When does an AI agent make sense for your business?

An AI agent makes sense when one recurring task takes more than about five hours weekly after subtracting review, work grows faster than your team, and you can catch mistakes before they leave. Below two hours weekly, it is practice rather than an investment.

This page makes “makes sense” measurable. Five signals (volume, repetition, growth, shortages, mistakes) each have a recognizable number. Then come calculations with real prices, unsuitable cases, a seven-question self-test, and what CBS and Gartner figures suggest for the next two years. Calculations use our stated assumptions, rather than customer results.

The short answer

• Below two hours saved weekly, calculating is not worth it. Above twelve months’ payback, the answer is “not yet.” Steding, August 10, 2026
• Payback below six months is strong for a bounded agent. Above a year, choose another task. BrightBox, June 25, 2026
• For dozens of weekly cases, manual work is often more efficient than configuring an agent. bConnect, accessed September 30, 2026
• Small businesses (5 to 50 employees) use automation against staff shortages half as often as large ones: 20.1 versus 40.4 percent. Nederland Digitaal on CBS, June 3, 2026
• The labor market is less tight than in 2022: 95 vacancies per 100 unemployed in Q2 2026, versus 142 at the peak. CBS, July 30, 2026
• 29.8 percent of Dutch small and midsize businesses with 10 to 249 working persons used AI in 2025, versus 66.2 percent of large businesses. CBS, March 16, 2026
• Gartner expects over 40 percent of agentic AI projects discontinued by the end of 2027, and counted only around 130 genuine vendors among thousands claiming it. RCR Wireless on Gartner, June 27, 2025
• No source we found independently measures time agents save small or midsize businesses. All examples come from agent vendors.

What is an AI agent, and how does it differ from an AI assistant?

An agent performs work itself end to end; an assistant helps while you do the work. PROMPTED summarizes, translated: “An AI assistant speeds up your work. An AI agent does work in your place.” (PROMPTED, June 5, 2026). An assistant waits for your question. An agent notices an incoming invoice, finds the project, and prepares an entry.

An assistant can be useful when one employee occasionally needs writing help. An agent requires a stream arriving and being processed repeatedly in the same way. The actual question is whether you have that stream and whether it is large enough.

Compare fixed software rules in AI agents versus traditional workflow automation. Autonomous working time appears in What is an AI agent, and how long does it work alone?.

Which five signals make an AI agent worthwhile?

Volume, repetition, growth, shortages, and interceptable mistakes make an agent worthwhile. Vendors discuss them separately; we compare them measurably. Volume and repetition determine capability. Growth and shortages determine demand. The mistake question determines how quickly you can start.

Third-column thresholds are our interpretation, rather than law. Use them to assess your task.

How much time must a task take before an AI agent pays off?

A task must take two to five net hours weekly, after review, before calculating. Above five, an agent is worthwhile for most small and midsize businesses. Steding supplies the lower limit, translated: “Below two hours of saved work per week, you do not even need to do the calculation.” (Steding, August 10, 2026). The upper limit follows the calculation below.

“Net” matters most. Preparing ten hours of work saves less than ten if each proposal needs three minutes’ review. Ignoring review inflates savings, for every vendor, including us. Count weekly cases, current minutes per case, and review minutes per proposal.

Invented inputs: 150 purchase invoices weekly, four minutes each to retype and match, take ten hours. One-minute proposal review leaves seven and a half net hours, above five. Thirty invoices leave one and a half net hours. Then wait.

Volume must sit in one task, rather than ten small ones. Ten half-hour tasks total five hours but each needs explanation, rules, and review. See Which administrative processes should I automate first?.

How do you calculate whether an AI agent pays off for your business?

Use net weekly hours × 52 × employee hourly cost, divided by annual agent cost. Steding describes, translated: “Hours per week × 52 × a realistic hourly rate, minus monthly costs × 12, divided into development costs.” If payback exceeds twelve months, “the answer is almost always: not yet.” (Steding, August 10, 2026).

Three vendors published examples this year. We compared and recalculated them. All sell agents and use assumptions without measured customer data.

Both Dutch-language examples assume four to fifteen hours weekly, neither subtracting review. BrightBox uses 45 working weeks, UnifyAI four weeks monthly. These choices move results ten to fifteen percent.

Now use a fixed price. Bombos’s pilot costs € 2,000 for three months, then € 1,000 monthly, the same for everyone: € 11,000 in the first twelve months and € 12,000 each later year. Required net weekly hours depend on hourly cost:

These are thresholds, rather than results: required work, rather than what Bombos takes over. Add your people’s explanation and review time separately. Full budgets including setup and hidden costs appear in What does automating administrative processes with AI cost? and What does AI cost in a small business?.

A clearer comparison than payback: five hours weekly is just over half a workday. An extra administrative employee is forty hours. The question is whether an agent removes enough work to postpone the next vacancy. See Hire someone or use AI as work grows?.

Are staff shortages still a reason to use an AI agent?

Yes, but weaker than in 2022. Growth without extra people is now stronger. CBS counted 95 vacancies per 100 unemployed in Q2 2026, versus 142 in Q2 2022. At Q2’s end, 375 thousand vacancies were open, translated: “3 thousand fewer than a quarter earlier.” (CBS, July 30, 2026).

Source: CBS, seasonally adjusted labor-market tightness series, July 30, 2026.

Tightness declined for four years with a small final-quarter increase. Still, 95 is high: 2016 had 22. Hiring an experienced administrator or assistant accountant still feels difficult.

Small businesses respond less than large ones. Overall, 29.7 percent automate to address shortages. For small businesses with 5 to 50 employees, it is only 20.1 percent, versus 40.4 percent for large ones. Productivity investment differs more: 53.2 percent for large businesses versus only 28 percent for small ones (Nederland Digitaal on CBS, June 3, 2026).

Small businesses have the fewest people to absorb vacancies and automate least. See reducing manual administrative work without hiring.

Why does the mistake question matter more than technology?

Suitability depends less on capability than on what happens after a mistake. bConnect says, translated: “In contexts where errors are acceptable and easy to correct, agents can be deployed sooner.” (bConnect, accessed September 30, 2026). Appfront sets a lower bar than many expect: tasks “where 80% good enough delivers major time savings” (Appfront, May 2026, translated).

A wrongly prepared purchase-invoice entry caught before posting costs a minute. An incorrect reminder already sent costs a call and trust. Same agent, same error, different risk. The distinction is whether someone reviews before it leaves.

Approval makes a task suitable rather than holding it back. A customer-reaching error makes a task unsuitable without approval, but suitable with it if review is faster than manual work. See preventing AI mistakes and who reviews them.

When does an AI agent not make sense for your business?

When a task is too small, rare, needs unique judgment per case, or nobody can assess correctness. Even Anthropic recommends the simplest solution: “This might mean not building agentic systems at all.” Agents trade speed and cost for results, “and you should consider when this tradeoff makes sense.” (Anthropic, December 19, 2024).

Specifically:

• Below two net hours weekly. The calculation fails whatever the price.
• Dozens of cases weekly or fewer, spread across work types. Manual work often wins (bConnect).
• Every case differs. Appfront lists rare experience-based exceptions and sensitive customer conversations as poor candidates (Appfront, May 2026).
• Binding legal or medical decisions without human review. Also listed by Appfront.
• Data so messy “even a person cannot make sense of it” (Appfront, translated). Clean up first.
• Ordinary software already solves it. Accounting rules or scan-and-recognize features cost less. See AI or traditional software for administrative processes.

Knowledge held only in one coworker’s head looks like a “no” but need not be. An agent cannot find it, but an interview can document that person’s decisions for everyone.

How do you test in five minutes whether an AI agent makes sense?

Answer for one task, rather than the whole business. Each “yes” earns a point.

1. Does this task recur weekly in the same way?
2. Are there over a hundred cases monthly?
3. Does it take over five hours weekly across everyone doing it?
4. Can you explain successful handling in a few sentences?
5. Is required information in email, documents, or packages, rather than only someone’s head?
6. Can you detect and repair errors before a customer, supplier, or bank receives them?
7. Is work growing faster than the people doing it, or is a vacancy open for it?

Six or seven: worthwhile; calculate using the table. Four or five: worthwhile with conditions; address a “no” through approval or recorded knowledge. Three or fewer: choose another task or wait. Our scoring is a source-based rule of thumb, rather than a measured standard.

Choose one task if several score highly. See Where do I start with AI with twenty employees? and Which administrative tasks can AI take over?.

What do these figures say together?

Our position: growing businesses find agents worthwhile because they do not want headcount growth, rather than inability to hire. Review time is the real threshold. Three previously uncombined series point there.

First, tightness fell from 142 to 95 (CBS). “We cannot find anyone” weakens. “Another hire adds more weight than the work warrants” remains.

Second, small businesses automate against shortages half as often: 20.1 versus 40.4 percent (Nederland Digitaal on CBS). Their work is not less suitable. They lack someone to lead a project. The barrier is organizational.

Third, public calculations assume four to fifteen hours weekly without subtracting review (BrightBox, UnifyAI). Gartner expects over 40 percent of projects stopped by the end of 2027, partly for unclear value (RCR Wireless on Gartner). Our interpretation: examples concern one bounded task; failures mainly broad projects without clear tasks or measurements. No source explicitly says this. It is our reasoning.

Start with one task above five net hours where errors are detectable before sending. Count review from day one. A task below the threshold has not failed; its turn has not arrived.

Where is this heading?

Small-business AI use grows rapidly from a low base. Microbusiness use doubled from 6.8 percent in 2023 to 13.8 percent in 2025. Businesses with 10 to 249 working persons reached 29.8 percent (CBS, March 16, 2026). Among AI users, 32 percent use it “for business administration or management tasks” (CBS, December 12, 2025, translated). In June 2025 Gartner expected a third of business software to contain agents by 2028, versus under 1 percent then (RCR Wireless on Gartner). See AI-agent trends for businesses.

Our expectation: CBS’s 2027 figures will show over 40 percent AI use in businesses with 10 to 249 working persons. The question will be which task first. Microbusinesses doubled in two years; this group already approaches 30 percent. Existing software increasingly includes agents, exposing businesses without separate projects.

We also expect the task threshold to fall. Wages rise annually with collective agreements; fixed agent prices do not automatically do so. Higher hourly costs increase a weekly hour’s value and lower break-even hours: 6.0 at € 35 versus 3.5 at € 60 in the table.

The expectation fails if CBS measures leveling below 35 percent in 2027, or discontinued projects deter small businesses. By then, Bombos wants your first task running for a year, and your team to have taught the second and third, answering suitability with your own figures.

What can this assessment not do yet?

It cannot prove your time savings. Nobody currently can. No source we read independently measures savings at a Dutch small or midsize business. Every calculation, including three-to-seven-month paybacks, comes from agent vendors using assumptions.

Thresholds are rules of thumb. Two hours comes from Steding; five from our calculation assuming € 45 hourly. At € 35 or € 60, the threshold shifts as shown. The self-test has not been validated on real businesses.

Review time is the largest unknown, established only using your own work. A trial with a predefined measurement beats examples. See testing an AI agent before automating and starting with an AI automation pilot.

Gartner’s June 2025 figures come through a news site because its page was inaccessible. Gartner sells advice. Treat them as direction, rather than measurement.

How does Bombos approach this?

Bombos asks which task takes enough net time and where approval sits. If none qualifies, we say so. An agent does not yet suit every business.

Chef distributes incoming work to specialists. Wegwijzer explains, helps set boundaries, and suggests tasks. Specialists are built around your tasks, systems, and rules. Bombos reads connected email and documents, finds customers and cases, and prepares supported proposals. You approve, change, or reject in Bombos. Corrections become rules for coworkers too. Review time per proposal therefore declines as Bombos learns more rules.

Payments, customer messages, and contracts wait for approval by default. This makes suitable tasks where errors would otherwise reach customers. Bombos interviews coworkers to record knowledge held only in their heads.

The goal is more work at higher quality with the same team. A first task such as retyping purchase invoices or distributing the shared inbox is the beginning. Your team then teaches Bombos the next without technical skills. We guide you; then you can do it yourself.

Sources

Each source was opened September 30, 2026, and each quotation appears literally in it. Dutch quotations above are translated. Vendors are identified.

1. Steding: AI, saving costs or investing in growth?, August 10, 2026. Sells agents.
2. BrightBox: calculating AI-agent ROI, June 25, 2026. Sells agents.
3. UnifyAI: What does an AI agent cost?, May 17, 2026. Sells implementation.
4. AgileSoftLabs: AI Agents for Customer Service, ROI Calculator, July 27, 2026. Sells customer-service software.
5. bConnect: AI Agents, complete overview, undated. Sells agents.
6. Appfront: what AI agents can mean for your business, updated May 2026. Sells agent development.
7. PROMPTED: What is an AI assistant, and when do you choose an agent?, June 5, 2026. Sells agents.
8. Kweekers: 5 signs your organization is ready for an AI agent, undated. Sells its own agent.
9. CBS: unemployment fell in Q2 2026, July 30, 2026. Includes labor-market tightness series 2016-2026.
10. Nederland Digitaal: CBS, more automation against staff shortages, including AI, June 3, 2026.
11. CBS: AI use in Dutch microbusinesses, chapter 2, March 16, 2026.
12. CBS: businesses use AI most for marketing or sales, December 12, 2025.
13. RCR Wireless: Gartner, more than 40% of agentic AI projects will fail by 2027, June 27, 2025. Secondary source; Gartner’s page was inaccessible. Gartner sells advice.
14. Anthropic: Building effective agents, December 19, 2024. Makes AI models.