Starting, testing, and adjusting

Which administrative processes should you automate first?

Start with a narrow, recurring process with fixed rules, quick review, and recoverable errors: usually a shared inbox or purchase invoices.

Piet Baudoin··14 min read

Automate a frequent administrative process with fixed rules, results reviewable in seconds, and mistakes recoverable without harm. For small and midsize businesses, this usually means the shared inbox or purchase invoices, starting with one narrow task within it.

This page gives six scoring criteria, an eight-process example, two exclusion rules absent from other models, and reasons to start narrowly. We compare six existing models, recalculate examples, and identify gaps. When an AI agent makes sense addresses whether to start at all; here we select one task afterward. Dutch adoption appears in AI workflow automation for Dutch small and midsize businesses.

The short answer

  • Score six criteria 1 to 5: volume, regularity, fixed rules, recoverability, review time, measurability. At least 22/30 is strong; below 15, repair the process (Forge RPA, March 11, 2026).
  • Recoverability or review time scoring 1 or 2 disqualifies a first task regardless of total.
  • Our twenty-person example gives purchase invoices and shared inboxes 27/30; contracts 9.
  • A vendor gives 2-to-3-minute manual documents versus 10-to-15-second AI. Yet 1,800 monthly documents yield about 30 hours after review rather than the calculated 69 (Crux Digits): review and exceptions explain the difference.
  • Gartner predicts over 40 percent of agentic projects canceled by end-2027 for costs, unclear value, or inadequate risk control (Digital Applied on Gartner, 2026). Narrow measurable tasks are the antidote.
  • 74.6 percent of Dutch businesses considering but not implementing AI cite inexperience (CBS, September 2025). First tasks chiefly build experience.
  • Mandatory B2B e-invoicing is approaching, between 2030 and 2032 according to peppol.nu, June 26, 2026. PDF retyping has an expiration date; inboxes do not.

Which criteria determine the first process to automate?

Frequency, regularity, fixed rules, mistake severity, review speed, and measurability determine selection. Most models contain the first three and omit the last three, which determine first-task success.

  1. Volume. Weekly frequency. Less than weekly offers little. TaskMesh, January 27, 2026 sets “more than once a week” (translated).
  2. Regularity. Same procedure and unchanged for a year? Quickomate, June 4, 2026 requires 12 months’ stability.
  3. Fixed rules. Written agreements versus judgment. Forge RPA: “Automations excel at following rules and fail at judgment.”
  4. Recoverability. Misrouted internal email takes a minute to repair. A demand sent to a paid-up customer does not.
  5. Review time. Human proposal-check time remains a daily cost.
  6. Measurability. Counts, elapsed time, corrections before/after. Without baseline, three months cannot show success.

Crux Digits says, translated, to automate high-volume, regular, easy-to-check work before judgment tasks. AFAS, Dutch business software, says start with simple processes and few exceptions (AFAS).

How do you score a process for automation?

Give each criterion 1 to 5, higher meaning more suitable, totaling at most 30. Forge RPA says: “Processes scoring 22+ are strong automation candidates. Those scoring 15-21 are worth investigating. Below 15, fix the process first.” 22/30 is about 73%; half the maximum only reaches investigation.

Two first-task exclusions:

  • Recoverability 1 or 2: not first. Customer/payroll errors in learning cost more trust than the task returns.
  • Review time 1 or 2: not first. If review nearly equals manual work, little remains.

Count each occurrence for a week. Have handlers record unusual features in ten random cases for exception rate and rules. Ask who notices errors first, you or customers. Time a coworker reviewing another’s work. This takes an afternoon.

Which administrative processes score highest?

Our twenty-person illustration gives invoices and shared-inbox distribution 27/30. We filled it to demonstrate the model, rather than measure a client. An installer with little purchasing scores invoices differently than a wholesaler.

Process Volume Regularity Rules Recovery Review Measurable /30 Judgment
Purchase-invoice retyping 5 5 4 4 4 5 27 Strong: check per line
Shared-inbox distribution 5 5 3 5 5 4 27 Strong: internal misrouting repaired quickly
Weekly timesheets 4 5 4 3 4 5 25 Strong, but errors affect payroll
Email to files 4 5 3 5 4 3 24 Strong; search-time effects harder to measure
Standard customer answers 4 4 3 2 3 4 20 Excluded: external messages
Payment demands 3 4 4 1 3 4 19 Excluded: customer relationship
VAT-return preparation 2 3 4 2 2 4 17 Excluded: low volume, serious consequences
Contracts 1 2 2 1 1 2 9 Below 15: judgment

Scores 1 to 5, higher better. Forge RPA thresholds 22/15. Bombos example, rather than measurement.

Standard customer questions score 20, nominally investigate, but recoverability 2 excludes them first. Crux Digits similarly places VAT checks under review and tax planning/final advice under assistance only. Customer questions can be second or third once approval/correction is familiar. See which administrative tasks AI can take over.

How do existing scoring models compare?

Four of six include error consequences, two do not. Only Crux names checkability; none asks review duration.

Model Criteria Scale/threshold Error consequences? Source
Forge RPA Rule complexity, volume, structure, stability, error impact, technology 1-5, max 30; 22+ strong, 15-21 investigate, below 15 repair Yes 11-03-2026
Activepieces Volume, complexity, business value Low 1, medium 3, high 5; weights ×3, ×-2, ×5 No 27-09-2026
Quickomate Volume, pain, structure, stability 1-5 summed No 04-06-2026
Kamyar Shah Benefit versus failure consequence 2×2 quadrants Yes, main axis Undated
JumpCloud Impact × probability × cost 1-125 Yes 01-07-2026
Crux Digits Volume, regularity, checkability Three layers Yes, layers 2/3 Undated
This model Volume, regularity, rules, recovery, review, measurement 1-5, max 30, two exclusions Yes, separate review time Bombos synthesis

All six sell automation, software, or advice. Useful, but recalculate. Activepieces uses “(Volume Score × 3) + (Complexity Score × -2) + (Business Value Score × 5)”. High/medium/high invoice processing gives 5 × 3 + 3 × -2 + 5 × 5 = 34, but its table says 28. Lead routing gives 28 versus listed 24. Order remains, figures do not. High and Very High both score 5, failing to distinguish ten invoices daily from a hundred.

Quickomate gives the strongest sentence: “Automate the process that’s bleeding the most hours into work no one needs to do, not the one that’s complaining the loudest.” Yet it also omits remaining review hours.

Why does review time matter more than promised savings?

Our position: select first tasks by review time and recoverability rather than greatest theoretical savings. Checks consume promised hours, absent from brochures.

Crux Digits gives 2 to 3 manual minutes versus 10 to 15 AI seconds, and about 30 reclaimed hours after validation for a fifteen-person firm processing 1,800 documents. Calculation: 1,800 × 2.5 minutes = 75 hours; × 12.5 seconds = over 6. Difference 69 hours versus claimed 30. Missing 39 hours, 57% of theoretical benefit, goes to validation; review/exception breakdown is absent.

Zenvoices says, translated, “Everything above, for example, 95 percent confidence can go through automatically. You want to review the rest.” Accountancy Vanmorgen, July 2, 2026 Return depends on the rest’s size and speed. AIDA warns automation creates speed rather than automatic certainty in the same journal. Both sell these services.

For twenty people, equal-volume processes differ. At 400 monthly items, 10-second reviews and five-minute exceptions: 5% exceptions cost almost 3 hours (67 + 100 minutes); 20% almost 8 (67 + 400). First-task learning requires checking everything. Fast review leaves capacity for task two.

Why must your first automated task be narrow?

Broad projects fail more often, while the first task mainly teaches the team. Narrow means one stream, document/message type, destination package: “regular suppliers’ purchase invoices into Exact, Dutch accounting software,” rather than “administration.”

Gartner predicted June 25, 2025: “over 40% of agentic AI projects will be canceled by the end of 2027”, citing costs, value, risk. Analyst Anushree Verma calls many “early stage experiments or proof of concepts that are mostly driven by hype” (Digital Applied, 2026; Gartner release inaccessible). Digital Applied sells advice and recommends the narrowest measurable agent first. It also reports about 75% of leaders engaging with agents, 17% deployed, 11% production-ready.

Among Dutch businesses with ten or more people considering but not adopting AI, 74.6% cite lack of experience, by far most important (CBS, September 2025). Microbusinesses: 71.6% (ICT Magazine on March 2026 CBS). Experience is acquired rather than purchased. Narrow tasks teach proposals and corrections within weeks. Our service-business AI-native blog says, translated, choose work someone is already paid for whose correctness you can check.

Purchase invoices or shared inbox: where do you start?

Choose the shared inbox when email work is distributed among people; invoices when purchasing volume is high and one person retypes. Equal example scores leave your situation decisive.

Four of nine reviewed pages put invoices first: aiagency.nl, March 20, 2026, Timmermans Media, June 6, 2026, TaskMesh, Forge RPA. VrijLeven and our guide start with intake/email. Invoices usually arrive by email: the same stream from different starting points.

Two omitted considerations:

  • Invoice, bank, VAT, customer communication ordering comes from accountancy, often accountancy-AI sellers. Installers, owners’ association managers, and wholesalers may start with inboxes, emailed orders, or resident reports.
  • PDF retyping has an expiration date. Peppol.nu, June 26, 2026 expects mandatory domestic B2B e-invoicing, draft legislation Q4 2026, obligation 2030 to 2032. Other sources differ; date is unsettled. Structured invoices remove retyping, leaving agreement discrepancies. Email remains.

Starting with invoices is reasonable, but value later shifts to review.

Which processes should you avoid automating first?

Avoid customer, payroll, or tax-authority-reaching errors, case-by-case judgment, and less-than-weekly work: demands, replies, VAT, contracts, annual work.

These can later be automated. Crux Digits allows direct posting, duplicates, bank matching; reviewed summaries, VAT checks, fraud flags; assistance only for tax planning, audit opinions, final advice. Kamyar Shah has a do-not-automate quadrant. Accounting firms also need professional review: June 2026 NBA/NOREA guidance requires reproducible, professional-validated AI figures according to Crux Digits.

Rare harmless tasks still make poor first choices. Activepieces gives password resets highest volume but lowest total because value is low.

Must you clean up a process before automating?

Yes if coworkers follow different procedures; no if only written descriptions are missing. Automating mess preserves it.

Zenvoices recommends standardize, centralize, automate in Accountancy Vanmorgen: office costs on ledger 4300 for one client and 4500 for another. Forge says below 15, “fix the process first.” Unwritten agreements are different and common in twenty-person firms; setup can uncover them.

Cleanup may reveal no AI need. PuraFact, September 14, 2026 separates fixed software, AI interpretation, and human decisions per step. Fixed accounting rules are cheaper. See AI or traditional software.

What does the first task deliver and how do you measure it?

Count weekly items, active minutes, elapsed time, and errors beforehand. Vendor figures vary widely and are self-reported.

Measure Manual AI Source
Per document 2 to 3 minutes 10 to 15 seconds Crux Digits, undated vendor
Per invoice cost €12 to €18 €3 to €5 Timmermans Media, June 6, 2026, vendor
Monthly hours, six-person firm, 90 clients 180 65 aiagency.nl, March 20, 2026, unnamed vendor case
Automatic-pass threshold N/A 95% confidence Zenvoices, July 2, 2026, vendor
Payback N/A 3 to 14 months TaskMesh, Retrii, aiagency.nl, all vendors

Payback differs by 14 / 3 = 4.7, mainly showing no independent measurement. TaskMesh calculates 5 weekly hours × €40 = €800 monthly using four weeks; 52 annually gives €867.

Collect four measures over a week and repeat after three months:

  1. Weekly items in the selected stream.
  2. Minutes per item, including package searches.
  3. Review time per proposal and correction share after starting.
  4. Elapsed time before processing.

Corrections matter most. A declining monthly share indicates learning; unchanged share suggests excessive scope or unrecorded rules. See preproduction tests and total costs.

Where is this heading?

Measured growth is high but absolute adoption low. Dutch businesses with ten or more people rose nearly 9 points to 22.7% in 2024; 10-to-19-person firms 17.8%, 500+ 59.2% (CBS, September 2025). Microbusinesses reached 13.8% in 2025; 25.9% of users apply AI to administration/management (ICT Magazine on CBS). Continued nine-point annual growth implies roughly four in ten by end-2026: crude extrapolation rather than measurement. See increasingly automated processes.

Article 50 AI Act transparency duties have applied since August 2, 2026. Regulation (EU) 2026/1744 postponed Annex III high-risk rules to December 2, 2027 (Niels Roest, after July 27, 2026). Narrow approved tasks do not need redesign for those dates. Mandatory e-invoicing eventually removes retyping.

Our expectation: by end-2027, most small and midsize businesses starting AI choose first tasks by review time and recoverability rather than largest promises. Broad projects will be substantially written off (Gartner over 40%); structured invoices remove easy typing, leaving exceptions and review costs. It fails if CBS shows broad projects succeeding without attrition or models make routine review unnecessary.

Bombos will make review time visible so future choices use figures rather than intuition.

What can this approach not do yet?

Scoring aids choice rather than predicting results:

  • Example scores are ours, not measurements. No published completed customer sheet. Forge’s 22/15 RPA thresholds are untested on Dutch twenty-person businesses.
  • Review and exceptions become known in operation. Estimates may be off by a factor two. Use month one for measurement and reversible work.
  • Unrecorded knowledge is invisible. Unwritten exceptions inflate apparent fixed-rule scores. Bombos interviews handlers to record them.

One task does not transform a business alone. Benefits continue through tasks two, three, and four your team chooses.

How does Bombos approach this?

Start with a narrow task you choose, usually inbox or invoices, teaching your team AI agents. The broader goal is more work at higher quality without headcount growth.

Chef reads connected addresses, recognizes tasks, and assigns specialists. They find client/files in Exact (Dutch accounting software), AFAS, Twinfield (accounting software), Syntess (installation-business software), or other packages, then prepare sourced proposals. You approve, change, or reject in Bombos before results enter packages. Corrections become coworker rules. Payments, messages, contracts wait by default. Bombos can disable a work-type boundary at your request and risk.

Review becomes visible through correction counts and decline. Wegwijzer then suggests next tasks your team teaches without technical skills. We guide the first; then you can do it yourself. See starting with an AI pilot.

Sources

Each source was opened September 30, 2026; quotations appear literally in it, with Dutch quotations translated above.

Free, no obligation

More work done, at a higher quality, with the same team.

That is what Bombos is for: companies that grow fast and want to keep the same team. We start with one task that keeps piling up and guide you until your team can handle it. Then your team teaches Bombos the next task. Leave your number and we will call you back to talk about your situation.

We read what you write. Within one working day you hear from the one of us who knows your kind of work best.

More about starting, testing, and adjusting

All questions in the knowledge base →