Make, Buy or Partner: AI Agent Procurement for DACH Mid-Market Executives 2026

Key numbers at a glance

• Cloud API €37 per month vs own GPU hardware €1,750 plus electricity at light usage (under 1M tokens per day). At this scale cloud is clearly economically superior.
• 30,000 to 80,000 EUR typical pilot budget for externally built agent in DACH mid-market 2026, 90,000 to 200,000 EUR for scaling. Internal build at 1-to-1 comparability costs 30 to 50 percent more gross engineering time.
• 1 in 10 is the success rate of pure internal agent builds in DACH mid-market 2026 without external accompaniment. Main reasons: skill library architecture and KPI measurement are missing from the internal repertoire.
• 8 to 12 weeks realistic timeframe for a robust pilot, 6 months for production scaling. Whoever wants to be productive in 4 weeks either has a very narrow use case or skips pre-production checks (see architecture failures post).
• 3 typical mid-market models: SaaS agent from vendor (€199-2,000/month), external implementation partner (€30-200k engagement), internal build (€80-300k initial investment plus ongoing FTE cost).

---

If you are a managing director, CTO or Head of Operations at a DACH mid-market company in 2026 deciding on procurement of a first AI agent, you face the make-buy-partner question. The answer is not ideological ("we want independence" or "we want speed") but along 5 decision axes. This post delivers the axes, a decision tree, and a cost estimate per model from DACH mid-market engagement practice 2026.

Important upfront: the three options are not exclusive. In most of our 2026 engagements we see mix models: SaaS agent for standard workflows, external partner for custom implementation, internal team for skill library care and KPI measurement. The question is not "which model" but "which shares per use case."

Who this post is for and who it is not

This post is for decision makers in DACH mid-market (30 to 500 FTE) who have identified a use case (see our 90-day use case matrix) and face the procurement decision. Concretely: you have a use case charter, a pilot budget between 30,000 and 200,000 EUR, and now you have to decide whether to buy a SaaS agent, hire an external partner, or build internally.

Not a fit for corporates above 500 FTE with a dedicated AI engineering team. For them the make decision is often default-rational because FTE costs scale differently.

The 5 decision axes

Axis 1: standardisation degree of the use case. Standard use cases (incoming invoices, mail triage, lager monitoring, standard outreach) have ready SaaS agents on the market. Non-standard use cases (industry-specific workflows, own data models, regulated special processes) need custom implementation because SaaS agents reach 60-percent coverage but do not cover the critical 40 percent.

Axis 2: data sovereignty requirements. Use cases with standard data (public information, standard CRM data) are cloud-API-suitable. Use cases with sensitive data (patient data, financial data of regulated industries, engineering IP) need on-premise or dedicated EU cloud hosting constellations, which makes the make option more attractive.

Axis 3: time to value. If you must be productive in 6 months, buy or partner is the default recommendation because internal build typically needs 9 to 18 months time-to-first-production. If you have 12 to 24 months and want strategic independence, make investment is worthwhile.

Axis 4: skill availability internally. Do you currently have at least 1 senior engineer with AI engineering experience (LangGraph, Claude Agent SDK, MCP, OpenAI Assistants API) who is dedicated for 12 to 18 months? If no, make is a several-hundred-thousand-EUR bet without sufficient internal capacity. If yes, it is worthwhile for strategic use cases.

Axis 5: strategic differentiation. Is the use case a differentiation factor against competitors (own customer service agent as USP, own engineering pipeline optimisation)? Then make is long-term valuable. Is the use case a standard efficiency optimisation (incoming invoice, travel booking, standard reporting)? Then buy is default-rational because differentiation does not lie in the workflow but in the core business.

Model 1: SaaS agent from vendor (Buy)

What it is: you buy a pre-built agent from the vendor (e.g. Salesforce Agentforce, ServiceNow AI Agents, industry-specific providers), configure it on your data flows, and use it as a service.

Costs 2026 in DACH mid-market:

• Licence costs: typically €199 to €2,000 per month per agent or per 100 to 500 actions
• Setup costs: typically €5,000 to €30,000 for initial configuration
• Hidden costs: API token costs at higher usage (can scale from €37/month to €2,000+/month), skill library build (see our cost spike post), compliance setup, onboarding workshops

When it fits: standard use cases (axis 1 standard), low data sensitivity (axis 2), fast time to value (axis 3 under 6 months), no internal AI skills (axis 4 no), no strategic differentiation (axis 5 efficiency).

When it does not fit: custom workflows that SaaS coverage does not reach, regulated data flows that vendor cloud does not allow, use cases with differentiation value.

Risks: vendor lock-in (migration costs 4 to 12 weeks engineering if at all possible), pricing changes (vendor raises prices every 6 to 12 months), roadmap dependency (what vendor does not build does not come).

Practical note: SaaS agents are very good in many standard domains in 2026. If 70 percent of your use cases are standard, you should evaluate SaaS before considering make investment.

Model 2: external implementation partner (Partner)

What it is: you hire an external implementation partner (consultancy, boutique AI firm, or specialised vendor like Sentient Dynamics) for custom implementation of an agent in your stack. The partner builds, hands ownership over to your team after 6 to 12 months.

Costs 2026 in DACH mid-market:

• Pilot engagement: typically €30,000 to €80,000 for 90-day pilot
• Production scaling: typically €90,000 to €200,000 for 3 to 5 workflows
• Ongoing maintenance: typically €30,000 to €80,000 per year for skill library care, drift monitoring, model updates
• Success-based models: some partners (incl. Sentient) offer output-measured compensation, that lowers initial risk

When it fits: medium to high custom requirements (axis 1 custom), medium data sensitivity (axis 2), time to value 6 to 9 months (axis 3), no sufficient internal skills (axis 4 no), strategic use cases with long-term internalisation plan (axis 5).

When it does not fit: pure standard use cases (SaaS is cheaper and faster), pure make strategy where internal sovereignty is absolute prerequisite.

Risks: partner lock-in if handover to internal team is not done properly, knowledge loss on partner switch, quality fluctuations between different partner engagements.

Practical note: partner models work best with clear handover clause: after 6 to 12 months productive phase your internal team takes ownership of the skill library, the partner stays as sparring resource for escalations. That avoids partner lock-in and builds internal capability.

Model 3: internal build (Make)

What it is: you build the agent with your internal engineering team, typically with 1 senior engineer with AI experience plus 2 to 3 junior to mid-level devs. Open-source frameworks (LangGraph, AutoGen, Claude Agent SDK), own skill library, own permissions architecture.

Costs 2026 in DACH mid-market:

• Initial investment: typically €80,000 to €300,000 for 6 to 12 months build phase, depending on use case complexity and engineering FTE cost
• Ongoing costs: typically 1 to 2 FTE for skill library care, model monitoring, drift detection (€100,000 to €240,000 per year)
• Cloud or on-premise infrastructure: €5,000 to €50,000 per year for cloud API tokens, or €50,000 to €200,000 for own GPU infrastructure
• Hidden costs: senior engineer opportunity costs (what would they have done otherwise), onboarding time for new team members, engineering plan trade-off against other initiatives

When it fits: highly custom workflows (axis 1 custom), high data sensitivity (axis 2), time to value 12 to 24 months acceptable (axis 3), strong internal AI skills (axis 4 yes), strategic differentiation (axis 5 USP).

When it does not fit: standard use cases (SaaS is 10x cheaper), no internal AI senior available, fast time to value required.

Risks: time to value often 50 to 100 percent longer than planned (typically 18 instead of 12 months), senior engineer risk on personnel change, skill library tech debt without external pattern experience flowing in.

Practical note: in 1 of 10 cases pure make works without external accompaniment in DACH mid-market 2026. Main reasons: skill library architecture and KPI measurement are specific know-how from multiple production rollouts that internal teams typically do not have. Make plus external sparring (1 to 2 days per quarter) is significantly more successful than pure make.

Decision tree: which model for which situation?

From DACH mid-market engagement practice 2026 this decision tree as orientation:

Question 1: is the use case standard (incoming invoice, mail triage, inventory monitoring, standard outreach)?

• Yes → Question 2
• No → Question 3

Question 2 (for standard): do you need on-premise or is cloud acceptable?

• Cloud acceptable → SaaS agent is first choice. If SaaS coverage above 70 percent → buy. If below 70 percent → partner for custom adjustment.
• On-premise needed → partner with on-premise experience, or make if internal AI skills are present.

Question 3 (for custom): do you have an internal AI senior with 12 plus months availability?

• Yes → make plus external sparring is optimal. Senior builds, external sparring brings pattern knowledge.
• No → partner is the default recommendation. Plan handover clause to internal team after 6 to 12 months.

Question 4 (for strategic use cases): is the use case a differentiation factor?

• Yes → make or partner with internalisation plan. Never SaaS because you become dependent on vendor roadmap.
• No → prefer SaaS, save engineering time for differentiation use cases.

60-minute sparring on your make-buy-partner decision →

Mix models from DACH mid-market practice 2026

In most of our 2026 engagements we see not pure make, buy or partner strategies but mix models. Three typical constellations:

Mix 1: SaaS for efficiency, partner for differentiation. Incoming invoices and mail triage via SaaS agents (buy), own customer service agent via partner with internalisation plan (partner-to-make). Advantage: fast ROI on standard workflows, long-term differentiation on the USP workflow.

Mix 2: partner for initial build, make for ongoing care. First use case is built with partner, skill library and permissions architecture are built up. After 9 to 12 months internal team takes ownership and builds further use cases on the existing architecture. Advantage: fast first success, then internal capability without vendor lock-in.

Mix 3: hybrid cloud with partner sparring. Standard workflows use cloud APIs (OpenAI, Anthropic, Google), sensitive workflows run on on-premise open-source models (Llama, Mistral). Partner delivers pattern knowledge and architecture sparring, internal team builds and operates. Advantage: data sovereignty for sensitive data, cost efficiency for standard data.

Frequently asked questions

What about the large consultancies (McKinsey, BCG, Capgemini, Accenture)? They have agent implementation practice but typically high list prices (€500 to €2,000 per day per consultant plus premium factors) and little DACH mid-market specificity. For 500 plus FTE organisations that can make sense, for 30-200 FTE mid-market smaller specialised partners are typically more efficient.

Is open source really worthwhile from 2026? For many use cases yes. Llama, Mistral, DeepSeek reach production quality in 2026 for standard workflows, with data sovereignty advantage and lower variable costs. Trade-off: higher fixed costs for GPU infrastructure, slower model updates, less tool integration. Break-even against cloud API costs typically from 5 to 10 million tokens per day, which is too much for most mid-market workflows.

What does a bad make decision cost? In our 2026 engagements we see typical costs of an aborted internal build initiative: 6 to 12 months engineering time (€60,000 to €200,000 FTE cost), opportunity costs of other initiatives, plus reinvestment for partner pivot or SaaS switch. Total typically €150,000 to €400,000 for a make decision revised as failed after 12 months.

What does a bad buy decision cost? Typically less than bad make decision: SaaS subscription can be cancelled, setup costs of €5,000 to €30,000 are lost. But: 6 to 12 months lock-in with insufficient tool, plus reinvestment for better tool. Total typically €50,000 to €120,000.

How do I check whether a partner is good? Four questions from our procurement advisory: (1) how many production engagements in DACH mid-market have they completed? (2) Which pre-post KPI data can they show? (3) How is the handover clause to internal team formulated? (4) Which vendor diversity do they know (only Anthropic, or also OpenAI, Google, open source)? More detail in the bullshit trainings post, the pre-procurement checklist applies analogously to agent implementation.

What is Agentic AI? The executive crash course →

Sources

• Kai Gondlach: Make, Buy or Partner – AI Products
• Hardwarewartung: Make or Buy AI infrastructure
• OneReach Agentic AI Stats 2026
• Salesforce / DMB KI-Mittelstandsindex 2026
• Bitkom AI Study 2026 (PDF)
• Kai Waehner: Enterprise Agentic AI Landscape 2026
• Mittelstand Digital: AI agents vs chatbots
• PwC AI Performance Study 2026

---

About the author

Sebastian Lang is co-founder of Sentient Dynamics and leads the Agentic University programme. Before Sentient he was responsible for AI workforce programmes at SAP's Strategy Practice with 15+ years of engineering leadership experience. Sentient Dynamics works on a success-based compensation model and is deployed across the SHD and Bregal portfolios.

Subscribe to the newsletter | Subscribe to the newsletter | Sebastian on LinkedIn