilities and combines the results.

Fiskl also uses interchangeable external frontier language models for broader reasoning, language understanding and complex interpretation. Each component has its own role:

• Fiskl's specialised models make defined accounting decisions.
• External language models add complementary reasoning and interpretation.
• Fiskl's accounting engine performs deterministic financial calculations and applies accounting rules.
• The orchestration layer selects and coordinates the right components.

Fiskl can therefore replace or upgrade external models without changing the ledger, accounting rules or historical records.

Where Fiskl's own models leave a decision unresolved, an external model can suggest a treatment. The suggestion is subject to Fiskl's confidence levels and accounting controls. This optional step can be turned off in Settings > AI & Automations.

AI contribution: Multi-step reasoning, model routing, context selection and tool orchestration let Fiskl handle tasks beyond the scope of any single accounting model.

See Fi, Fiskl's AI CFO.

Layer 5. AI Governance, Policy and Execution

Purpose: Turn AI decisions into authorised accounting actions while enforcing financial controls.

Reasoning and execution are separate responsibilities. An AI recommendation needs permission before it becomes an action.

Fiskl applies:

• User and agent permissions
• Accounting policies and business rules
• Approval requirements
• Locked accounting periods
• Validation checks
• Action provenance and audit records

AI actions operate within the permissions of the relevant user or integration.

For example, an agent can identify an accounting adjustment, assemble the evidence and prepare a journal entry. If approval is required, the entry waits for an authorised user before it is recorded.

Fiskl blocks AI changes inside a locked accounting period until an authorised person unlocks it.

The Fi Close Agent shows this in practice. It identifies period-end issues, prepares corrections, obtains approval and checks the resulting accounting state.

AI contribution: This layer provides the controls that turn model decisions into governed financial operations, with a person involved where required.

Layer 6. Deterministic Accounting Engine

Purpose: Turn authorised decisions into mathematically and accounting-correct financial records.

AI reasoning is probabilistic. Fiskl's deterministic accounting engine keeps every calculation exact and enforces ledger integrity.

The engine implements the accounting rules that calculate, validate and record financial activity. Its capabilities include:

• Double-entry accounting and journal validation
• Multi-currency accounting and FX calculations
• Tax treatment and calculations
• Accounts receivable and payable
• Inventory and cost of goods sold
• Fixed assets and depreciation
• Revenue recognition
• Accounting periods and period locks
• Financial reporting under supported GAAP and IFRS requirements

Before an entry is saved, the engine applies the relevant validations.

AI contribution: The engine gives AI a reliable execution environment. Models interpret evidence and propose treatments, and the engine enforces accounting calculations and ledger rules in code.

Layer 7. Financial System of Record and Audit Infrastructure

Purpose: Hold the authoritative financial record and the evidence to verify AI-assisted operations.

Fiskl's ledger records your financial position and the relationships between transactions, ledger accounts, documents, counterparties and accounting entries. It provides the authoritative data for context retrieval, AI models, reporting and execution.

The audit infrastructure records each financial change and its origin, including user actions, AI actions and request IDs. These records support:

• Traceability from source documents and transactions to financial entries
• Verification of recorded accounting outcomes
• Review of AI and user changes
• Investigation of exceptions
• Financial reporting and audit evidence
• Evaluation of AI-assisted decisions

A request ID links all the changes that belong to one recorded action.

AI contribution: The ledger grounds model decisions in your actual financial position. It also lets Fiskl compare predictions and actions with verified accounting outcomes.

See Audit Log.

Recursive Learning and Continuous Improvement

Fiskl's learning architecture connects AI decisions to validated financial outcomes, so the system becomes more accurate, relevant and effective over time.

In Fiskl, a correction updates the transaction or journal entry and also becomes a feedback signal. For example, Fiskl assigns a transaction to one expense category, and an accountant then confirms a different one. The signal includes the original prediction, its confidence, the transaction context, the correction and the verified treatment. Fiskl evaluates these signals to find where the decision process should improve.

The closed-loop learning cycle

1. Capture: Fiskl records the input, the relevant context, the model prediction and its confidence. It also records the selected action, any human intervention and the accounting outcome.
2. Validate: Fiskl compares the decision with ledger validations, completed workflow results, exceptions and authorised corrections. This separates a model prediction from a confirmed accounting result.
3. Evaluate and learn: Fiskl analyses the validated signals to find recurring errors, weak classifications, unresolved matches, confidence-calibration issues and gaps in retrieved context. The findings drive improvements to models, retrieval, orchestration and decision policies.
4. Improve and release: Fiskl applies tested improvements through company-specific rules, refreshed retrieval knowledge, updated orchestration logic and approved model releases. Changes are evaluated before they are deployed broadly.

Three ways Fiskl improves

A. Immediate company-specific adaptation. Fiskl uses your completed transaction history to improve recurring categorisation and to suggest rules. When you treat the same merchant consistently across completed transactions, Fiskl suggests a rule. This adapts to your accounting patterns without retraining the global model.

B. Specialised model improvement. Validated predictions, corrections and exceptions are used to evaluate Fiskl's accounting models. They drive improvements in classification, matching, anomaly detection and other specialised capabilities. Each model update is tested against evaluation datasets, accuracy and error metrics and regression tests, then released under control. A single correction never changes a production model directly.

C. System-level optimisation. The same feedback improves components beyond the specialised models:

• Context retrieval: Improving which financial records, documents and knowledge are selected for a task.
• RAG quality: Evaluating retrieval relevance, source grounding and answer quality.
• Model routing: Identifying when a specialised model, external language model or deterministic tool fits best.
• Confidence calibration: Improving how model confidence relates to observed outcomes.
• Exception handling: Identifying when further evidence, another model or human review is needed.
• Orchestration: Improving task plans, tool selection and workflow reliability.
• Policies: Refining automated handling and approval thresholds within authorised controls.

Each of these improvements is evaluated and approved before it changes how Fiskl works.

What recursive improvement achieves

Fiskl connects financial understanding, decisioning, execution and outcome verification. Each completed accounting workflow provides evidence for handling future workflows better. Over time, more decisions complete accurately, repeat exceptions fall and more tasks run with less intervention.

The objective is measurable improvement in accounting decision quality and execution reliability, with the integrity of the financial record maintained.

From Financial Question to Verified Accounting Action

Consider a request to investigate a payment that has not reconciled correctly.

1. Understand: Fi or an automated workflow identifies the request and the affected financial entities.
2. Retrieve context: Fiskl assembles the payment, related invoices, bank transactions, documents, currency information and accounting history.
3. Decide: Specialised matching models evaluate possible relationships and return candidate decisions with confidence signals. Broader reasoning is invoked where needed.
4. Validate and authorise: Fiskl checks the proposed treatment against accounting rules, permissions, period locks and any approval requirements.
5. Execute: The deterministic accounting engine applies the authorised match or accounting changes.
6. Verify: Fiskl checks the resulting ledger state and records the action with its supporting evidence.
7. Improve: The outcome, including any correction or exception, feeds Fiskl's evaluation processes.

This shows how financial understanding, AI decisioning, controlled execution and verification work within one accounting platform.

One Intelligence Architecture, Three Ways to Access It

The same financial intelligence and accounting infrastructure supports three ways to work with Fiskl.

The Fiskl App

The Fiskl app is the complete accounting and financial management application. AI runs throughout accounting workflows, with automatic categorisation, matching, reconciliation, document processing and confidence-based review.

Fi, Fiskl's AI CFO

Fi is the conversational and agentic interface, and it works in more than 200 languages. Fi analyses financial information, explains results, recommends actions and coordinates permitted workflows using Fiskl's accounting capabilities.

External Platforms, API and MCP

Fiskl exposes supported financial capabilities to external applications and AI systems. Banks, fintechs and payment providers integrate accounting functionality through the API. You can also connect your preferred AI through MCP, with access governed by your Fiskl permissions.

See Connect Your Preferred AI and Webhooks.

Fi is one interface to Fiskl's intelligence. The specialised models, context retrieval, orchestration, controls and deterministic engine also run independently within the app and through supported integrations.

AI, Deterministic Logic and Financial Truth

Fiskl assigns each kind of problem to the component built for it.

Component	Primary responsibility
Specialised accounting AI	Predicting, classifying, matching, interpreting and detecting financial patterns
External language models	Broader reasoning, natural language understanding and complex interpretation
Financial context and RAG	Supplying relevant financial relationships, evidence and retrieved knowledge
Agent orchestration	Selecting models and tools and coordinating multi-step workflows
Policy and execution controls	Determining whether and how an action may proceed
Deterministic accounting engine	Calculating and validating accounting outcomes
Ledger	Maintaining the authoritative financial record and audit evidence
Recursive improvement system	Using verified outcomes to evaluate and improve future decision processes

Each component does the work it is built for, and together they form one accounting platform.

Frequently Asked Questions

Does Fiskl send all my transactions to external AI models?

No. Fiskl's specialised models make accounting decisions directly. An optional external language model adds reasoning where Fiskl's own models leave a decision unresolved. You can turn it off in Settings > AI & Automations.

Does every correction immediately retrain Fiskl's AI?

No. Your categorisation history and rule suggestions adapt to completed transactions straight away. Updates to Fiskl's production decision models are evaluated and approved before release.

Is Fiskl's financial context a separate ledger?

No. The ledger remains the authoritative financial record. The financial context layer assembles the relevant entities, relationships and evidence into a structured view for retrieval and AI reasoning.

Can Fi or a connected AI change a locked accounting period?

No. Fiskl enforces period locks and user permissions. An authorised person must unlock the period before changes can be recorded.

Are external language models part of Fiskl's proprietary models?

No. External language models are interchangeable reasoning components. Fiskl's decision models, deterministic accounting logic, financial records and execution controls stay separate from any external provider.

How does Fiskl verify AI decisions?

Verification depends on the task. Fiskl validates journal entries, confirms payment matches, checks tax and currency rules, reviews resulting balances and compares predictions with confirmed outcomes.

Related Topics

• Fiskl AI Overview: What Fi and Fiskl AI can do
• Fi, Fiskl's AI CFO: Ask questions and run analysis on live data
• Fi Close Agent: Run the month-end close with approvals
• How Auto-Categorisation Works: Confidence bands and the learning cycle
• Tax Intelligence and Nexus: Tax rules by jurisdiction, with sources
• Audit Log: See who changed what, and when
• Connect Your Preferred AI: Use your preferred AI with Fiskl through MCP
• Webhooks: Send Fiskl events to your own systems