Most build-vs-buy debates end in stalemate. The champion of the off-the-shelf platform cites speed and low upfront cost. The in-house tech lead argues for control and long-term flexibility. The CEO wants a recommendation by Thursday. Nobody leaves the room with one.
This article gives you a structured way out. Run through five scoring dimensions, tally a number, and you will have a defensible position — not a gut feeling, not a vendor pitch — to put in front of your board.
One important caveat up front: “buy” does not mean a single thing. The market today splits into at least three distinct options, and choosing between them matters as much as the build-or-buy headline question.
The Three Options You Are Actually Choosing Between
Before scoring anything, be precise about what “buy” means for your situation:
Off-the-shelf SaaS agents — pre-built AI assistants from vendors (Intercom, Zendesk, HubSpot, etc.) that ship with AI features baked in. Fastest to deploy; least customisable; you work within their UX and data model.
No-code / low-code agent builders — platforms like Zapier AI, Make, Voiceflow, or Botpress that let non-engineers assemble agent logic visually. More flexible than SaaS, but they hit ceilings quickly as complexity grows. We covered this ceiling in detail in When No-Code AI Agent Builders Hit Their Ceiling.
Custom development — an agent built on open frameworks (LangGraph, CrewAI, Mastra) or from first principles, scoped and shipped by a development team. Highest upfront cost; highest long-term flexibility and integration depth.
Hybrid approaches exist too — a no-code builder for the front-end workflow, a custom model or integration layer underneath. The scorecard handles this.
The Five-Dimension Scorecard
Score each dimension 1–3. Add up the total.
- Score 5–8: Off-the-shelf or low-code platform is likely the right call.
- Score 9–11: Hybrid approach. Consider a platform for the front-end with a custom integration layer.
- Score 12–15: Custom development is probably the correct answer.
This is a guide, not a mandate. Use it to anchor the conversation.
Dimension 1 — Differentiation Value (1–3)
Ask: Does this agent touch a process that is a source of competitive advantage for our business?
- 1 point — The process is generic (scheduling, FAQ deflection, document routing). Competitors in your industry handle it the same way, and the way you handle it is not a reason customers choose you.
- 2 points — The process matters to customers but is not uniquely yours. You have specific preferences about how it works that no off-the-shelf product quite matches.
- 3 points — The process is a core differentiator. How you handle it is part of your product, your brand, or your margin story. Handing it to a platform means making your competitive advantage look identical to every other company on that platform.
A logistics firm with a proprietary quoting algorithm scores 3 here. A law firm deploying an agent to handle new-client intake questionnaires might score 2 — it matters, but it is not the firm’s legal expertise.
Dimension 2 — Integration Depth (1–3)
Ask: How deeply must this agent connect to internal systems?
- 1 point — The agent can do its job with data available via standard APIs or public integrations. Your CRM, calendar, or ticketing system is mainstream enough that every platform supports it.
- 2 points — The agent needs access to two or three internal systems, at least one of which has a non-standard interface, an on-premise component, or a bespoke data schema.
- 3 points — The agent must read from or write to ERP tables, legacy databases, proprietary internal APIs, or regulated data stores with specific access control requirements. Off-the-shelf connectors do not exist or are not trustworthy.
Swiss mid-market companies often discover they score 3 here when they start mapping actual data flows — a Sage 200 integration or a custom-built ERP is not on any platform’s native connector list, and the effort to bridge that gap belongs in the cost side of the build-vs-buy equation.
Dimension 3 — Data Sensitivity (1–3)
Ask: How sensitive is the data the agent will process, and where must it live?
- 1 point — Data is non-sensitive or already lives in a SaaS environment. Processing via a third-party platform is acceptable under your data policy and any applicable regulations (nFADP, GDPR, sector rules).
- 2 points — Data is moderately sensitive. You are willing to use a cloud platform, but you need contractual guarantees around data residency, and you would not want data stored on the platform’s shared infrastructure indefinitely.
- 3 points — Data is highly sensitive (health records, financial data, legally privileged communications, personnel files) or there are explicit requirements for on-premise or private-cloud processing (typically from sector regulators such as FINMA or medical confidentiality rules, rather than the nFADP alone). Many platforms cannot satisfy this without enterprise add-ons that radically change the cost picture.
Healthcare operators in Switzerland, fiduciaries, and law firms almost always score 3 here.
Dimension 4 — Change Frequency (1–3)
Ask: How often will the agent’s logic, outputs, or integrations need to change?
- 1 point — The process the agent handles is stable. You expect to configure it once and make minor adjustments quarterly at most. The vendor’s update cadence will not disrupt your workflow.
- 2 points — The process changes a few times a year. You can tolerate the platform’s change management overhead, but you have had bad experiences with vendor-driven updates breaking things.
- 3 points — The process changes frequently — new product lines, regulatory updates, market moves. Every change on a no-code platform means pulling a visual workflow apart and reassembling it. That overhead compounds. On a custom system, a developer changes a config or a prompt template; on a platform, it is a rebuild.
If your compliance requirements change with Swiss regulatory updates or if your product catalogue changes seasonally, score yourself honestly here.
Dimension 5 — Internal Capability (1–3)
Ask: What can your team realistically maintain after go-live?
This dimension runs in reverse. A high score here is not always better.
- 1 point — Your team has developers comfortable with Python or TypeScript, familiarity with APIs, and some experience with LLM-based systems. They can own a custom-built agent after handover.
- 2 points — You have tech-savvy operations staff who can work in a no-code environment but not in code. Custom development would need ongoing vendor support for any changes beyond configuration.
- 3 points — No meaningful technical capacity in-house. You need something a business user can maintain, or you need a fully managed service model.
For scoring: if you scored 1 here, subtract 1 from your total (internal capability reduces the cost of custom). If you scored 3, add 1 (it increases the real cost of custom, even if every other dimension points that direction).
Running the Scorecard: Two Worked Examples
Example A — A 40-person Swiss insurance broker deploying an agent to handle policy renewal reminders
- Differentiation Value: 1 (generic process)
- Integration Depth: 2 (connects to their policy management system via a well-documented REST API)
- Data Sensitivity: 2 (personal data, needs GDPR-compliant processing but cloud is acceptable with contractual guarantees)
- Change Frequency: 1 (renewal process changes rarely)
- Internal Capability: 3 (no developer on staff)
Total: 9 — Hybrid. A no-code builder like Voiceflow or Make, with careful data processing agreements, is a reasonable starting point. If the policy system’s API turns out to be less clean than documented, revisit.
Example B — A 120-person Swiss manufacturing company deploying an agent to handle incoming order confirmations, cross-reference stock in their on-premise ERP, and flag exceptions to the logistics team
- Differentiation Value: 2 (order management is central to their operations, not a differentiator per se, but their specific process is idiosyncratic)
- Integration Depth: 3 (on-premise ERP, custom stock tables, internal exception-management workflow)
- Data Sensitivity: 2 (commercial data, no hard regulatory constraints beyond standard data protection)
- Change Frequency: 2 (product mix changes quarterly, some process adjustments expected)
- Internal Capability: 1 (small internal dev team, comfortable with APIs)
Total: 10 (minus 1 for capability) = 9 — Hybrid, leaning custom. Here we would recommend starting with a custom integration layer connected to the ERP, with a lightweight front-end workflow the internal team can adjust. Full bespoke development is justifiable if budget allows.
What the Scorecard Does Not Tell You
No scorecard replaces judgment on cost. A company scoring 14 on this framework may still look at the upfront cost of custom development and decide a platform is the pragmatic choice for the next twelve months — with a plan to migrate when the limitations bite. That is a valid call.
What you want to avoid is making that trade-off without acknowledging it. A platform decision made without scoring integration depth is not a decision; it is a hope. When the platform can’t talk to your ERP in month four, the emergency custom integration costs more than getting it right the first time would have.
The total cost of ownership picture is worth examining separately — The Real Cost of AI Agents: Custom vs Platform TCO goes into the numbers that rarely appear in vendor proposals. And if platform lock-in is a concern, AI Agent Platform Lock-In: The Risks Nobody Prices In covers what the switching costs actually look like.
Who Should Build, and Who Should (Probably) Buy
| Signal | Likely direction |
|---|---|
| Process is generic and stable | Platform or SaaS |
| No internal dev capacity, limited budget | Low-code platform (with eyes open on ceiling) |
| Data must stay on-premise or in a specific jurisdiction | Custom or private-cloud deployment |
| Agent is a core part of your product or service | Custom |
| You have a working internal dev team | Custom or hybrid |
| You need a proof-of-concept in two weeks | Platform to validate, custom to scale |
| Integration requires non-standard internal systems | Custom |
| You have been burned by vendor lock-in before | Custom |
Where Orange ITS Fits
We work across all three delivery modes — we help clients validate whether a platform genuinely fits before they commit, and we build the custom layer when it does not. Our bias is toward shipping working agents, not toward any particular technology.
The teams we work with best are ones who want a second opinion on the decision before they sign a platform contract or kick off an in-house build. Getting the build-vs-buy call wrong at the start is the single most common reason AI agent projects fail to deliver — it is cheaper to get it right in a structured conversation than to unwind a six-month platform deployment.
If you are working through this decision now, a 30-minute call with our team at Orange ITS gives you a clear recommendation for your specific agent initiative — not a generic pitch for one approach. We will ask about your stack, your data, and your internal capacity, and tell you honestly where we think the risk sits.
Book a 30-minute build-vs-buy scoping call
We are based in Chiasso, work across Switzerland and Europe, and have shipped custom AI agent development for companies from 15 to 500 staff. No obligations, no slideware.