01Overview
SavvyTalk is an AI call analytics platform built for call centers. It collects your agents' call recordings, transcribes them, scores them against the quality criteria you define, and makes the results available in dashboards, reports and exports.
Where a traditional quality team listens to a small sample of calls, SavvyTalk analyzes every call with the same scorecard and backs every verdict with a quote from the conversation.
For supervisors
A complete view of quality, operation by operation and agent by agent, without re-listening to calls.
For quality managers
Scorecards written in plain language, applied consistently to every call.
For trainers
Progress reports that list each agent's recurring weaknesses and the actions to work on.
For management
Reliable indicators (volumes, valid-call rates, call reasons) and Excel exports for your own analysis.
02Key concepts
| Term | Definition |
|---|---|
| Operation | A campaign or activity in your call center (for example an outbound sales campaign). It holds the evaluation criteria, the report instructions, optional web search instructions and the list of enabled metrics. Every call belongs to an operation. |
| Call / transcription | An audio recording received by SavvyTalk, with its transcript, analysis, status and metric values. It also carries its context: record ID, agent, phone number, duration. |
| Criteria | The operation's quality scorecard, written in plain language. The AI evaluates each criterion, in order, without skipping any. |
| Instructions | The structure of the analysis report: sections, order and expected format (for example a "Recommendations" section at the end). |
| Status | A binary qualification of the call, VALID or UNVALID, extracted automatically from the analysis. It is typically used to qualify leads or opportunities. |
| Custom metric | An indicator you define (Yes/No, Number, or a value from a list), evaluated by the AI on every call of the operations where it is enabled. |
| Agent | The person who made or took the call, identified by a code and a name sent with the recording. |
| Credits | SavvyTalk's unit of usage. Each processing step uses a number of credits known in advance. |
03Quick start
- Sign in to your SavvyTalk workspace with the credentials provided at onboarding.
- Create an operation from Operations › New operation: give it a name, write your criteria and instructions, or start from a template. MrSavvy can help you write them.
- Create your metrics under Metrics and enable them on the operation.
- Send calls: a few files from the web app to test, then switch on automatic collection with SavvyAgent or the API.
- Review the results in each transcription's details, then on the operation dashboard and the Agents page.
- Refine your criteria if needed, then re-run the analysis on existing calls to compare.
04Call collection
SavvyTalk accepts calls three ways. All of them lead to the same processing.
SavvyAgent (automatic collection)
SavvyAgent is a Windows program that runs continuously on a computer in your call center. For each configured operation:
- SQL polling: at regular intervals it runs a query on your call database. Each row describes a call: ID, recording path, agent, phone number.
- Queue: each new call is added to a local queue. The same call is never added twice.
- Check: before sending, the agent asks SavvyTalk whether the operation is active. If it is disabled, the call waits.
- Download: it fetches the audio file(s) from your recording server and merges them if needed.
- Upload: the file and its details are sent to the SavvyTalk collector.
- Retry: on error, the upload is retried automatically.
A test mode (--dryrun) lets you check the query and file paths without sending anything. All activity is written to a readable local log.
Collection API
Your own system can send calls directly. Each request is authenticated with a token (Authorization: Bearer …) issued by our team.
| Method | Path | Purpose |
|---|---|---|
GET | /v1/operations/{id}/enabled | Tells whether the operation accepts new calls. |
POST | /v1/operations/{id}/upload | Sends a WAV recording and its details; analysis starts immediately. |
Fields sent with the file (multipart):
audio_file the call's .wav file
record_id the call's ID in your system
agent_id the agent's code
agent_name the agent's name
phone_num the customer's phone number
The collector address and your token are provided at onboarding.
Manual upload from the web app
From an operation's page you can upload audio files directly. Processing runs in the background, so you can leave the page.
- Each upload's progress is shown and refreshed every 3 seconds (queued, transcribing, analyzing, done).
- An upload in progress can be cancelled.
- Formats other than WAV are converted automatically.
05Transcription
Each recording is converted to a standard audio format (mono, 16 kHz) if needed, then transcribed by an AI speech recognition model. The actual call duration is measured directly from the audio file.
- Long calls: beyond 20 minutes, the audio is split into 20-minute segments that overlap by a few seconds. Segments are transcribed in parallel and then stitched back together without duplicates.
- Result: the full text of the conversation, viewable in the call details and exportable.
- Reliability: if a step fails, the call is rolled back cleanly: no partial result is kept and no credits are used.
06Operations and templates
The operation is the heart of the setup. It defines how its calls are evaluated.
| Field | Use |
|---|---|
| Name, description, client | Identifies the campaign in the app. |
| Evaluation criteria | The quality scorecard, as formatted text. Example: "Introduce yourself and state the reason for the call", "Confirm the decision-maker". |
| Instructions | The structure of the analysis report: sections, order, level of detail. |
| Web search instructions | Optional: what the AI should check online for each call (see Web search). |
| Enabled metrics | The custom metrics computed for this operation. |
| Active / inactive | An inactive operation rejects new calls; SavvyAgent keeps them queued. |
Operation templates
Save a complete setup (criteria, instructions, client, web search instructions) as a template. A template can stay private or be made public for other users. When creating an operation, pick a template to prefill every field. Unused templates can be deleted in bulk.
07AI quality analysis
Once the transcript is ready, the AI acts as a rigorous call center evaluator. It receives your instructions, criteria and the transcript, and produces a full report.
Rules the analysis follows
- The report follows the structure of your instructions (sections, order, format).
- Every criterion is evaluated, in the given order, without skipping any.
- For each criterion: a clear verdict, a detailed reason and at least one exact quote from the transcript.
- No invention: when information is missing, the report says so ("Not mentioned", "Absent from the transcript").
- No introduction or filler: the report goes straight to the evaluation.
Call status
A second step reads the analysis and assigns a VALID or UNVALID status. It feeds the valid-call rate on dashboards and can be used to qualify your leads.
Re-analysis
After editing your criteria or instructions, you can re-run the analysis on calls that are already transcribed, without re-sending the audio.
08Custom metrics
Metrics turn the analysis into comparable figures. Each metric has a name, a description, a type, an optional constraint and an instruction for the AI.
| Type | Constraint | AI answer | Example |
|---|---|---|---|
Yes / No (YesNo) | — | Only "Yes" or "No" | Was the decision-maker confirmed? |
Number (Number) | Range <min..max> | Only a number within the range | Courtesy score from 0 to 10 |
List (String) | Possible values (A, B, C) | Exactly one of the values | Reason: Sign-up, Information, Cancellation… |
- Create the metric under Metrics › New metric.
- Write a precise instruction: what the AI should check and how to decide.
- Enable it on one or more operations.
The metric is then computed on every new call of those operations. It uses the transcript, the analysis and, when enabled, the web search. Fixed answer formats guarantee values you can use in charts and exports.
09Web search
Some criteria need an outside check: an address in a covered area, a company that exists, public information quoted by the customer. Add web search instructions to the operation and, for each call, the AI runs the search and produces:
- a one-paragraph summary, ready to use;
- the list of sources consulted (up to 5).
The result is stored with the call and can be used by your metrics. If the main provider is unavailable, a backup provider takes over automatically.
10Dashboards
Home
An overview: total calls analyzed, active operations, configured metrics, and a summary of each operation (calls, active metrics).
Operation details
- Key indicators: call volumes, valid calls, trends over time.
- One chart per custom metric: Yes/No split, score distribution, breakdown of list values.
- Tracking of uploads in progress and completed.
- Aggregates are recomputed automatically several times a day (per operation and per agent).
11Transcriptions
The Transcriptions page lists every call you have access to.
- Filters: operation, file name, record ID, period (from / to), rows per page.
- Call details: context (agent, number, duration, record ID, operation), full transcript, analysis, status, metric values and web search result.
- Comfortable reading: any text block opens full size for long analyses.
- Excel export straight from the list (see Exports).
12Agent tracking
The Agents page gathers every agent's results.
- Two views: cards showing each agent's key metrics, or a full table.
- Sort on any column (calls, valid rate, metrics) and instant search by name or code.
- Agent page: all their metrics in detail and direct access to their reports.
- CSV export of the list as displayed, Excel-compatible.
13Progress reports
The progress report summarizes, for one agent and one period, the weaknesses that come up most often in their calls, with concrete actions to work on. It replaces reading dozens of analyses before a coaching session.
Generate a report
- Open Agents, then the Reports tab.
- Choose the agent (search by name or code, ignoring accents and case).
- Choose the period: 30 days, 90 days, 12 months or custom dates.
- Click Check: this free step shows the number of usable calls and the price.
- If the agent has more than 100 usable calls, choose how many to analyze (the most recent, up to 500).
- Click Generate. The report appears within seconds.
Reading the report
| Part | Content |
|---|---|
| Header | Period, number of calls analyzed, calls without a usable section, generation date, credits charged. |
| Summary | The main area to improve, in two sentences. |
| Priority actions | The 3 to 5 actions to work on first. |
| Weaknesses by theme | Each weakness with its finding and actions, from most to least frequent. |
| Isolated observations | The number of findings too rare to form a theme. |
"16 calls out of 35": 35 is the number of calls analyzed, 16 the number of calls concerned by the theme. A call counts once per theme, but can belong to several themes.
- The report is built from the analyses' recommendations, not from the transcripts.
- Only fully analyzed calls count; at least 5 are needed to draw conclusions. A report based on fewer than 5 calls is flagged as "weak".
- Reports are saved in the History tab, filterable by agent, operation, sample size, generation date and analyzed period. Reopening them is free.
- The same calls produce the same report: SavvyTalk then offers to reopen the existing one at no cost.
The price depends on the number of calls analyzed and is shown before you confirm (see Credits).
14MrSavvy, the AI assistant
MrSavvy is the assistant built into the operation create and edit pages. It helps you write and improve your criteria and instructions.
- Ask in plain language: "Add a criterion about restating the need", "Make the criteria more measurable", "Adapt this scorecard to a collections campaign".
- It shows its suggestions as diffs: additions in green, deletions struck through in red, with a full-size view for long texts.
- You apply or dismiss each suggestion, and a button lets you undo the last change.
- The conversation history is kept for each operation.
15Exports
Excel export of transcriptions
- From Transcriptions, click Export to Excel.
- Choose the operations and the period.
- Tick the data to export: date, operation, file, record ID, phone, agent, duration, size, status, transcript, analysis, each of your custom metrics and the call context fields.
- Check the preview (number of matching rows), then download the
.xlsxfile.
An export can contain up to 5,000 calls; beyond that, narrow the period.
CSV export of agents
On the Agents page, the Export CSV button downloads the list as displayed (filters and sorting included).
16Users and security
| Role | Access |
|---|---|
| User | Their own operations and metrics, and the calls, agents and reports of those operations. |
| Super-administrator | All data, user management, the audit log and bulk deletion. |
- Sessions: automatic sign-out after 30 minutes of inactivity.
- Audit log: logins, report generation, deletions and configuration changes are recorded in a monthly log.
- Authenticated collection: every call upload must present a valid token; tokens can be disabled at any time.
- Operations you can switch off: a disabled operation accepts no calls.
- Hosting: the platform is cloud-hosted and served over HTTPS.
17Credits and billing
SavvyTalk works with credits. The balance is checked before each processing step, and credits are only charged once processing succeeds.
Call analysis
| Item | Credits |
|---|---|
| Call up to 10 minutes | 10 |
| Each additional 5 minutes | +5 |
| Each custom metric evaluated | +1 |
| Web search | +4 |
Example: a 12-minute call with 4 metrics and web search costs 10 + 5 + 4 + 4 = 23 credits.
Progress report
| Calls analyzed | Credits |
|---|---|
| Up to 100 | 25 |
| Each additional 50 calls | +10 |
| 451 to 500 | 100 (maximum) |
Free of charge: checking the price, reopening a saved report, a report with no weakness found, and a generation that fails.
18Architecture and AI
SavvyTalk is not a single call to an AI model: it is a structured processing pipeline where each step has a precise role and a controlled output format.
- Collection: authenticated intake of recordings.
- Transcription: speech recognition, with long calls split into segments.
- Analysis: a full report based on your criteria and instructions.
- Status: VALID / UNVALID extraction.
- Web search (optional): outside verification with sources.
- Metrics: one fixed-format evaluation per metric.
Multi-model: the analysis layer can use several leading AI providers. We pick the best-suited model for each step, and the architecture lets us switch without service interruption.
Asynchronous processing: uploads are processed in the background, with progress tracking and cancellation.
19FAQ
Which audio formats are accepted?
The collection API expects WAV. From the web app, other common formats are converted automatically.
How long does it take to analyze a call?
From a few tens of seconds to a few minutes, depending on the call's length and the number of metrics. Upload tracking shows progress live.
Can we change a scorecard after launch?
Yes. New calls use the new scorecard, and you can re-run the analysis on existing calls.
How do we get support?
Our team installs SavvyAgent with you, sets up your first operations and trains your staff. Contact us or email savvy@123.tn.
No results. Try another word.