Most AI agent discussions stop at the upfront question: “How much does it cost to build?” That is the wrong question. The number that actually hits your P&L is the three-year total cost of ownership — and for most business workloads, that number looks very different from what a platform’s pricing page suggests.
This article replaces the “no-code is always cheaper” intuition with an actual cost model. By the end, you will know the break-even volume for your specific workload and which path makes financial sense before you sign anything.
Why the First-Year Number Misleads
Platform agents look inexpensive at launch. Monthly subscriptions start low, there is no engineering work to commission, and you can have something running in days. Those are real advantages — for the right workload.
The problem is that platform pricing compounds. Per-task fees, token consumption charges, and connector fees grow with every workflow you automate. The degree of compounding varies sharply by platform category: lightweight workflow automation tools (Make, n8n) scale more gently than customer-service AI agent platforms (Zendesk, Intercom), which charge per resolution at rates that add up quickly at volume. Maintenance also accumulates: prompt engineering after each model update, manual workarounds when the platform deprecates a feature, and re-wiring when your CRM or ERP releases a new API version. None of those line items appear on the pricing page.
Custom development works the opposite way. Year one is the expensive one — design, build, testing, deployment. Years two and three are mostly infrastructure costs (compute, model APIs) and periodic enhancement work. The cost curve is front-loaded, then flat.
The crossover point is the question worth asking.
A 3-Year TCO Model: Where the Lines Cross
The numbers below are illustrative scenarios, not case study data. Use them to build your own model — substitute your actual task volumes and vendor quotes.
Scenario: A Support Triage Agent at Mid-Volume
Assume a company processing roughly 2,000 inbound support interactions per month, routed and classified by an AI agent before reaching a human team.
Platform path (illustrative):
| Cost item | Year 1 | Year 2 | Year 3 |
|---|---|---|---|
| Platform subscription | CHF 6,000 | CHF 7,200 | CHF 8,400 |
| Per-task / token fees | CHF 4,800 | CHF 5,760 | CHF 6,912 |
| Integration setup (internal time) | CHF 5,000 | CHF 1,500 | CHF 1,500 |
| Prompt maintenance & updates | CHF 2,000 | CHF 3,000 | CHF 3,500 |
| Year total | CHF 17,800 | CHF 17,460 | CHF 20,312 |
| 3-year cumulative | CHF 55,572 |
Platform costs vary enormously by tool category: lightweight automation platforms (Make, n8n) start from €20–€50/month at this volume; customer-service AI agent platforms (Zendesk, Intercom) charge $1.50–$2.00 per automated resolution plus per-seat add-ons, reaching CHF 2,000–10,000+/month at 2,000 interactions. These figures are composites representing the CX-platform tier; get a direct quote before modelling.
Custom development path (illustrative):
| Cost item | Year 1 | Year 2 | Year 3 |
|---|---|---|---|
| Custom build (design, dev, test) | CHF 22,000 | — | — |
| Infrastructure & model API costs | CHF 2,400 | CHF 2,880 | CHF 3,200 |
| Monitoring & maintenance | CHF 3,000 | CHF 2,500 | CHF 2,500 |
| Feature enhancements | CHF 2,000 | CHF 3,000 | CHF 2,000 |
| Year total | CHF 29,400 | CHF 8,380 | CHF 7,700 |
| 3-year cumulative | CHF 45,480 |
At this volume and these illustrative figures, custom development reaches parity somewhere in month 22–26 and finishes the third year roughly CHF 10,000 ahead — and that gap widens as volume grows, because the platform’s per-task fees keep scaling while the custom agent’s marginal cost is almost zero.
What Shifts the Break-Even
Four variables move the crossover point significantly:
- Task volume. Below a few hundred tasks per month, platforms win almost every time. Above 3,000–5,000 tasks per month, custom typically wins. The exact threshold depends on the platform’s per-task rate.
- Complexity and integration depth. Platforms charge separately (or cap functionality) for multi-step workflows, non-standard connectors, and high message volumes. Complex workflows close the gap faster than simple ones.
- Model update cycles. Every time an underlying LLM is deprecated or a platform changes its prompt interface, someone spends time fixing prompts. That labour is invisible in year one but compounds noticeably by year three.
- Migration risk. If you eventually outgrow the platform, you rebuild. That rebuild cost is a hidden liability sitting on the platform-path balance sheet from day one. See AI Agent Platform Lock-In: The Risks Nobody Prices In for a full treatment.
The Costs That Never Appear in a Vendor Pitch
Failure and Rework
Platforms fail in specific ways: rate limits hit at peak load, classification breaks after a model update, a connector goes down mid-process. Each failure has a cost — either in staff time to investigate and patch, or in customer experience damage. Custom agents fail too, but the failure modes are yours to design around, and fixing them does not require navigating a vendor’s support queue.
The Rebuild Tax
The most under-priced risk in the platform path is having to rebuild later. A team that spends CHF 15,000 running a platform agent for 18 months and then migrates to custom development does not save CHF 15,000 — it pays twice: once for the platform years and once for the rebuild. The 3-year TCO model only captures this if you assign a probability-weighted cost to migration.
A related read: Outgrowing Your Agent Platform: The Migration Path to Custom.
Internal Labour
Prompt engineering, connector configuration, and testing workarounds all consume internal time that rarely appears in a formal cost model. In our experience, teams underestimate this by a factor of two in year one and a factor of three by year three, as the agent scope expands.
When Platforms Are the Right Answer
This is not an argument that custom development always wins. Platforms make sense when:
- Volume is low and genuinely expected to stay low. Under ~500 tasks per month for a simple workflow, the build cost rarely pays back within three years.
- You need to validate a use case first. A platform can run a proof of concept in weeks. If the business case does not materialise, you have not sunk significant development budget. (Though be deliberate about the boundary between proof-of-concept and production commitment — they often blur.)
- The workflow is genuinely simple and stable. If your agent does one well-defined thing that will not change for years, a platform’s standardised tooling is sufficient and sensible.
For a structured way to think through the build-versus-buy decision beyond pure cost, Build vs Buy: A Decision Framework for AI Agents covers the non-financial factors: data sensitivity, competitive differentiation, and internal capability.
Running Your Own Numbers
The model above is a scaffold. To make it yours:
- Count your monthly task volume today and estimate a realistic growth rate over 36 months.
- Get a real platform quote for your exact workflow — not the website’s headline price. Ask specifically about per-task fees, connector costs, and what happens to your bill if volume doubles.
- Price the internal labour honestly. Who maintains the prompts? Who handles failures? Multiply hours by a fully-loaded hourly rate.
- Assign a migration probability. If there is a 40% chance you outgrow the platform in three years and face a rebuild, add that expected cost to the platform column.
- Compare year-by-year, not just three-year total. If your cash position matters, the front-loaded nature of custom development may be a constraint even if the three-year number favours it.
If you want help with the numbers for your specific workload, that is exactly what a short discovery call is for. No obligation, no sales deck.
What Determines the Real Winner
The ai agent total cost of ownership question is not ideological — platforms are not lazy, and custom development is not always the premium choice. The honest answer depends on three things your vendor will not calculate for you: actual task volume over time, the real cost of internal maintenance labour, and the probability of needing to rebuild.
For most Swiss SMBs running above moderate task volumes on processes that touch core systems (CRM, ERP, customer-facing workflows), custom development reaches parity by year two and is cheaper by year three. Below that threshold, platforms are a rational starting point — provided you go in knowing where the ceiling is.
See When No-Code AI Agent Builders Hit Their Ceiling for the specific signals that tell you when it is time to move.
Upfront build costs — what custom development actually costs to commission — are covered separately in What AI Agent Development Really Costs in 2026.
Ready to run this model against your actual workload? Orange ITS works with SMBs across Switzerland and Europe to size AI agent projects before any commitment is made. A 30-minute call is enough to map your task volumes, estimate TCO for both paths, and tell you honestly which makes financial sense. Book that call here.