ERP with Artificial Intelligence: What It Actually Means in 2026

In 2026, many ERP products describe some part of their workflow as AI-assisted. The phrase can cover very different designs and operational outcomes.

In one design, AI is added as a question-and-answer layer over existing data. It may retrieve a figure such as last month’s invoicing or prepare a summary.

In another design, AI tools are connected to specific workflows, with business context, permissions, and review points defined alongside each operation.

The practical distinction is where context, permissions, and human review sit in the workflow.

The problem with “adding AI” to an existing ERP

Imagine a building constructed in the 1990s that someone wants to retrofit with solar panels. The panels can work, but their integration depends on how the original electrical system and the new components are connected.

An AI layer added to an existing ERP may be limited to reading data or generating summaries, or it may be connected more deeply. The useful question is which workflows it can access, what context it receives, and where a person reviews the result.

An AI-native ERP builds its data architecture from the start around what the AI needs to be useful: what context each operation requires, how data is modelled so the AI can reason about it, how AI tools connect to the actual workflows of the business.

What function tools are and why they matter

Frihet documents 55+ function tools that its assistant can use on authorised workspace data. A function tool has a bounded purpose and follows the permissions of the underlying workflow.

A function tool is a concrete capability: retrieve recorded invoices, extract fields from a document, suggest an expense category, calculate a tax estimate, detect possible duplicates, or prepare a cash-flow forecast from available data.

Unlike a generic answer, a tool result can use current workspace context. Depending on the tool, it can return data or prepare a reviewable draft or action.

In practice, this shows up in capabilities like:

get_invoices, get_expenses, get_clients — An agent connected over MCP or the API can read the current state of your invoices, expenses and clients under the workspace permissions. From the chat inside the app, the copilot does not read that data.

create_invoice, update_expense — These tools can prepare a create or update action from a plain-language instruction, subject to permissions and review before confirmation.

calculate_tax_summary, get_vat_report — Tax estimates and reports based on the available workspace data.

detect_anomalies, predict_cashflow — Possible anomalies and general forecasts based on recorded data; the output is an estimate to review.

With function tools, the assistant can retrieve relevant records and prepare a result within the user’s permissions. The user checks the inputs, the proposed action, and any estimate before relying on it.

OCR: the difference between digitising and understanding

OCR (optical character recognition) has existed since the 1990s. What changed with AI is not that text can be read — that was already possible — but that the system can understand the structure of the document.

Traditional OCR reads: “INVOICE — Company XYZ — Net amount: 1,000.00 — VAT 20%: 200.00 — Total: 1,200.00.”

AI-assisted document extraction can map that text to invoice fields such as supplier, net amount, VAT, total, and description, then present those fields for review.

The operational difference is that extracted fields can arrive as a reviewable draft instead of starting from an empty form.

Frihet extracts from receipts and invoices fields such as supplier, date, issuer tax ID, description, net amount, VAT rate, tax amount, total, and payment method. It also suggests an accounting category.

The extracted fields and suggested category remain a draft. Compare them with the original document and confirm or correct them before recording the expense.

Suggested categorisation: context with human review

Expense categorisation has a complexity that is invisible until you do it manually: the same supplier can invoice you for things that belong in different categories. Amazon can be “office supplies” on one invoice and “IT equipment” on the next. The petrol station can be “transport” or “client entertainment” depending on context.

Frihet uses the current document and the available expense context to suggest an accounting category. The same supplier can still require a different category from one invoice to the next, so the proposal is not treated as a final entry.

Review the supplier, description, amounts, taxes, and suggested category together. If the context is ambiguous, choose the category that fits the actual purchase and your accounting policy.

Frihet presents categorisations as suggestions for review. The system proposes; the user approves or corrects. The final judgment remains with the user or responsible professional.

Reactive fiscal hints and tax estimates

This is the feature where the gap between added AI and native AI becomes most visible.

A chatbot can tell you “VAT in the UK is 20% of the net amount.” That is generic information you could find on a government website.

Frihet’s fiscal hints are reactive. In the invoice workflow, they load from the current invoice and fiscal context and refresh when relevant fields, such as the buyer’s country, change. They can surface guidance about reverse charge, e-invoicing, or SEPA-related cases when that context applies.

Separately, the fiscal dashboard can provide general VAT or income-tax estimates and cash-flow context from recorded data. These are estimates, not a filed return or a determination of which deductions a business is entitled to.

Review each hint and estimate against the underlying documents and applicable rules. A qualified professional remains responsible for interpretation, filings, and decisions with legal or tax consequences.

General forecasting: an estimate, not a fixed scenario

Cash-flow estimates can help a small business inspect recorded trends and pending invoices before making a decision.

Frihet’s revenue forecast uses available history and pending invoice data to provide a general forward-looking view. Its usefulness depends on the completeness and quality of the recorded data.

The forecast does not promise a fixed horizon or named client-by-client scenarios. Treat it as an estimate, review its assumptions, and combine it with known commitments that may not yet be recorded.

The user decides whether to defer an expense, follow up on a payment, or take another action; the projection does not make that decision.

The conversational assistant: from question to action

Frihet’s conversational assistant can use documented function tools for user-initiated questions and actions within workspace permissions.

The practical difference:

Chatbot: “How much did I invoice in January?” → Response: “€1,340”

Assistant with function tools: “Prepare the January summary for my accountant” → Retrieves the relevant recorded data and prepares a reviewable summary or export. The user checks it before saving, exporting, or sharing it.

Frihet’s assistant has access to 55+ function tools. The assistant selects a relevant tool for the request, while permissions and human review remain part of the workflow.

It works in 17 languages. A Japanese user, a German user, and a Spanish user all use the same assistant in their native language, with the same operational capabilities.

The autonomous Agent Loop remains beta, draft-only, and not customer-visible. It should not be presented as an autonomous production workflow.

The MCP server: when the AI already has your ERP integrated

For developers and technical users, Frihet publishes an MIT-licensed MCP (Model Context Protocol) server with 157 available tools.

MCP is a protocol that lets compatible AI assistants connect with external systems. Frihet’s MCP tools expose documented operations; access controls, workspace permissions, and review still apply.

The tools and their inputs are already defined, so a developer can compose a workflow using the documented operations and permissions.

The MCP server is MIT-licensed and open source. Check the current product documentation for access, authentication, and workspace permissions.

What the AI in Frihet does not do

Honesty is part of the design. There are things Frihet’s AI deliberately does not do:

It does not make decisions for you. It can project cash flow, but the decision to defer an expense or accelerate a collection is yours. It can detect an anomaly, but what to do about it is yours.

It does not replace a tax adviser. It can automate data capture and calculate estimates. Strategic tax planning, interpretation of complex regulations, and decisions with legal implications require a professional who takes responsibility for them.

Document quality matters. Illegible or ambiguous uploads can reduce extraction and categorisation quality, so compare suggestions with the source document.

It does not commit data without human review. Categorisation suggestions and AI-generated drafts are marked as such. The user reviews and approves before anything is committed to the accounts.

These are not embarrassing limitations. They are design principles. An ERP that automates critical decisions without human oversight is not a good ERP — it is a problem generator with a nice interface.

How to evaluate the AI in an ERP before you buy

If you are evaluating an ERP that claims AI, these are the questions that separate real AI from marketing AI:

Does the AI access my actual data, or does it just give generic answers? Ask about a specific case from your business. If the response is generic, the AI does not have context from your data.

Which function tools does the system expose? Ask which records each tool can access, which outputs are estimates or drafts, and what review is required.

Can the AI prepare an action, or can it only inform? A reviewable invoice draft from a plain-language instruction is different from an explanation of how to create one.

Is there human oversight in the workflow? A system that automatically commits results without review is a system that can make errors in your accounts without you knowing.

Which data controls are documented for the deployed service? Review the current privacy notice, data-processing agreement, and subprocessor list for purposes, legal basis, retention, transfers, and deletion. Do not infer a model-training policy, local processing, EU-only hosting, or regulatory compliance unless the applicable documents and contract state it explicitly.

For Frihet, the product manifest documents server-side workspace access controls and human review for drafts and suggestions. It does not establish a no-training policy, local processing, EU-only hosting, or GDPR compliance. Verify those points in the current legal and contractual documents, and ask support to demonstrate the relevant workflow and permissions before you commit.