Gravitel Launches Neural-Network Predictive Dialer: Up to 10,000 Calls per Hour

Gravitel has rolled out a predictive auto-dialing feature built on neural networks. According to a statement from company representatives passed to CNews, the system is designed to handle a flow of up to 10,000 outbound calls per hour. The initial announcement is sparse: it does not name the model architecture, pricing or customers. This article therefore combines the facts from the announcement with an explanation of how solutions of this class work.

How predictive dialing works

Predictive dialing predates the neural-network wave by a long way. The idea is simple: the system dials numbers ahead of time instead of waiting for an operator to become free. It factors in the number of available agents, the average call length and the share of people who pick up at all. If the forecast is too high, the customer answers but no live agent is on the line, and the call becomes "abandoned." If the forecast is too low, operators sit idle.

The whole economics rests on forecast error. Classic dialers run on static coefficients that an administrator tunes by hand. The neural-network approach replaces them with a model that estimates the probability of an answer from the time of day, contact history and campaign type. The announcement does not say which task Gravitel's models actually handle: only call-answer forecasting, or also a voice bot that conducts the conversation.

What is known about the system

• Feature: predictive auto-dialing built on neural networks.
• Claimed capacity: up to 10,000 calls per hour, about 167 per minute, or 2.8 per second.
• Use case: outbound campaigns in contact centers.
• Source: Gravitel representatives, in a comment to CNews.
• Not disclosed: model type and size, forecast accuracy, acceptable abandoned-call rate, CRM integrations, price.

Read the 10,000 figure as a platform ceiling, not as the result of any single customer. Real-world output depends on the number database and the number of operators. To give a sense of scale: at a hypothetical answer rate of 25%, that flow yields about 2,500 conversations per hour. Handling them takes a contact center with hundreds of seats, or a bot that takes over the first part of the dialogue. This is an illustrative calculation, not company data.

Where the line of what is permitted lies

Mass dialing runs into limits that go beyond technology. In Russia, advertising calls are permitted only with the subscriber's consent (Article 18 of the Law "On Advertising"), and the number database must be compiled in line with the personal data law. Telecom operators, for their part, are rolling out anti-spam filters and may block numbers with abnormally high outbound activity. A platform capable of producing 10,000 calls per hour can quickly fall under such filters, along with its customer, if configured carelessly.

That makes the question of restriction settings as important as dialing speed. A good predictive dialer can keep the abandoned-call rate within set limits, respect calling windows across time zones and exclude numbers on the opt-out list.

What to check before deployment

If you are considering such a solution for sales, debt collection or surveys, start with a pilot on part of your database rather than the full volume. Compare it with your current mode on four metrics: answer rate, abandoned-call rate, operator utilization and cost per successful contact.

Ask the provider separately how the model is trained: on your data or on the platform's aggregated data, and where that data is stored. Find out how the system behaves when the answer rate drops, for example during holidays or when operators block numbers. And check whether the call result is passed to the CRM automatically: without that, the gain from dialing speed is quickly lost to manual processing.

Until Gravitel publishes more details, it is reasonable to treat the figure of 10,000 calls per hour as a claim of a serious tool, but not as a proven result.