AI agents
Machine learning for contact center tasks
Machine learning in CONTACT.WORK means models that help your contact center spend calls and agent time where they pay off: they suggest who to call first and at what hour, screen out answering machines, and identify the topic of an inquiry right away. We measure the gain in your own metrics – contact rate, first-call sales, agent time – and compare it with working without the model on identical lists.
Connect ratemore conversationsrisingSalesfor the same callsupEmpty dialsfewer minutesdownReviewfaster to the teamfasterMeasurable results
How we measure the gain
We agree on the metrics with you up front and sign off on the model based on them.
Tasks
Where models pay off
We start with the task where the gain shows fastest – usually outbound calling.
Customers most likely to buy come first
The model scores each lead based on the history of similar customers and moves those most likely to say yes to the top of the list. Agents make the same number of calls per shift, but more of them end in a deal.
- A score for every lead
- Dialing order
- More deals from the same calls
First 100high purchase oddsat startMiddle300 numbersin queueLast 80deals unlikelylastShift resultmore salessame number of callsCalling when they'll answer
For each customer, the model picks the time they most often answer and moves the attempt to that hour, within the calling hours you set. Fewer empty dials, more connects, fewer repeat attempts.
- A time for each customer
- Rescheduled attempts
- Fewer repeats
Emily K.weekdays, 6-7 PManswers most oftenVector LLCTue-Thu, 10 AM-12 PManswers most oftenPeter S.weekends, afternoonanswers most oftenAttemptmoved to 6:15 PMin queueAgents hear only people
During an outbound campaign, the model detects who's on the line: a person, an answering machine, a bot, or no one. Only people reach agents, so the shift goes to conversations with customers, not voicemail greetings.
- Live answer – to an agent
- Answering machine – screened out
- Silence – retry later
Live answer1,240 callsto agentAnswering machine610 callsfiltered outSilence95 callsretry laterAgentslive conversations onlyno dead airStraight to the people who can resolve it
A call or email gets a topic as soon as it arrives and goes to the team that handles those issues. No manual sorting and no passing inquiries between departments.
- Topic right away
- The right team
- No manual sorting
ChannelemailinboundTopicproduct returnby modelGroupreturns teamscheduledStafftopic correctconfirmedThe same approaches work in speech analytics and inside AI agents, and we can build a custom model for your specific task.
How it works
Why you can trust the gain
Data
Your own history, not a generic one
We train each model on a specific company's calls and deals: what helps one sales team may not work for another.
- Call recordings and outcomes
- Deals and customers
- Inquiries and their topics
Telephonycalls and outcomesAsteriskCRMdeals and customersvia APIInquiriestopics and resolutionsplatformRecordingscall transcriptsoptionalTesting
We prove the gain first
The model learns on one part of the history and is tested on another part set aside in advance. Then it runs on part of your outbound calls alongside your usual order. The model stays only if your metrics are better with it.
- Held-out history
- Pilot run
- Only if it's better
Holdout history20%unseen by modelPilot runpart of campaignalongside standardComparisonvs. standard orderequal listsModel staysif results are betterruleHosting
Cloud or your network
Models run wherever the platform runs: our cloud or your own servers. In the second case, call recordings never leave your company.
- Cloud
- Your network
- Recordings stay in-house
CloudCONTACT.WORKoption 1Your serveralongside the platformoption 2Call recordingsstay with youwith option 2Choicecustomer's choiceyou decideLaunch
How we work
- 01TaskWe pick what to improve and the metric we'll use to compare the result with the baseline.
- 02DataWe collect your call, deal, and inquiry history and check that nothing is missing.
- 03First modelWe train the model and check its predictions against the held-out part of the history.
- 04Pilot runWe turn on the model for part of the lists or queue and see where results are better.
- 05RolloutWe roll the model out across all work and show the results in the platform and reports.
- 06Ongoing supportWe track metrics and retrain the model when accuracy drops.
Requirements
What it takes to launch a model
- History
- Calls, deals, and inquiries from past months
- CRM
- Deals and call outcomes via API
- Telephony
- Asterisk connected to the platform
- Hosting
- Our cloud or your own hardware
- Oversight
- An employee who flags model errors
- Cost
- Development and training billed in support plan hours
Preparation
What to discuss in advance
Q&A
Frequently asked questions
How is machine learning different from training an AI agent?
An AI agent talks to customers or employees. A machine learning model doesn't talk – it calculates: who to call first, when to call, what an inquiry is about, and whether the call reached a person or an answering machine.
What other tasks do models handle?
Besides outbound calling and inquiry topics: speech analytics and AI agent tasks such as speech recognition and picking a response. If the model you need doesn't exist, we build one for your data and processes.
What data do you need?
Calls with outcomes, deals, and inquiries with topics. The volume depends on the task – we'll tell you once we see your data.
How do models work with the auto dialer?
The auto dialer places the calls, and the models tell it the order of numbers, when to try, and which calls to connect to an agent. Answering machine detection is built right into the auto dialer.
How often do models need retraining?
It depends on how quickly your products and customers change. Usually we review metrics four times a year, sometimes monthly.
What if the model gets it wrong?
An employee flags the error, such as a misidentified topic or customer phrase. The flag triggers retraining, and the next model version accounts for that case.
Do we need our own team of analysts?
No. We build and maintain the model; from you, we need data and an employee who knows the process well and can spot the model's errors.
Can we keep the models in-house?
You can. Models are installed on your company's infrastructure along with the platform, and call recordings stay with you.
Who owns the trained model?
The customer who paid for the hours to develop and train it.
Where does the model get deal data?
From your CRM via API. It sends results back there or shows them in the platform.
How much does machine learning cost?
Model development, training, and support are billed in support plan hours. We give a preliminary estimate after reviewing your data.
Glossary
Terms on this page
- Cloud deployment (cloud)
- The platform runs on the vendor's servers, and agents connect over the internet. Learn more →
- Call recording
- Calls are saved so they can be replayed and disputes settled. Learn more →
- CRM – customer relationship management
- Where a company keeps its customers, deals, and full communication history. Learn more →
- Speech analytics
- The system analyzes recorded calls automatically: topics, script adherence, customer emotions. Learn more →
- Hosting
- You rent a server from a provider instead of buying one. Learn more →
- API (application programming interface)
- The way programs exchange data with each other without a person in the loop. Learn more →
- Answering machine detection (AMD)
- Tells whether a person or an answering machine picked up an outbound call. Learn more →
- Voicemail
- A recorded message when the person does not pick up. Learn more →
- On-premises deployment
- The system runs on the company's own servers, not in the vendor's cloud. Learn more →
- Speech-to-text
- Turns spoken conversation into text. Learn more →
Updated: