A vendor's AI demo always looks great — accurate transcription, organized summary, beautiful dashboard. The problem is the demo runs in lab conditions: clean audio, a short call, a pre-defined classifier. In a real call center — with background noise, fast-talking agents, long and complex calls — performance can be very different. The five questions below expose the gap before you pay.
1. What is your accuracy on telephony audio (8kHz)?
Most demos use studio-quality audio — 16kHz and above. But a regular phone call is 8kHz, compressed, sometimes with background noise. Models trained on audiobooks or studio recordings drop significantly in accuracy on real call-center audio.
How to check: ask the vendor to run a demo on a real audio file of yours, not on their standard file. Ideally a 5+ minute call with background noise and mild speaker overlap.
2. What happens when an agent speaks with a non-native accent?
Your agents are likely a mix of native Hebrew speakers, immigrants, and second-language speakers. Customer calls may include Arabic, Russian, or French native tongue. A model trained only on native speakers drops in accuracy on everyone else.
How to check: in the demo, give the vendor a real call file with a non-native accent. If accuracy drops dramatically, that's a red flag.
3. Which Israeli telephony platforms do you support out of the box?
Marketing claims about "VoIP integration" are vague. A vendor that supports your Israeli telephony out of the box (Voipe, Origami and others) saves months of development. A vendor that doesn't support, but "can integrate after custom development" — that's more cost and more time.
How to check: ask for a specific list of supported platforms, and where the vendor has actually seen customers deploy successfully.
4. What is the pricing model — per agent, per call volume, or per transcription minute?
Different pricing models create very different cost outcomes for the same call center. Per-agent pricing fits stable organizations. Per-volume pricing fits call centers with fluctuation. Per-minute pricing can be especially expensive in organizations with long calls.
How to check: ask for a price forecast based on your specific numbers (X agents, Y calls per day, Z minutes per call on average). A vendor that cannot provide a forecast — that's a red flag.
5. How do we add a new call type 6 months after deployment?
After the system has run for a few months, it becomes clear that there are call types you didn't anticipate during deployment. It's important to know the process for adding a new classifier:
- Can you define it yourself through a UI, or do you need to file a support ticket?
- How long does configuration take — hours, days, or weeks?
- Is there an additional cost per new classifier?
A vendor that says "contact us for every change" — means the system isn't flexible. A vendor that shows a UI where you can define a classifier yourself — much better for the long term.
Bonus — 6. Which of your customers have been running the system for over a year?
New AI vendors will show "active" customers, but the real measure of success is customers who stayed over a year. Worth asking for 2-3 referrals from customers running the system for over a year — and asking them directly.
These five (or six) questions won't save you from all the friction of deploying an AI transcription system. But they will expose vendors that aren't ready to support a real call center — before you pay.
For the full list of questions and category-level comparison, see our comparison guide for Hebrew transcription tools.
Get conversation-intelligence insights
Practical writing on call-center performance, QA and coaching - straight to your inbox.
