Voice bots in the mid-market: buy or build?

A voice bot takes calls, answers standard questions and captures requests in a structured way, around the clock. The question is rarely whether it works technically. The question is whether an off-the-shelf product fits your case or a tailored in-house build is more economical. A mid-market company tested exactly that in a proof of concept. The result and the criteria behind it.

Where voice bots work in the mid-market

Three use cases come up again and again in enquiries:

• Residential property management: Tenant enquiries and damage reports arrive in the evening and at weekends too. A voice bot captures them in a structured way, clarifies standard questions on service charges or appointments and hands the rest to the team, with a full log instead of a voicemail gap.
• Procurement: Suppliers ask about order status, invoice details or contacts. The bot handles recurring answers; the procurement team handles cases that need negotiation.
• Customer service: Opening hours, order status, spare parts requests. The bot resolves frequent cases and triages the rest.

The patterns transfer across industries; what matters is not the sector but the structure of calls: high volume, recurring requests, clear data sources.

Voice bot instead of off-the-shelf software

Starting point

A mid-market client (several hundred employees, double-digit million revenue) wanted to automate tasks in their phone communications. The obvious option, an off-the-shelf product on the market, was too expensive and too rigid for the specific requirements. The open question: is build cheaper than buy here?

Approach

We proved the feasibility of a custom voice bot in a PoC. Fast, cost-controlled, with clear acceptance criteria. No "feels doable" gut call, but a solid basis for the investment decision.

Delivery

• Requirements analysis for the prioritised phone tasks
• Prototype covering the most critical voice flows
• Cost and effort comparison build vs. buy
• Handover into the production build

Outcome

Feasibility confirmed, economics proven. The in-house build comes out cheaper and more tailored than the off-the-shelf product. The PoC has moved into a production project at wonk.ai.

Note: Anonymised case from a funded project context. Technical details and client name on request.

Buy or build: four criteria

1. Call volume and licensing model. Off-the-shelf products usually charge per call, minute or agent. At high volume, running costs grow with usage; a custom build has higher upfront cost and flatter operating cost. The crossover point can be calculated.
2. Special logic. The closer the bot needs to work with your processes and systems (ERP, ticketing, domain vocabulary), the tighter the fit for standard software.
3. Data sovereignty. Conversation data is customer data. Hosting in Germany and full control over data often favour custom build or an on-premise option.
4. Operations. Who runs and maintains the solution? Without internal or external capacity for operations, the off-the-shelf option can be the right choice despite higher cost.

An honest take: with standard requirements and modest call volume, buying is often the better answer. That is exactly what the proof of concept as a pilot project with fixed acceptance criteria is for: it answers the question with numbers instead of gut feel, in 4 to 8 weeks. If the answer is "buy", we will say so. The same approach with a fixed frame is shown in the automating donor communications case.

Frequently asked questions

What does a voice bot cost?

Off-the-shelf products typically charge per call, minute or user on an ongoing basis; costs scale with volume. A custom build costs development once plus operations. In the project described here, the PoC proved the cost comparison concretely and the in-house build was cheaper. Which side wins for you shows up in the PoC numbers.

Can a voice bot understand technical terms and dialect?

Modern speech recognition handles everyday language well. Domain vocabulary is trained and tested in the project; that is exactly what the PoC acceptance criteria are for.

How long does implementation take?

The proof of concept with the critical voice flows takes 4 to 8 weeks. After that you can decide reliably whether and how to move into production.

Can this be done in a GDPR-compliant way?

Yes. Hosting in Germany, no use of conversations for model training, data processing agreement. The architecture is designed for this from the start.

When is the off-the-shelf option the better choice?

With standard requirements, low call volume and no capacity for operations. That is not a rhetorical option: if the PoC shows it, the recommendation is to buy.