Build or buy is the first strategic question in almost every AI project, and in 2026 the market has given a fairly clear answer: buy more, build less. In 2025, 76% of enterprise AI use cases were bought rather than built, up from 53% a year earlier, and AI deals reached production in 47% of cases against 25% for traditional SaaS (Menlo Ventures). MIT NANDA found that solutions bought from specialised vendors or built with partners succeeded in about 67% of cases, while internal builds succeeded far less often (Fortune).
But “buy” covers very different things — from a $49-a-month AI receptionist to a custom agent that a partner builds inside your own cloud account. And “build” can mean hiring a team or commissioning one system that you own completely. This article gives you a framework to decide: four options instead of two, seven questions, a scoring matrix and a three-year cost comparison.
Four options, not two
| Option | Examples | Time to value | Upfront cost | Running cost | Control and ownership | Best when |
|---|---|---|---|---|---|---|
| Platform-native agent | Intercom Fin, Zendesk AI, HubSpot Breeze, Salesforce Agentforce, Copilot Studio | Days to weeks | Low | Per outcome or credits | Low: lives inside the vendor platform | Standard use cases inside a tool you already use |
| Vertical SaaS | AI receptionists such as Rosie ($49–299 per month) or Smith.ai ($150); real-estate ISA Structurely ($499 per month plus $2,000 setup) | Days | Low | Subscription | Low to medium | Common workflows in your industry |
| Partner build on your stack | Self-hosted n8n plus custom code, models via API, MCP connectors | 2–8 weeks | Medium ($10–50k+) | Hosting, models, retainer | High if built in your accounts | Custom actions, integrations, data residency |
| In-house build | Your own engineering team | Months | High (hiring) | Salaries | Full | Core differentiation and a strong internal team |
Sources for the prices: Rosie, Smith.ai, Rework, Lets-Viz.
The two middle options are where most SMB and mid-market decisions land. Platforms keep absorbing simple scenarios — n8n is now embedded in SAP Joule Studio (Tech.eu), agents ship inside every major CRM — so a plain chatbot is no longer a product. Value moves into integrations, custom actions, data and process design.
Seven questions that decide it
1. Is this process a differentiator? If customers choose you because of how you handle it — your pricing logic, your expert knowledge, your service promise — owning the system matters. If it is hygiene, such as password resets or order status, buy.
2. Does a platform agent already cover most of it? Test first. If the agent in your helpdesk resolves the bulk of cases, Fin’s $0.99 per resolution sets the ceiling for what a custom solution may cost (Fin). Build only for the gap.
3. How deep are the integrations and actions? Answering questions is easy; issuing a refund in Shopify, changing a booking in your field-service software or posting an invoice to the ERP is not. Integrations account for 40–60% of a RAG project budget (ment.tech), and that is exactly where off-the-shelf agents stop.
4. How sensitive is the data, and where must it live? In Germany, 66% of AI users name data protection as their top barrier (Bitkom). Across Europe, 77% of companies consider the AI vendor’s country of origin and 66% worry about dependence on foreign AI (Deloitte data via DEV). If you need EU hosting or self-hosted models, your options narrow quickly.
5. What volume, and how will cost scale? Per-outcome fees grow linearly with volume; a custom build has mostly fixed costs. The crossover point is often lower than people expect (see the cost model below).
6. Do you have the people? Among German companies not yet using AI, 85% cite a lack of skills. A senior AI developer costs about $8,300 per month fully loaded (Lets-Viz); AI specialists in US agencies bill $175–350 per hour (Digital Applied), and AI/ML skills carry a 40–60% premium over general development rates in Western Europe (Index.dev). One person is also a single point of failure.
7. How easily can you leave? Ask what happens to your data, prompts, workflows and evaluation sets if the vendor raises prices, changes its model or disappears.
A simple scoring matrix
Score each criterion from 1 (clearly favours buying) to 5 (clearly favours building), multiply by the weight and add up.
| Criterion | Weight | 1 = buy | 5 = build |
|---|---|---|---|
| Differentiation | 25% | Commodity process | Core of your offer |
| Integration depth | 20% | Read-only answers | Write actions across several systems |
| Data sensitivity and residency | 15% | Public or low-risk data | Special-category data, strict residency |
| Volume economics | 15% | Low or uncertain volume | High, stable volume |
| Internal capability | 10% | No technical team | Experienced AI engineers in-house |
| Lock-in risk | 10% | Easy to replace | Critical, hard to migrate |
| Time pressure | 5% | Needed next month | Can wait a quarter |
Reading the result: below 2.5 — buy a platform agent or SaaS. Between 2.5 and 3.5 — hybrid: buy the platform, have a partner build the edge. Above 3.5 — build, with a partner or in-house depending on your capability score.
Two illustrations. An online shop’s order-status assistant scores roughly 1 / 3 / 2 / 3 / 1 / 2 / 1, a weighted total of 1.95: test the helpdesk’s native agent and add a small custom action if needed. A logistics company’s document pipeline (CMR, invoices and customs papers into the TMS and ERP, EU data) scores 3 / 5 / 4 / 4 / 2 / 4 / 3, a weighted total of 3.7: a build — and with a capability score of 2, a partner build in the company’s own accounts.
A three-year cost comparison
Consider a support agent. Option A is a per-outcome platform agent at $0.99 per resolution. Option B is a custom agent on top of your existing helpdesk: an assumed $30,000 build (inside the $10–50k market range) plus monthly running costs for support, hosting and model usage that grow with volume — assumed at $3,500, $4,000 and $5,000 per month for the three volumes below. Model usage per agent at scale typically runs $50–500 per month (Lets-Viz), so the retainer dominates.
| Resolutions per month | Per-outcome SaaS, 3 years | Custom agent, 3 years | Cheaper |
|---|---|---|---|
| 2,000 | $71,280 | $156,000 | SaaS |
| 5,000 | $178,200 | $174,000 | Roughly equal |
| 15,000 | $534,600 | $210,000 | Custom |
Caveats matter: the SaaS figures exclude helpdesk seats and platform fees; the custom figures exclude your internal time; resolution rates may differ between the two; and the custom agent can perform actions the platform agent cannot. Run the model with your own volume and your own quotes — the point is the shape of the curve, not the exact crossover. Our guide to outcome-based AI pricing explains how to compare per-resolution, per-conversation and per-action offers.
The hidden risks of building in-house
- Speed is not what it feels like. In a METR study, experienced developers using AI tools were 19% slower, while believing they were 20% faster (METR). 66% of developers complain about AI solutions that are “almost right” (Stack Overflow).
- Security debt. A scan of 5,600 public apps built with AI coding platforms found more than 2,000 serious vulnerabilities, 400 exposed secrets and 175 personal-data leaks (SoftwareSeni).
- Operations never end. Evals, monitoring, model updates and knowledge-base maintenance continue after launch. Budget for them or the system decays.
The hidden risks of buying
- Agent washing. Gartner estimates that only about 130 of the thousands of vendors marketing “agentic” products offer real agentic capabilities (MarTech).
- Vendors that are not what they claim. Builder.ai raised hundreds of millions of dollars, including from Microsoft, and became insolvent in May 2025 after it emerged that people did much of the work and revenue had been overstated (Wikipedia).
- Vendors change course. OpenAI shut down its Sora video app in April 2026 and the API in September 2026, despite a high-profile partnership deal with Disney (Wikipedia). Anything built on a single vendor inherits that risk.
- Pricing changes. HubSpot switched its AI agents to outcome-based pricing in April 2026 (Resolve247). Usually good for buyers, but your budget follows someone else’s decision.
The hybrid that usually wins: buy the platform, build the edge
For most mid-sized companies, the answer is neither pure build nor pure buy:
- Buy the commodity layers: foundation models via API, your helpdesk or CRM and its native agent, and an automation platform. n8n costs €20, €50 or €667 per month in the cloud, and the self-hosted community edition is free (n8n); see our comparison of n8n, Make and Zapier.
- Build the differentiating edge: custom actions, integrations with your systems, evaluation sets, guardrails and your own knowledge pipeline.
- Use open standards for the joints. The Model Context Protocol was donated to the Linux Foundation’s Agentic AI Foundation in December 2025, and there are more than 10,000 public MCP servers (MCP blog, Digital Applied). More in MCP explained.
- Keep ownership. Run the solution in your accounts, with your keys, and make sure code, prompts and data transfer to you contractually — the topic of avoiding AI vendor lock-in.
Vendor due-diligence checklist
When you buy — a platform, a SaaS tool or a partner — ask for:
- References with before-and-after metrics, not just logos.
- A data processing agreement, the list of subprocessors and an EU hosting option.
- Export of your data, conversation logs, prompts and configuration in usable formats.
- A written definition of billable outcomes, caps and price-change terms.
- Evaluation and quality reports, plus how regressions are detected.
- Security evidence (for example SOC 2 or ISO 27001) and the incident process.
- An exit clause with a handover period.
When to revisit the decision
Build vs buy is not a one-time verdict. Put a review in the calendar once a year, and earlier when one of these triggers occurs:
- Volume crosses the break-even point you calculated in the three-year model, in either direction.
- A vendor changes pricing or terms, for example a switch from seats to outcomes or new minimums.
- The platform catches up. If the native agent now covers what you built custom, retire your code and save the maintenance.
- A new compliance requirement appears, such as EU hosting, disclosure duties or a high-risk classification of the use case.
- The process becomes strategic. What was hygiene last year may now be part of your offer and worth owning.
Companies that review regularly avoid both classic traps: paying platform fees for years after the volume justified a build, and maintaining custom code long after a platform made it obsolete.
Typical mistakes
- Building what the platform already does. Always test the native agent first.
- Buying a demo, not a workflow. If the tool cannot act in your systems, staff will copy and paste.
- Comparing year-one costs only. Per-outcome pricing and custom builds have different cost curves; use three years.
- Hiring one AI engineer and calling it a team. Production needs review, operations and a deputy.
- Ignoring exit costs. Migrating prompts, data and workflows later is expensive if you never owned them.
- Letting the vendor define success. Your baseline, your metric, your data source.
Conclusion
Build vs buy is not a matter of principle; it is a matter of fit. Buy where the process is standard and a platform does it well. Build where the process differentiates you, where deep integrations or data residency are required, or where volume makes per-outcome fees expensive. For most SMB and mid-market companies, the winning setup is hybrid: buy the platform, build the edge, own the data.
That is how we build in Lotus Libre Build and Lotus Libre Agents: in your accounts, on open standards, with code, prompts and evaluation sets that belong to you. If you are facing a build-or-buy decision right now, book a free 15-minute audit and we will score your use case with you.