AI call centers and core systems: automating after-call work
AI answers the phone. It transcribes the call and summarizes it. Call centers have taken AI this far. Yet the after-call work of entering what was agreed into the core systems for customers, orders, or repairs is often still done by hand. When a core system has no API, there is no way in for the results AI produces.
This article covers how to connect an AI call center to core systems with no API, and automate the entry that follows each call.
The short answer: separate what happens during the call from what happens after it
- During the call, leave it to AI phone agents and speech recognition services
- After the call, automate entry with something that receives the call’s results and enters them in the core system
- If the core system has no API or import function, enter the data by operating its screens
- Before anything hard to undo, such as an address change or a cancellation, a person checks
- Calls contain personal information, so decide first where the automation runs and which AI model providers the data goes to
AI may answer, but the entry remains
Gartner predicts that by 2029, agentic AI will autonomously resolve 80% of common customer service issues without human intervention, cutting operational costs by 30%. It explains that while earlier AI was limited to generating text or summarizing interactions, agentic AI can act on its own to complete tasks.
Resolving an issue means finishing the work, such as changing an address or arranging a repair. For that, the results AI produces have to reach the core system.
Gartner also predicts that by 2027, 50% of organizations that expected to significantly reduce their customer service workforce will abandon those plans. In a March 2025 poll of 163 customer service leaders, 95% planned to keep human agents. AI and people will be answering side by side for some time.
Either way, the outcome of each call has to go into the core system. If the core system has an API, connecting it is not hard. If not, that entry stays manual (Integrating AI with legacy systems).
Where manual work remains in a call center
| Stage | What AI can do | What remains without a connection to the core system |
|---|---|---|
| During the call | First-line answers by an AI phone agent, suggested answers for operators | Looking up customer and contract details on another screen |
| Right after the call | Transcription, summaries, sorting by request | Reading the summary and re-entering it on core system screens |
| After-call work | Listing the arrangements needed next | Entering address changes, repair requests, and callbacks in each system |
This article covers the stages right after the call and after-call work. Looking up the core system to answer during a call is covered below, under “Not for lookups during the call”.
Four ways to receive call results
There are four ways to receive call results from AI phone and call recording services.
| How you receive them | Suits | Watch out for |
|---|---|---|
| Webhook | Services that can send the transcript or summary the moment a call ends | Each service has its own rules for how the receiver must answer |
| API (you fetch) | Services that let you fetch call records through an API | Delayed by the polling interval |
| Services that can email call summaries | The message format may change | |
| Files (CSV) | Call systems on your own premises that can export call records as CSV | Delayed by the export interval |
More services now send a webhook when a call ends. ElevenLabs, for example, sends data including the transcript and analysis results to a URL you specify once a call’s analysis is complete. MiiTel, a call analytics service in Japan, can send call history to a URL you specify, for instance when transcription finishes.
The receiving side has to follow each service’s rules. ElevenLabs treats only a 200 status code as success, and disables a webhook that keeps failing. MiiTel sends a verification token when you register the destination, and expects the receiver to return it as is. Check that whatever receives the webhook meets these rules.
Example: entering an address change after the call
Here is the work of taking address change calls and entering them in a customer management system with no API, split step by step.
| Step | Who does it | How it connects |
|---|---|---|
| Receive the transcript and summary when the call ends | Fixed steps | Webhook or email |
| Pick out the customer number, name, new address, and requested date | AI | None |
| Look up the customer in the customer management system and match the name and phone number | Fixed steps | Reads from the screen |
| Check requests that don’t match, and those where identity was not verified | A person | None |
| Enter the new address | Fixed steps (AI only when the screen changes) | Enters it on screen |
| Record the reference number and report failures | Fixed steps | A list and alerts |
AI handles only picking out the fields to enter from a transcript of spoken words, and what to do when a screen changes.
Putting a human check on every request only swaps after-call work for checking work. Enter requests that matched as they are, and send to a person only those that don’t match, those where identity was not verified, and those hard to undo, such as cancellations. Decide how much passes through automatically for each process, based on the impact of a mistake.
Not for lookups during the call
Looking things up by operating the core system’s screens takes from a few seconds to tens of seconds per item. That is too slow for AI to answer while looking at customer details during a call.
If you need lookups during the call, use the core system’s API, read its database, or use a customer list exported in advance. Use screen operation for entry after the call and for work done before calling back.
Volume and security: run it on dedicated VMs
After-call work in a call center comes in volume and contains personal information. Decide where it runs from these two.
- Volume: screen operation proceeds one item at a time on each computer. Estimate how many computers you need from the time per item, the number of items a day, and the peak hours
- Personal information: when AI reads a transcript or a screen, the content is sent to the provider of the AI model. Check the provider’s terms on whether it is used for training, and how long it is kept
- Network: if the core system is on the company network or in a closed network, put the computer that operates it on the same network
What to check:
- The number of calls a day, and the peak hours
- Whether transcripts contain information with handling rules of its own, such as card numbers
- Whether call content may be sent to external AI models
- Which network the VMs that operate the core system will sit on
Frequently asked questions
Can AI answer while looking at the core system during a call?
Not by operating screens; it is too slow. For lookups during a call, use the core system’s API or read its database. Use screen operation for entry after the call.
Which AI phone or call recording services can this work with?
Any service that can send call results out by webhook, API, email, or file. Each service has its own rules for receiving webhooks, and some cannot be received as they are, so the PoC settles which way to receive them.
Will operators’ jobs disappear?
No. In Gartner’s poll, too, 95% of customer service leaders plan to keep human agents. With after-call work automated, operators can spend their time checking requests that didn’t match, and on inquiries a person should handle.
Connecting with Kitewell
Descarty’s Kitewell is AI RPA that AI builds and AI fixes. Receiving call results, picking out fields with AI, and entering them in core systems with no API all run in one workflow.
- Receive call results: a webhook (on paid plans) or a REST API call from inside your network starts a run. Email (Gmail, Microsoft 365, IMAP) and CSV files are checked and read at set intervals. Webhooks arrive through Kitewell’s relay, so no port is opened. In a closed network, your call system starts runs through the REST API
- Pick out fields: questions to AI models and judgments become steps in the workflow. You choose the model for each step, including in-house and local models
- Enter data in the core system: AI reads the screen and operates customer management and order systems with no API, on the web and in desktop apps. Operations that worked are remembered and repeated without AI while the screen stays the same
- Send only what needs checking to a person: approval can be set on any step, and pending approvals are sent to Slack, Microsoft Teams, and email. Send one back, and the run restarts from the step you choose
- Alerts when it stops: failures are sent to Slack, Microsoft Teams, email, and more. With projects synced through Dagu Cloud, you are also alerted when a computer you chose stops reporting for 10 minutes
- Passwords stay out of AI’s hands: core system passwords are referenced by name, and Kitewell types them in without showing the value to the AI
- Everything on record: each run keeps its logs, screenshots, and approval decisions
Under an enterprise plan, the computers that run workflows can be VMs or dedicated PCs in your own cloud, on your premises, or in a closed network. Licensing, team sync, and updates can all stay in your environment, and with in-house or local AI models nothing is sent outside. AI models that can read and operate a screen are limited, though, and small local models usually cannot operate one. Discuss which model to use in a closed network, and how many VMs your volume needs, at the PoC stage.
You can start by choosing one type of call that still leaves entry work afterward, and testing it in a two-week PoC, quoted individually. If the agreed success criteria are not met, you pay nothing. See Pricing & PoC for details.
Summary
- Call center AI can answer and summarize, but when core systems have no API, the entry after each call stays manual
- Leave the call itself to AI phone agents and speech recognition, and automate the entry after it
- Receive call results by webhook, API, email, or file. Each service has its own rules for webhooks
- Send to a person only what doesn’t match and what is hard to undo
- Decide how many dedicated VMs you need, and where they sit, from the volume and the handling of personal information
Sources
- Gartner: Gartner Predicts Agentic AI Will Autonomously Resolve 80% of Common Customer Service Issues Without Human Intervention by 2029 (March 5, 2025)
- Gartner: Gartner Predicts 50% of Organizations Will Abandon Plans to Reduce Customer Service Workforce Due to AI (June 10, 2025)
- ElevenLabs: Post-call webhooks
- MiiTel: Outgoing Webhook: call history specification and usage (in Japanese)
Prove it on your own work in a two-week PoC
We run Kitewell workflows on the same screens and data as production, and measure processing time, success rate, and how easily a stopped run is fixed.
Choose one robot that stops when screens or systems change, or one task done by hand, and we run it in Kitewell.
What we need from you
- One contact person for the work being automated
- Test accounts and access to the screens involved
- Sample data close to production
- A machine to run Kitewell
- An API key for the AI model you will use, or an in-house model server
Quote
- If the agreed success criteria are not met, you pay nothing
- Agree on the task and success criteria in the first interview
- If you go on to a contract, the fee is credited against the first year
- A results report: items processed, time per item, AI usage cost, and how often and why runs stopped
- Kitewell is in public beta (0.x); features and screens may change