In short: automated call analysis works in three steps. The call recording is turned into text (Deepgram, OpenAI, Google and Lithuanian solutions already do this well in Lithuanian), a large language model scores the conversation against your criteria, and the results, with timestamps, go to your CRM or a report. Instead of a small share of randomly picked calls, every call gets scored.
I checked vendor prices, language support and legislation on October 17, 2026. I am not affiliated with any of the vendors mentioned. The legal information is general and is not legal advice.
How AI analyzes calls: the short answer
- Recording. The phone system (3CX, Baltneta PBX, Tele2 mobile PBX, Aircall and others) records the call and reports through an API or webhook that a new recording is ready.
- Transcription. The audio is turned into text, separating the agent from the customer (diarization). It works best when the phone system records the two speakers on separate channels.
- Scoring. An LLM (GPT, Claude, Gemini) scores the call against a rubric: greeted the customer, clarified the need, offered a solution, used the required phrases. Each score comes with a quote and a timestamp.
- Results. The summary and scores are written to the CRM or helpdesk, complaints and risky calls are sent to the manager right away, and a team report arrives once a week.
The key difference from listening by hand is not only volume. AI scores every call against the same criteria, so comparisons between employees and weeks become fairer. But AI makes mistakes, so the manager must always be able to click a timestamp and hear the exact moment.
Which AI speech-to-text understands Lithuanian best?
Transcription is the foundation of the whole process: if the system "hears" the wrong amount or surname, even the best language model will reach the wrong conclusion. As of October 2026, these solutions support Lithuanian:
| Solution | Lithuanian | Price | Notes |
|---|---|---|---|
| Deepgram Nova-3 | Yes; the Lithuanian model was updated in August and September 2026 | $0.0043/min for recordings (Pay As You Go) | Fast, speaker separation, keyterm prompting |
| OpenAI gpt-4o-transcribe / gpt-4o-mini-transcribe | Yes (multilingual models) | $0.006/min / $0.003/min | Simple integration; check accuracy on numbers and surnames |
| Whisper (open source) | Yes | Free license, you pay for your own server (GPU) | Data does not leave your infrastructure; needs maintenance |
| Google Cloud Speech-to-Text (Chirp 3) | Yes, lt-LT | Per the Google Cloud price list | Speaker separation is not supported for every language; check lt-LT |
| AssemblyAI | Multilingual models; check Lithuanian support per model | Universal-2 $0.15/hr, Universal-3.5 Pro $0.21/hr | EU region at the same price |
| Neurotechnology Call Analyzer (Vilnius) | Yes, its own Lithuanian speech-to-text technology | Custom | Ready-made solution: transcription + LLM analysis + charts |
Sources: Deepgram pricing and changelog, AssemblyAI pricing, Google Chirp 3 documentation, Neurotechnology Call Analyzer. Microsoft Azure AI Speech also supports Lithuanian.
What to check before choosing. There are almost no public comparisons of Lithuanian phone call transcription, and the accuracy figures vendors publish are measured on clean audio. Take 20–30 of your own recordings (with noise, fast speech, Russian or English words mixed in), run them through 2–3 vendors and mark errors in numbers, surnames and product names by hand. Keyterm prompting and a product glossary often improve the result more than switching vendors.
How to build customer service quality scoring criteria
Most companies already have a quality form that the manager uses to score calls. That is what you turn into an AI rubric: each criterion has to be checkable from the text, have a clear "yes / no / partly" definition and examples. A sample rubric for a customer service line:
| Criterion | Weight | How AI checks it |
|---|---|---|
| Greeting and introduction | 5% | Were the company and the agent's name mentioned in the first 15 seconds |
| Identifying the customer per procedure | 10% | Were the required details asked for, and no excessive ones (e.g. the full personal ID number) |
| Clarifying the need | 20% | Were clarifying questions asked before a solution was offered |
| First-call resolution | 25% | Was the problem solved, or was it clearly said who will do what and when |
| Accuracy | 15% | Do prices, deadlines and terms match the knowledge base |
| Required phrases and disclosures | 15% | Were the recording, return periods and contract terms mentioned |
| Closing the call | 10% | Was the call summarized, was "is there anything else I can help with" asked |
This is an example, not a standard; set the weights to match your goals. Practical tips:
- Require a quote. For every score, AI has to give the sentence and its timestamp. A score without a quote does not count.
- Calibrate. For the first two weeks the manager and AI score the same 30–50 calls; where they disagree, you fix the rubric, not the employee.
- Separate call types. Sales, complaint and tech support calls need different rubrics: AI first classifies the call, then applies the right one.
- Measure the business, not just people. The most common call reasons, recurring questions and product issues are often more valuable than individual scores.
Call sentiment analysis: what is allowed and what the EU AI Act bans
Many platforms advertise "emotion analysis". Here EU law draws a clear line that is worth knowing before you buy. Since February 2, 2025, Article 5(1)(f) of the AI Act has applied: it prohibits AI systems that infer people's emotions in the workplace, except for medical or safety purposes.
- Prohibited: analyzing agents' tone of voice, pace or intonation to infer their emotions ("irritated", "tired", "disengaged"). The European Commission's July 2025 guidelines on prohibited AI practices stress that the ban applies to the employer regardless of whether the vendor has prohibited such use in its terms.
- Not prohibited, but high-risk: emotion recognition from the customer's voice. The AI Act classifies such systems as high-risk, with additional requirements.
- Generally allowed: text analysis, i.e. whether the customer expressed dissatisfaction in the transcript, mentioned a complaint or threatened to cancel the contract. The AI Act links emotion recognition to biometric data, so "sentiment" based only on text does not fall under this ban. GDPR still applies.
In practice: when choosing a platform, switch off or do not buy features that assess employees' emotions from their voice, and calculate "sentiment" from the customer's words and the outcome of the call.
Is it legal to record and analyze calls? GDPR and consent
Recording business calls is allowed, with conditions. In a 2022 decision (in Lithuanian), the State Data Protection Inspectorate (VDAI, Lithuania's data protection authority) reminded businesses that under the Law on Electronic Communications a call may be recorded as evidence of a business transaction only after informing the other person about the recording and its purpose, and GDPR Article 13 requires providing information about the data processing. A checklist:
- Notice at the start of the call. "This call is recorded to improve service quality and as evidence of the contract. More in our privacy policy." If AI analyzes the recordings, mention it in the privacy policy.
- Legal basis. Usually legitimate interest (GDPR Art. 6(1)(f)) with a documented balancing test; in some cases, performance of a contract or a legal obligation.
- Informing employees. Employees must know in advance that calls are recorded and scored, by which criteria and who sees the results. Systematic monitoring of employees often requires a data protection impact assessment (GDPR Art. 35).
- Vendors and location. A data processing agreement (GDPR Art. 28) with every speech-to-text and LLM vendor; choose the EU region and make sure the vendor does not train models on your data.
- Data minimization. Card numbers, personal ID numbers and health data are masked in the transcript before it is sent to the LLM.
- Retention period. Keep audio recordings and transcripts only as long as the purpose requires (for example, 90 days for quality, longer only as contract evidence).
Call analytics software: ready-made platform or custom workflow?
| Lithuanian platform (WiseAudio, AINORA, 3CX Cloud LT, Neurotechnology) | Foreign platform (Aircall, CloudTalk, Gong and others) | Custom workflow (n8n + STT + LLM) | |
|---|---|---|---|
| Launch | Fast; the vendor handles integrations with 3CX and Lithuanian phone systems | Fast, if that platform is also your phone system | 1–2 weeks |
| Lithuanian | Optimized | Varies; often works best in English | You choose and test |
| Price | On request; WiseAudio offers a 14-day trial with up to 10 hours of calls (as of October 2026, wiseaudio.ai) | Per-user license + AI add-on | Setup from €450 + vendors' per-minute price |
| Flexibility | Scoring criteria: yes; integrations: depends on the vendor | Limited | Full: any rubric, CRM, reports |
| Best for | Call centers with several agents or more | Companies that already use that phone system | Companies that need results in their own CRM or helpdesk |
If you are just starting and only want to try it, begin with a Lithuanian platform's trial. If the results have to end up in Pipedrive, HubSpot, Freshdesk or your internal system and the rubric is specific, a custom workflow is often cheaper over a year. I use the same technical foundation for voice AI agents that answer calls themselves.
- 1The call is recorded3CX · Aircall · other phone systems
- 2The recording is turned into textDeepgram · OpenAI · Google
- 3AI scores it against your rubricGPT · Claude · GeminiEach score comes with a quote and a timestamp.
- 4Results and complaints go to your systemsCRM · Helpdesk · Slack
n8n connects everything, so you can swap the transcription or language model vendor without changing the whole workflow.
How much does AI call analysis cost?
An illustrative example (a calculation, not data from a specific client): 10 agents, 2,000 calls a month, 4 minutes on average, 8,000 minutes in total.
| Item | Assumption | Cost per month |
|---|---|---|
| Transcription | 8,000 min × $0.0043–0.006 | ~$35–48 |
| LLM scoring | ~4,000 input and ~500 output tokens per call, a smaller model (e.g. gpt-5-mini: $0.25 / $2 per 1M tokens) | ~$4–10 |
| Automation server (n8n) | Own server or cloud | ~€10–25 |
| Setup | One-off: integration, rubric, calibration | From €450 |
Prices exclude VAT, from vendors' October 2026 price lists. As you can see, the ongoing technology costs are low; the most expensive part is a good rubric and the integration. For comparison: a manager who listens to 20 calls a week, 6 minutes each including scoring, spends about 8 hours a month on it and still sees only 4% of calls. All my prices are on the pricing page.
Manager listens by handAutomated analysis
Share of calls scored
Figures from the example above: in ~8 hours a month the manager listens to about 4% of calls, while AI scores all of them against the same rubric.
How to roll out AI call analysis in 2 weeks: 6 steps
- Audit (1 day). Which phone system, whether recordings are available via API, where the results need to go, what the current quality form looks like.
- Legal part. Notice at the start of the call, privacy policy update, informing employees, agreements with vendors.
- Speech-to-text test. 20–30 recordings through 2–3 vendors, error comparison, building a glossary.
- Rubric and calibration. AI and the manager score the same calls, and the rubric is adjusted until agreement is good enough.
- Integration. Recording → transcription → scoring → CRM / helpdesk / Slack alert about a complaint.
- Launch and reports. A weekly team report, a monthly analysis of call reasons, a rubric review after a month.
Write to me with the phone system you use and how many calls you get a day. I will read it and reply by email on whether the recordings can be analyzed in your system. We agree on the price in writing before I start. Related reading: AI voice agent for business, AI in customer service.
Frequently asked questions
Does AI understand Lithuanian phone calls well?
In 2026, Lithuanian is supported by Deepgram Nova-3, OpenAI transcription models, Whisper, Google Chirp 3, Microsoft Azure and Lithuanian solutions (for example, Neurotechnology). Phone audio (8 kHz), dialects, numbers and surnames still cause errors, so before choosing a vendor it is worth checking transcripts of 20–30 of your own recordings by hand.
Can you record calls and analyze them with AI?
Yes, if the other person is told about the recording and its purpose before recording starts, there is a legal basis (usually legitimate interest) and the information duty under GDPR Article 13 is met. If recordings go to AI services, you need a data processing agreement with the vendor, and it is best to process the data in the EU.
Can you analyze employees’ emotions?
Not from their voice. Since February 2, 2025, Article 5(1)(f) of the EU AI Act prohibits using AI systems in the workplace that infer people’s emotions from biometric data (tone of voice, face), except for medical or safety reasons. Assessing from the call text whether an employee followed the script and offered a solution is allowed.
How much does automated call analysis cost?
The technology itself is cheap: transcribing and LLM-scoring 2,000 four-minute calls a month costs roughly $40–70 a month. Lithuanian platforms quote prices on request. A custom workflow connected to your phone system and CRM costs from €450 one-off.
How many calls do you need for AI analysis to pay off?
If a manager listens to 10–20 calls a week and the team handles more than 300 calls a month, automatically scoring 100% of calls usually pays for itself in time saved. A company with fewer calls often only needs summaries and complaint flagging, without scores.
Can AI scores be used for employee pay or dismissal?
I do not recommend doing that automatically. GDPR Article 22 restricts decisions based solely on automated processing that have a legal or similarly significant effect on a person. AI scores should be a signal for the manager and a coaching tool, and decisions about employees are made by a person who has listened to the recording.