Divide the steps: rules and traditional software for fixed structure and zero error tolerance, such as posting, calculation, and payment; AI for variable input, such as unfamiliar PDF layouts, natural-language email, and exceptions; and human approval between them.
This page draws that boundary: existing accounting features, strengths, invoice costs with our calculations, reliability, retention, and the AI Act. See agents versus workflow automation for definitions.
The short answer
- Fully describable if-then steps are cheaper, faster, and more predictable with rules.
- Ortem quotes $0.001-$0.01 per rule execution and $0.05-$0.50 per agent execution. Our model-based invoice-reading calculation is about one cent.
- Thinking Machines received 80 different answers to 1,000 identical temperature-zero requests. Use rules for arithmetic, VAT, and payments.
- AI helps when templates break: varied invoices, ordinary-language email, new suppliers without posting rules.
- Moneybird, online accounting software, matches bank transactions only at 100% certainty. Exact, Dutch accounting software, claims 98% Purchase Agent accuracy and includes automatic invoice processing.
- Review dominates AI cost: 10-15 seconds per document costs over ten times the model.
- Seven-year administration retention and AI Act accountability require recording proposals and approvals.
What does your accounting software already do?
Start with existing features. Three major Dutch packages do more than many owners realize, partly through AI.
Moneybird automatically processes bank transactions only when, in translation, “we are 100% certain the link is correct” (help center, read September 30, 2026). Otherwise you choose. Invoice recognition is “not always 100% perfect”; it asks you to check recognized data. It recommends UBL invoices because fields have fixed structured positions (help center, translated).
Exact’s Purchase Agent detects deviations and proposes ledger accounts with claimed 98% accuracy, without a measurement method (read September 30). Automatic processing is included in Exact Online Accounting editions and available across its solutions. It promises up to seven-times-faster invoice and payment processing (Scan & Recognize).
| Package | Standard behavior | External AI approach | Source |
|---|---|---|---|
| Moneybird | Certain bank matching; invoice recognition with review advice | Open MCP connection since Sept. 16, 2025, read-only or read/add | Help center and blog |
| Exact Online | Included recognition; seven agents: Purchase, Bank, Debtor, Financial, Support, CRM and Excel in beta | Switchable; Financial added Sept. 3, 2026 | Exact; ICT Magazine |
| AFAS, Dutch business software | Jonas for preconfigured work | No certified external AI connection in June 2026; uncertified access called unchecked and risky | AFAS customer portal |
Invoice reading is already package work. Additional purchases must do more. See built-in AI or a coworker alongside and AI with existing software.
When is traditional software better?
Rules suit stable structure, high volume, and zero error tolerance. DIDEV advises, in translation: “If you can fully describe the task in if-then statements, choose traditional automation” (undated). Tax Administration bank descriptions go to VAT accounts; supplier X invoices to account Y. No understanding is required.
Make says “RPA works best when your source system, target system, and business rules remain stable” and “It is built for stability, not flexibility. It does not make judgments. It follows instructions exactly” (April 20, 2026). Zapier recommends RPA for predictable high-volume work requiring absolute consistency, especially legacy software without APIs (January, updated July 2026). Anthropic recommends the simplest solution and complexity only as needed (December 19, 2024). All sell AI or automation. RPA repeats fixed screen actions; see orchestration versus RPA.
Determinism is the strongest reason. Rules return identical results for identical input. Models do not, even at temperature 0. Thinking Machines got 80 answers from 1,000 identical requests because hosted-model “the load (and thus batch-size) nondeterministically varies” (2025). Modified kernels made all identical, but took 42 rather than 26 seconds. Hosted customers cannot choose that. Use code for VAT, sums, and duplicates.
When is AI better?
AI suits variable input requiring meaning. Make says: “If it depends on structure, RPA works. If it depends on meaning, you need agentic AI” (April 20, 2026).
Three areas: variable documents, language, and exceptions. Vellum claims OCR reaches 99% on standardized documents but needs “template creation and rule definition” per type (December 3, 2025). One hundred suppliers mean a hundred templates vulnerable to redesigns. Models read like coworkers. “Same as last time, but to the new address” needs context. New suppliers lack rules; amounts can differ from orders.
Vellum recommends OCR for standard forms, models for receipts/contracts, and hybrids for invoices, statements, and resumes. Even invoices require a combination.
Which tool fits each step?
An invoice process is a chain, not one step. This is our synthesis.
| Step | Characteristic | Tool | Reason |
|---|---|---|---|
| Tax Administration bank description to VAT account | Fixed text/account | Package rule | Moneybird requires certainty |
| UBL or Peppol invoice | Structured | Traditional software | Fields already fixed |
| Arbitrarily laid-out PDF | Variable | AI, possibly built-in, then rules | Supplier templates break |
| Known supplier’s ledger account | Repetition | Supplier X → account Y rule | Explainable, nearly free |
| New supplier’s account | Meaning needed | AI proposal, human approval | No rule yet |
| “Same as last time” email | Language and history | AI | Meaning rather than structure |
| VAT, sums, duplicates | Arithmetic | Code | Models are nondeterministic |
| Payments and customer messages | Irreversible | Rule plus human approval | Zero tolerance, human responsibility |
| Amount differs from order | Judgment | AI flags, person decides | Rules detect difference rather than reason |
Corrections can become rules: today’s new supplier is next month’s known supplier. Money/customer final steps never depend on the model alone. See first processes.
Why is “AI or traditional software” the wrong question?
The boundary lies between variable-input steps, reading, understanding, searching, and output steps allowing no errors, posting, calculation, payment, sending. Use AI for interpretation, rules for execution, and people at the boundary.
Three facts align: model invoice reading costs about one cent; identical requests vary, 80 answers in Thinking Machines’ 1,000; Exact already includes reading. The expensive mistake is adding AI review where rules suffice. Conversely, AI can remove hundred-supplier template maintenance.
The hardest decision is which steps can proceed without people.
What does AI cost per invoice?
AI computation costs more than rules, but building, testing, and review dominate.
| Provider | Plan | Monthly price | Included |
|---|---|---|---|
| Zapier | Professional | $19.99 annually paid | 750 tasks |
| Make | Core | $9 | 10,000 credits |
| Make | Pro | $16 | 10,000 credits, extra features |
| n8n | Starter | €20 annually paid | 2,500 executions |
| n8n | Pro | €50 | 10,000 executions |
Sources, September 30, 2026: Zapier, Make, n8n.
Zapier counts successful steps and programmatic calls. Five steps across 500 monthly invoices mean 2,500 tasks, above entry’s 750, requiring a higher tier. n8n counts full workflows. Billing structure matters more than subscription headline.
Our calculation uses Anthropic prices, September 30, assuming 2,000 input and 500 output tokens per invoice. A token is a text fragment roughly a syllable.
| Model | Input per million | Output per million | Per invoice | 500 monthly invoices |
|---|---|---|---|---|
| Sonnet 5.5 | $2 | $10 | $0.009 | $4.50 |
| Haiku 4.5 | $1 | $5 | $0.0045 | $2.25 |
Seller Autoboeker quotes 2-3 manual minutes versus 10-15 AI-supported seconds (February 14, 2026). At assumed €40 hourly, retyping costs €1.33-€2.00 each. AI is about half a percent of that. Reviewing for 10-15 seconds still costs €0.11-€0.17, over ten times model costs.
Ortem quotes rules $0.001-$0.01 and agents $0.05-$0.50, claiming “50-500x more per transaction” (2026). Its ranges actually imply 5-500 times. Its $800-$2,500 monthly inference for 500 daily documents compares with our roughly $95: 500 × 21 working days × $0.009. The gap is overhead or rough estimation. Unreferenced RPA development $25,000-$50,000 versus agents $60,000-$120,000 concerns large companies. See SME automation costs.
How reliable is administrative AI?
Reliable enough to prepare, requiring a safety net before posting. Seller figures measure different things.
Exact’s 98% account proposals and Moneybird’s imperfect extraction can coexist. At 500 invoices, 98% leaves about ten wrong proposals, unidentified beforehand.
CRMArena-Pro’s best older agents achieved about 58% single-turn, 35% multiturn, but “over 83%” on fixed workflows (Salesforce arXiv, May 24, 2025). Older models still illustrate fixed-path advantages.
Rules stop on no match; models can provide Vellum’s “plausible but incorrect information.” UK’s FRC identifies hallucinations and says responsibility cannot be shifted to technology (Accountancy Vanmorgen, April 7, 2026, translated). Among 100 Dutch professionals, 44% cite accuracy/reliability/quality checks versus 33% cost (September 15; Silverfin-commissioned seller research). See errors and review.
How do you combine AI and rules?
Four layers:
- Rules first: UBL, known suppliers, fixed bank rules.
- AI for the remainder: sourced proposals from emails, invoices, or earlier entries.
- Code after AI: check totals, VAT, duplicates.
- Humans at irreversible actions: posting, payment, sending after approval. Corrections expand rules.
AI covers exceptions while rules constrain variability. See agents versus workflows.
What do retention and the AI Act change?
Both favor recorded combinations. The Dutch Tax Administration requires 7-year administration and 10-year property-data retention (read September 30). Retain posting reasons too.
The AI Act was amended May 2026. Annex III high-risk duties begin December 2, 2027; Annex I August 2, 2028. Article 50 transparency remained August 2, 2026 (Gibson Dunn, May). Lawyers assess posting proposals’ classification. Traceable system actions and decisions are easier with rules and approval logs than autonomous posting.
AFAS reported no certified AI connection in June, called uncertified MCP unchecked and risky, and advised restricted permissions (June 2026). Moneybird says users decide how to use their AI tool (September 16, 2025). See ChatGPT operating accounting software.
Where is this heading?
Software itself incorporates AI. CBS reports 27% AI use in businesses with 10 to 49 employees in 2025, versus 11% in 2023; 32% of users apply it to administration/management (December 2025). Eight annual percentage points gives roughly 43% in 2027, our linear estimate.
Exact has seven agents and launched Financial Agent September 3, 2026 for about 675,000 customers, describing financial-context reasoning rather than dashboards or chatbots (ICT Magazine, translated). Moneybird opened AI access; AFAS has Jonas. The government chose phased B2B Peppol e-invoicing in 2030-2032, with final legislation expected mid-2028 (Peppol.nu, June 26). Structured invoices need no reading AI.
| Date | Event |
|---|---|
| Sept. 16, 2025 | Moneybird opens accounting to chosen AI |
| June 2026 | AFAS reports no certified AI connections |
| June 26, 2026 | Netherlands chooses Peppol, rollout 2030-2032 |
| Aug. 2, 2026 | AI Act Article 50 transparency |
| Sept. 3, 2026 | Exact Financial Agent |
| Dec. 2, 2027 | Annex III high-risk obligations |
Our expectation: by late 2027, every major Dutch accounting package includes AI invoice and email reading, making the binary choice disappear. Remaining questions concern approval and cross-package work. Exact already includes reading, small-business adoption more than doubled, and Peppol reduces reading work from 2030.
This fails through major erroneous-posting incidents causing retreat, stricter financial-administration AI Act interpretation, or closed package AI encouraging external alternatives. Bombos aims to keep work flowing with your chosen approval boundaries.
What can administrative AI not yet do?
- Guarantee identical outputs through hosted models, even at temperature zero. Calculation needs code.
- Read error-free. 98% leaves ten plausible errors per 500 invoices. Review determines costs.
- Offer independent Dutch-package error measurements. Exact’s 98%, Vellum’s 99%, and Autoboeker’s times are seller claims.
- Prove payback. Pegamento gives 12-24 months (undated); others claim under six weeks. Neither cites sources.
- Take responsibility: it stays with you or your accountant, including under FRC guidance.
- Beat an already reliable rule on cost. AI adds expense and review.
Rules also break when suppliers redesign invoices or banks change descriptions, requiring maintenance.
How does Bombos approach this?
Bombos retains working steps and handles gaps between rules and software, targeting more work at higher quality with the same team.
Exact recognizes suppliers and amounts already. Missing job allocation requires emails, agreements, or receipts. Bombos prepares that investigation, as in invoices to the right project. Reliable package steps stay there.
Chef distributes incoming jobs. Wegwijzer explains and helps set boundaries. Specialists read emails/documents, find customers/files in Exact, AFAS, Moneybird, or other software, and prepare sourced proposals. You approve, change, or reject in Bombos. Results reach software afterward. Corrections become coworker rules, moving exceptions into next-time rules.
Payments, customer messages, and contracts wait by default. Bombos removes that boundary only at your request and risk. Interviews record undocumented knowledge such as unusual customer payment terms.
The first task begins the process. Your team teaches subsequent tasks without technical skills after our guidance.
Sources
Each source was opened September 30, 2026, and each original excerpt appears verbatim in it. Dutch excerpts above are translations.
Packages
- Moneybird help center, Automatic bank processing. Living document. Seller.
- Moneybird help center, Incoming-document recognition. Living document. Seller.
- Moneybird MCP connection. September 16, 2025. Seller.
- Exact, Artificial intelligence. Living document. Seller.
- Exact, Scan & Recognize. Living document. Seller.
- ICT Magazine, Exact launches SME Financial Agent. September 3, 2026.
- AFAS customer portal, AI integration. June 2026. Seller.
Rules, AI, and costs
- Make, RPA vs agentic AI. April 20, 2026. Seller.
- Zapier, Agentic AI vs RPA. January, updated July 2026. Seller.
- Anthropic, Building effective agents. December 19, 2024. Model seller.
- Thinking Machines, Defeating Nondeterminism in LLM Inference. 2025.
- Vellum, Document Data Extraction: LLMs vs OCRs. December 3, 2025. Seller.
- DIDEV, AI versus traditional automation. Undated. Seller.
- Autoboeker, AI-ready administration setup. February 14, 2026. Seller.
- Ortem Technologies, AI Agents vs Traditional Automation 2026. 2026. Seller.
- Pegamento, RPA versus AI. Undated. Seller.
- Zapier pricing, Make pricing, n8n pricing, Anthropic pricing. Living pages.
Reliability, adoption, and regulation
- Salesforce, CRMArena-Pro, arXiv 2505.18878. May 24, 2025. Agent seller.
- Accountancy Vanmorgen, UK regulator: accountant retains AI responsibility. April 7, 2026.
- Accountancy Vanmorgen, AI leads expected accounting changes. September 15, 2026. Commissioned by software seller Silverfin.
- CBS, AI most used for marketing or sales. December 2025.
- Tax Administration, Retaining administration. Living document.
- Gibson Dunn, EU AI Act Omnibus Agreement. May 2026.
- Peppol.nu, Dutch B2B e-invoicing toward Peppol 2030. June 26, 2026.
Our invoice-cost, review-time, Ortem-check, and CBS-extension calculations state their assumptions in the text.