AI for call centers and BPOs: voice, QA and agent assist.

02What can AI do for an operator that runs queues for many brands?

Four jobs repeat across nearly every operation. Forecasting is on the list because people assume a language model handles it. It doesn’t.

03How do you keep one client’s data apart from another’s?

An operator is usually a processor of personal data for each client. The EDPB’s guidelines on controllers and processors say a processor must not process data except on the controller’s instructions, and one that starts setting its own purposes becomes a controller for that processing. That argues against the tempting shortcut of pooling every client’s calls to improve one shared model or one shared knowledge base. Check each contract, and design for separation from the first day:

• A tenant ID on every object, enforced below the model. Recordings, transcripts, embeddings, prompts, scorecards, eval sets, summaries and logs all carry the client ID. Retrieval filters on it in the data layer, through separate indexes or enforced row-level filters. Telling the model to “only use Client A’s documents” is not a control.
• Per-client configuration, not per-client code. Scorecard, disposition codes, knowledge base, escalation rules, retention and storage region live in a tenant configuration, so adding a client is a configuration job.
• Keys, retention and deletion per client. Separate encryption keys, retention set by each contract, and an exit process that deletes transcripts, embeddings and caches, and records that it did. The UAE’s data protection law, for one, requires a processor to erase data after the processing period or on handing it to the controller.
• Agents who work for several brands. The assist panel must switch tenant when the dialed number or campaign changes, and drop the previous client’s context. Bleed between brands is a failure your floor staff will notice quickly.
• Sub-processors and regions. The speech-to-text, language model and voice vendors are sub-processors, and each client may need to approve them and the region they run in. Saudi Arabia’s transfer regulation (version 2.0, August 2024), for example, requires the controller, here your client, to run a risk assessment before certain transfers outside the Kingdom.
• A leak test that runs on every release. A suite that signs in as Client A and tries to retrieve, summarize or score Client B’s material, including through shared caches and hostile prompts. The pass mark is zero leaks.

04Which systems does it need to connect to?

An operator’s stack has two halves. Contact center platforms such as Genesys Cloud, NICE CXone, Five9, Amazon Connect, Twilio Flex, Avaya and Cisco hold routing, dialers, recordings and the agent desktop. Each client’s own systems, such as Salesforce Service Cloud, Zendesk, ServiceNow, Freshdesk and the order or billing system behind them, hold the records agents read and edit.

Audio usually arrives one of two ways: a recording fork where the platform or session border controller supports SIPREC (RFC 7866, which streams media and call metadata to a recording server), or the platform’s own media-streaming API. Because clients bring their own platforms, most operators end up running several. We put an adapter layer in front, so every call becomes one internal record (call ID, tenant, speaker-separated audio, metadata, disposition) and the QA and assist logic is written once. Integration with the client’s CRM and an older platform is usually the largest piece of work, not the AI.

05How does QA on every call work, and can you trust the scores?

Sampled QA reviews a slice of calls and hopes it is typical. Scoring every call shows patterns by client, queue, shift and new hire, but only if the scores are right. The pipeline we’d build:

1. Transcribe with speakers separated. Use dual-channel audio where the platform offers it, because separating two voices from a mixed recording afterward is harder and adds errors.
2. Redact before any LLM sees it. The PCI Security Standards Council’s telephone payment guidance (a 2018 supplement, still the current version) says it is trivial to convert audio into queryable data. A transcript of every call does that for you, so spoken card numbers become searchable text. Vendor redaction helps but isn’t a guarantee: Amazon Connect’s documentation on sensitive data redaction says it may not remove every instance, and for voice it runs after the call ends. Keep card data out of the recording in the first place (see the payment row in the voice table below).
3. Put must-pass items in code. Was the recording notice read? Was identity verified before account details were shared? Was a forbidden phrase used? These are pattern checks on the transcript, and they should not be left to a model’s judgment.
4. Use the LLM for judgment items, with evidence. “Did the agent acknowledge the problem?” gets a score and a timestamped quote. A score with no quote is not shown.
5. Calibrate per client. Human analysts score a set of calls on the client’s form. Measure agreement question by question, and keep any question below your threshold human-scored. Re-run this whenever the model, prompt or scorecard changes (our note on evals explains the method).
6. Let agents dispute. Low-confidence and disputed calls go to a person, and agents see the evidence behind each score.

Test the speech layer on your own audio. A peer-reviewed study of five commercial speech recognition systems found an average word error rate of 0.35 for Black speakers against 0.19 for white speakers. It dates from 2020 and the systems have changed, but the lesson stands: measure error rates by accent, language, line quality and code-switching (Arabic and English, or Spanish and English), and by each client’s product names. Otherwise QA quietly penalizes the agents the transcriber understands least.

06What does agent assist do, and where does it go wrong?

Live assist reads the transcript as it forms. It surfaces the right article from the client’s knowledge base, prompts required steps (a disclosure, a verification question) and drafts the after-call summary and disposition. It fails in predictable ways:

• Wrong brand. The multi-brand agent gets Client A’s refund policy on a Client B call. The tenant switch above is the fix, and the leak test should include this case.
• Confident invention. Assist must cite the article it used and say “nothing on file” when retrieval finds nothing. That is the behavior described in our guide to retrieval that admits when it doesn’t know.
• Lag and unchecked summaries. A suggestion that arrives after the customer has moved on is noise, so measure the delay. A wrong summary written straight into the client’s CRM becomes the client’s record, so write summaries as drafts the agent confirms.

07Which calls can a voice agent take, and which should it never take?

The plumbing is well understood: speech-to-text feeds your application, a model decides what to say, and text-to-speech plays it back. Twilio’s ConversationRelay documentation shows the shape, including a WebSocket to your own backend, interruption settings, keypad (DTMF) events and a warning that not every speech vendor is guaranteed PCI compliant. The hard part is deciding what the agent may do:

Three engineering rules matter more than the voice. Limits sit in code, not in the prompt: a refund cap, a list of actions the agent may take, and identity checks before any account lookup. Handoff is a feature: after two misunderstandings, an angry caller or a request for a person, the call goes to an agent with the transcript and the caller’s intent already on screen. And stream the reply: Twilio’s best practices advise sending model tokens as they arrive, to shorten the silence a caller hears.

08Containment or customer experience: which number should you manage to?

Containment is the share of calls the bot ends without a transfer. It’s easy to measure and easy to game, because a caller who gives up and hangs up counts as contained. Manage to contained and resolved instead: no repeat contact about the same issue inside a window agreed with the client. Track abandon-in-bot rate, time to reach a person, and survey scores for bot-handled calls against person-handled ones.

Set targets per call type and per client. One containment target across all clients pushes the bot to hold on to calls it should pass on. Gartner’s poll of 163 service leaders found 95% plan to keep human agents, so design for a bot-first service where a person is always reachable.

09How does the operator’s commercial model change?

Much operator revenue is still tied to agent time. Automation that cuts handle time, after-call work and call volume therefore cuts billed hours, so quality improvements shrink revenue unless the contract changes with them. Buyers are already moving: Forrester’s July 2026 analysis says clients are buying outcomes, not hours of effort. An HFS Research survey of more than 500 executives (the report was produced in partnership with a BPO provider) expects reliance on FTE-based models to fall from 42% to 28% over three years, and on outcome-based models to rise from 20% to 39%.

An outcome contract only works if the engineering can back it:

• An agreed definition of “resolved,” and a log that can prove it for every call.
• An audit trail per call: what the AI said, which tools it called, and which prompt and model version were live.
• Clear ownership of AI errors, such as a wrong refund the voice agent granted. Limits in code and wording in the contract should match.
• Data terms: no training on a client’s calls without permission, and per-client reporting drawn only from that client’s tenant.

Gartner’s October 2025 survey of 321 service leaders found only 20% had cut agent headcount because of AI, while 42% were hiring specialized AI roles. Plan for a changed workforce, not an empty floor.

10Where AI is the wrong tool, and where a person stays in charge

• Forecasting and scheduling. Time-series and optimization models do this better. An LLM can explain a forecast to a planner, not make one.
• Routing. Skills, priority and language are rules and classic models. Use the LLM only to read the caller’s intent.
• Required wording and card data. Deterministic matching checks that a disclosure was read. Pause-and-resume or DTMF masking, both covered in the PCI guidance above, keeps card numbers out of recordings.
• Decisions about people. A QA score is evidence for a conversation with an agent, not a verdict. A person decides coaching, pay and dismissals.

Score what was said, never how the agent is feeling. The first is quality control; the second risks falling under an outright ban in the EU.

11Which rules apply? (as of October 2026)

This describes what the rules say. It is not legal advice, and a qualified professional should confirm what applies to each client and market.

• US, AI outbound calls. The FCC’s February 2024 declaratory ruling confirms the TCPA’s “artificial or prerecorded voice” rules cover AI that generates human voices, so such calls need the called party’s prior express consent unless an emergency or exemption applies. For outbound campaigns, check the client’s consent records before an AI voice dials.
• US, recording. Federal law allows recording with one party’s consent (18 U.S.C. 2511), but California requires all parties’ consent for confidential calls (Penal Code 632, as the state supreme court explains). With callers in many states, announcing recording on every call is the safer default.
• UK. The ICO’s monitoring workers guidance says call monitoring for training or quality control can be justified, that workers must be told, and that callers must be told a call is recorded and why. It also says an organization cannot justify monitoring its workers solely because a customer makes it a condition of business, which matters when a client’s contract demands call scoring. The ICO notes it is reviewing the guidance after the Data (Use and Access) Act.
• EU AI Act. Article 50, applying since 2 August 2026, requires providers to design AI systems that talk to people so that those people are told it is an AI, unless that is obvious, and requires deployers of emotion recognition to inform those exposed. Article 5(1)(f) prohibits inferring emotions of people in the workplace, except for medical or safety reasons, which is why agent-tone scoring from voice is risky. Annex III lists AI that monitors and evaluates workers’ performance and behavior as high-risk, with those duties now due from 2 December 2027 under the updated timeline. See our EU AI Act guide.
• UAE. The data protection law sets the processor duties above. Marketing calls by licensed companies fall under Cabinet Resolution 56 of 2024: callers must tell the consumer at the start that the call is recorded, respect the Do Not Connect register and call only between 9:00 am and 6:00 pm. Licensed financial institutions also face the Central Bank’s telemarketing consent rules, where consent records whether the customer accepts AI-based agents. See also our guide to building AI in the UAE.

12How should you start, and what should you measure?

1. Pick one client and one queue with steady volume and a QA form that already exists.
2. Build the tenant boundary and the redaction step before any model touches a call.
3. Run automated QA in shadow beside human QA for several weeks, and compare question by question.
4. Add agent assist for one call type, with that client’s knowledge base only.
5. Add a voice agent for one call type from the top of the autonomy table, with a one-step route to a person.
6. Change the commercial terms for that queue last, once the numbers below are stable.

Track five numbers, per client:

• Share of calls scored, and agreement with human calibrators on each question.
• After-call work time and average handle time on assisted calls.
• Contained-and-resolved rate next to raw containment, plus repeat contact.
• Time to reach a person, and survey scores for bot calls against person calls.
• Cross-tenant leak tests passed (the target is no failures, ever).

This work sits within our AI systems and integration services, including the evals and monitoring after launch.

13Questions call center operators ask

Can one AI system serve all our clients?

One platform can, but the data can’t be shared. Run the same code for every client, with separate data, keys, retrieval indexes, scorecards and retention per tenant. Test that boundary on every release.

Will automated QA replace our QA analysts?

It replaces sampling, not judgment. Analysts move to calibrating the scores, handling disputes and coaching. Gartner’s survey, cited above, found organizations adding specialized AI roles rather than only cutting headcount.

Do we have to tell callers they’re talking to an AI?

In the EU, the AI Act requires it unless it’s obvious from the context. In the US, outbound AI voice calls need prior express consent, and state rules vary. Disclosing at the start of every AI call is the safer default, and a qualified professional should confirm it for each market.

How long until a first system is live?

A focused QA or agent-assist system for one client and one queue is the realistic first release, and it includes a shadow period beside your current process. The tenant boundary and the integration with the client’s platform and CRM drive most of the effort, so we scope the timeline after seeing those.