AI Agents in Accounting: What's Real, What's Hype, and What SMEs Should Do Now

Goldman Sachs Has AI Doing Its Books. Should You?

In early 2026, Goldman Sachs deployed AI agents — built on Anthropic’s Claude — to handle accounting and compliance tasks that previously required teams of analysts. The agents don’t just generate reports. They reconcile transactions, flag anomalies, prepare regulatory filings, and escalate edge cases to humans. According to internal reports, some processes that took days now take minutes.

At the same time, a survey by Gartner found that 82% of midsize companies are actively exploring agentic AI for finance operations. Not chatbots. Not dashboards with “AI-powered insights.” Actual agents that execute tasks autonomously, with human oversight.

This isn’t the usual AI hype cycle. Something genuinely different is happening in accounting — and it matters for SMEs, not just banks.

This article breaks down what AI agents actually are, how they differ from the “AI features” in your current accounting software, what’s real today, and what you should be doing about it.

What AI Agents Actually Are (And Aren’t)

Let’s be precise, because the term “AI” gets slapped on everything from auto-categorisation to chatbots to tax calculators.

Traditional accounting software

QuickBooks, Xero, Sage, DATEV — these are tools. They give you forms, menus, dashboards, and reports. You enter data. You press buttons. You make decisions. The software executes what you tell it to. Some newer features auto-categorise bank transactions or suggest matches, but the human drives the process.

”AI-powered” features

Many accounting tools now advertise AI. Usually this means:

• OCR (Optical Character Recognition): scans a receipt or invoice and extracts text into fields
• Auto-categorisation: suggests expense categories based on merchant name or description
• Anomaly detection: flags transactions that look unusual compared to your history
• Chatbot: answers questions about your data (“What was my revenue last quarter?”)

These are useful. They save time. But they’re features within a tool. You still initiate every action. The software doesn’t decide to reconcile your bank feed at 2am because it noticed unmatched transactions accumulating.

AI agents

An AI agent is fundamentally different. It has:

1. Goals: not just functions. An agent is given an objective (“keep the books current”, “prepare Q2 VAT filing”, “process incoming invoices daily”) and determines how to achieve it
2. Autonomy: it can take actions without being asked. When a bank transaction comes in, the agent matches it. When an invoice PDF arrives, the agent extracts the data, creates the journal entry, and flags anything unusual
3. Judgement: it can handle ambiguity. “This invoice has a line item that could be classified as either office supplies or marketing materials” — an agent evaluates context and makes a call, or asks a human when it’s genuinely uncertain
4. Memory: it learns from corrections. If you reclassify a transaction, the agent adjusts its future behaviour
5. Multi-step execution: it can chain actions. Receive invoice → extract data → check for duplicates → verify supplier → create journal entry → match to purchase order → schedule payment

The difference isn’t speed. It’s agency. Traditional software waits for you. An agent works alongside you.

What AI Agents Can Actually Do for SME Accounting Today

Let’s separate the real from the aspirational. As of March 2026, here’s what AI agents can genuinely handle for small and medium businesses:

Proven and working

Invoice processing (received): An agent receives a PDF invoice (by email, upload, or OCR), extracts all structured data (supplier, amounts, VAT, line items), cross-references against known suppliers and purchase orders, checks for duplicates, creates the journal entry with correct account codes, and presents it for human confirmation. The human reviews and approves — the agent does the 15 minutes of work that previously sat in someone’s inbox for days.

Bank reconciliation: An agent monitors bank feeds, matches incoming payments to outstanding invoices, categorises expenses based on learned patterns and merchant data, and flags anything it can’t confidently match. What used to be a monthly chore becomes a continuous background process.

Invoice generation (issued): “Create an invoice for Müller GmbH for the March maintenance contract, €5,000 net.” The agent creates the invoice, applies the correct VAT rate based on the transaction type and jurisdiction, generates the journal entry, and produces the document in the required format (PDF, Factur-X, XRechnung, CFDI — depending on the country).

Compliance calendar management: An agent tracks filing deadlines across jurisdictions (VAT returns, corporate tax, payroll submissions), prepares the required data, and alerts you before each deadline. In multi-country setups, this alone saves hours of coordination.

Expense categorisation and anomaly detection: Continuous monitoring of transactions against historical patterns. New vendor? Unusual amount? Duplicate payment? The agent flags it in real time, not during a quarterly review.

Working but requires human oversight

Month-end close: An agent can prepare accruals, reconcile sub-ledgers, generate trial balance, and draft financial statements. But the judgement calls — provisions, impairments, unusual items — still need human review. The realistic improvement: month-end close goes from 12 working days to 2-3 days, with the agent handling 80% of the mechanical work.

Tax preparation: An agent can aggregate data, calculate liabilities, and pre-fill returns (VAT, income tax, payroll taxes). But tax filings have legal consequences. Most SMEs (and their accountants) want a human to review before submission. The agent makes the review faster, not unnecessary.

Payroll processing: Calculate gross-to-net, apply tax tables, generate payslips, prepare filings. Agents handle the computation reliably. But edge cases (new hires mid-period, terminations, retroactive adjustments, statutory changes) benefit from human verification.

Not yet reliable for SMEs

Strategic financial advice: “Should I lease or buy this equipment?” “Is this the right time to expand into Germany?” AI agents can provide data to inform these decisions, but the judgement is still yours. Anyone telling you their AI can make strategic business decisions for you is selling something.

Audit-grade accuracy without review: Regulators and auditors still expect human accountability. An agent can prepare everything, but a human must sign off. This won’t change soon.

Complex multi-entity consolidation: Intercompany eliminations, currency translation, minority interests — agents can handle the mechanical parts, but the accounting policy decisions require expertise.

The “Human + Agent” Model

The most effective approach isn’t full autonomy. It’s what practitioners call the “Human + Agent” model:

• The agent handles volume: routine transactions, data entry, matching, categorisation, deadline tracking. The 80% of accounting work that’s mechanical and repetitive.
• The human handles judgement: unusual transactions, policy decisions, client relationships, strategic advice, regulatory interpretation. The 20% where expertise matters.

This isn’t a compromise. It’s the optimal design. Goldman Sachs doesn’t let its AI agents file regulatory returns unsupervised — they prepare everything, and a human reviews and submits. The same model works for a 10-person company.

The practical result: 4-8 hours recovered per week for a typical SME owner or bookkeeper. Not because the agent is faster at any single task, but because it handles dozens of small tasks that would otherwise accumulate and wait for a human to get around to them.

How to Evaluate AI Agents in Accounting Software

If you’re shopping for accounting software or an ERP in 2026, here’s how to tell real AI agents from marketing AI:

Signs of a real agent

• It takes actions without being asked: processes invoices as they arrive, reconciles bank feeds continuously, sends deadline alerts automatically
• It chains multiple steps: receive invoice → extract data → journal entry → payment scheduling, without you clicking through each step
• It explains its reasoning: “I classified this as marketing expense because the supplier Acme Design was categorised as marketing in the last 12 transactions”
• It improves from corrections: reclassify a transaction once, and it adjusts future categorisation
• It knows when to ask: flags uncertain items for human review instead of guessing silently

Signs of marketing AI

• “AI-powered dashboard”: it’s a chart. Charts aren’t agents
• “AI insights”: a summary of data you can already see in the reports
• “Smart suggestions”: auto-complete in a search bar
• Requires you to click a button to trigger it every time: that’s automation, not agency
• Can’t explain why it did something: black box categorisation with no audit trail

Questions to ask the vendor

1. Can the agent process an incoming invoice end-to-end without me opening the app?
2. Does it reconcile bank transactions automatically or only when I press “reconcile”?
3. If I correct a categorisation, does it learn for next time?
4. What does it do when it’s uncertain? Does it guess or does it ask?
5. Can I see an audit trail of every action the agent took and why?

How Odiverse Approaches AI Agents

Odiverse was built around a specific philosophy: GPU where you need flexibility, CPU where you need accuracy.

The AI agent (Odi) handles interpretation, conversation, categorisation, anomaly detection — tasks where judgement and natural language understanding matter. The deterministic engine handles tax calculation, journal entries, hash chains, regulatory formatting — tasks where precision is non-negotiable and a 99.5% accuracy rate isn’t good enough.

Odi processes invoices: upload a PDF or just tell Odi “invoice Acme Corp for €3,000 for March consulting.” Odi handles the rest — correct VAT rate, journal entry, document generation in the right format for the right country.

Odi reconciles continuously: bank feeds are matched in real time. Odi categorises, matches payments to invoices, and flags discrepancies. You review a clean list, not a raw bank statement.

Odi tracks deadlines: VAT returns, payroll filings, corporate tax — Odi knows what’s due, when, and in which country. It prepares the data and alerts you before the deadline.

Odi speaks your language: Spanish, English, Portuguese, French, German, Italian. Not through translation — through context. It understands that “hazme factura” and “create an invoice” and “erstell eine Rechnung” all mean the same thing, with different fiscal implications depending on the jurisdiction.

Odi doesn’t guess on compliance: tax rates, hash chains (VeriFactu), e-invoice formats (Factur-X, XRechnung, CFDI), filing calculations — all handled by deterministic code, not the AI model. The AI decides what to do. The code ensures it’s done correctly.

What SMEs Should Do Now

If you’re already using accounting software

1. Audit your time: Track how many hours per week you spend on routine bookkeeping (data entry, categorisation, reconciliation, filing prep). If it’s more than 4 hours, an AI-powered system will pay for itself
2. Check your current tool’s AI capabilities: Does it actually take actions, or does it just suggest? There’s a meaningful difference
3. Don’t wait for your current vendor to “add AI”: Retrofitting agents onto legacy architecture rarely works well. Purpose-built systems outperform bolt-ons

If you’re choosing software for the first time

1. Start with agent-first, not features-first: The question isn’t “does it have a VAT calculator?” The question is “will it handle my VAT filing end-to-end with minimal input from me?”
2. Prioritize compliance automation: With e-invoicing mandates hitting across Europe, Mexico, and Brazil, compliance that “just works” is worth more than a prettier dashboard
3. Think multi-country from day one: If there’s any chance you’ll operate internationally, choose a system that already handles multiple jurisdictions natively

If you’re an accountant serving SME clients

1. AI agents aren’t replacing you: They’re replacing the manual data entry and reconciliation work that you’d rather not be doing. Your value is in judgement, advice, and client relationships
2. Clients using AI-powered ERPs will need different services: Less data entry, more advisory. Prepare for that shift
3. Learn to work alongside agents: Review the agent’s work, refine its categorisation, set rules — this is the new workflow

The Bottom Line

AI agents in accounting aren’t science fiction and they aren’t marketing fluff. They’re working today — at Goldman Sachs and at 10-person companies. The technology is real. The question isn’t whether to adopt it, but when.

For SMEs, the calculus is straightforward: if you spend more than 4 hours a week on routine bookkeeping and compliance, an AI-powered ERP will give you that time back. Not by making the work faster, but by doing it for you — with you reviewing the output instead of producing it.

The 82% of midsize companies exploring agentic AI aren’t wrong. They’re just early. In 12 months, this will be the default expectation.

See how Odi works — or read about how AI ERPs compare to traditional software and why AI agents are different from accounting tools.