QA Samples Four Calls a Month on Your PCI Program. On-Site AI Now Reviews All 118,000 Without the Audio Leaving the Building.

It is 9:15 on a Tuesday and Dana, the quality analyst on the card-services program, has the scorecard open on her second monitor and a call from Friday afternoon playing back at 1.4x. She will get through maybe six calls before the 11 a.m. calibration session with the client. Her monthly target is four calls per agent. There are 120 agents on the program. That is 480 calls reviewed out of roughly 118,000 handled.

Four-tenths of one percent. Every operations manager in this business knows that number and nobody says it out loud on a client governance call. When the retailer's vendor risk team asks how you monitor for a card number being read aloud outside the pause-and-resume window, the honest answer is: "we sample, and we have never caught one, which is not the same thing as there never having been one."

The reason coverage is 0.4% and not 100% is not the cost of the software. Speech scoring got cheap two years ago. The reason is Schedule B of the data processing addendum.

The clause that keeps the recordings inside the building

Pull your master services agreement for any regulated program and find the subprocessor list. On a payment program under PCI DSS 4.0.1, on a payer program running under a HIPAA business associate agreement, on a county or state program with a criminal-justice data clause, the wording is some version of the same thing: client data may only be processed by entities named in writing, in the approved location, and any addition requires consent.

So when your operations director asks to run every recording through a cloud speech vendor, what actually happens is a change request, a security questionnaire, an updated SOC 2 Type II mapping, six to ten weeks of back-and-forth with the client's third-party risk team, and — on the healthcare and public-sector programs — a flat no. Meanwhile the client's own RFP says "100% of interactions monitored for compliance." Everyone signs it. Everyone samples.

On-site AI means the call recording, the screen capture and the transcript are processed on a machine your own badge opens, in the same building as the agent who took the call, with nothing crossing the firewall to an outside vendor. That is the whole idea, and in 2026 it stopped being a science project.

What actually changed on the hardware in 2026

Two things moved at once. Processors built for local work — the Qualcomm Dragonwing class of chips — got good enough to run capable models on a machine sitting in your own comms room rather than in somebody's data center. And large enterprises stopped treating on-premises as the legacy option. Cisco's rollout of a personal AI assistant to roughly 90,000 employees leans deliberately on on-premises processing for control and data protection, which matters to you for one very practical reason: it is now a normal answer on a security questionnaire instead of an eccentric one.

The cost side helps too. For high-volume repetitive work — and scoring every call on a 120-seat program qualifies — running locally comes in around 90% cheaper than sending the same work out. The economics finally match the compliance argument instead of fighting it.

flowchart TD
  A["Agent takes a card payment call"] --> B["Recording lands on the floor server"]
  B --> C{"Restricted program under the DPA?"}
  C -->|Yes| D["Scored on the on-site box, audio never leaves"]
  C -->|No| E["Scored in the client-approved cloud tenant"]
  D --> F["Only critical-fail flags enter the QA queue"]
  E --> F
  F --> G["QA analyst reviews flagged calls, writes coaching card"]
  G --> H["Friday calibration session with the client"]

Dana's Tuesday, run the other way

The overnight run finished at 5:40 a.m. Every call from Monday — all 5,400 of them — has been scored against the same form the client signed off on in the QA methodology exhibit: verification steps completed in order, disclosure read, pause-and-resume triggered before the payment page opened, no card or security code repeated back aloud, hold longer than two minutes without a check-back, after-call work opened within the window.

Dana does not open a list of 480 random calls. She opens a queue of 71 flagged ones, sorted by severity, with the timestamp of the moment in question already marked. Nine are pause-and-resume misses. Two of those are real. The other seven are the agent saying the word "card" while the screen was on the order summary, which the machine flagged and Dana clears in eleven seconds each.

The bigger change is the question she can now answer. At 10 a.m. her operations manager asks whether a specific disclosure was read on every enrollment call in June. Under sampling, the answer is a shrug and an offer to pull thirty calls. Now it is a search across all 118,000, with the eleven exceptions listed by agent, team lead and date. Those eleven get coached, and the answer that goes back to the client is a number instead of an assurance.

The arithmetic: same QA headcount, all the calls

Assumptions, all illustrative: 120 agents on the restricted program, 22 working days, 45 calls per agent per day. Manual review takes 12 minutes per call including scoring and notes. A loaded QA analyst costs $58,000 a year, call it $28 an hour fully burdened. Two on-floor servers cost $18,000 and get written down over three years.

Read that carefully, because the pitch you will hear from vendors is wrong. This does not cut your QA headcount. It costs you about $500 a month more. What it buys is that the same analyst hours are now spent on calls the machine already believes are failures rather than on calls picked out of a hat, and that the coverage line in your next RFP response stops being creative writing.

The money shows up elsewhere. If a single missed disclosure on a regulated program costs you a corrective action plan, a client-side audit, thirty hours of your compliance officer's time and a service credit against a monthly invoice, one avoided event pays for the servers several times over. That is the case to take to your CFO — not headcount.

Where the box is worse than Dana

It cannot judge whether an agent was actually rude. It can spot an interruption, a raised voice, dead air, a missing empathy statement; it cannot tell the difference between an agent who cut a customer off and an agent who cut a customer off because the customer was mid-way through reading a full card number aloud on a recorded line. Dana can. That distinction shows up several times a week on payment programs.

It will over-flag on strong accents, on noisy home-agent environments, and on any program where customer and agent switch between English and Spanish mid-call. Budget for false positives in the first two months and make somebody own tuning them down.

It must not deliver a disciplinary outcome. Coaching conversations, performance improvement plans and terminations stay with the team lead and HR, with a human-reviewed score behind them, because on the day an agent disputes a write-up you need a person who listened to the call. And the client's own QA team will score some of these calls differently — that is what calibration is for, and calibration stays a meeting with humans in it.

What to do Monday

Pull the data processing addendum for your two most restricted programs and highlight every clause about processing location and subprocessors. That single page tells you which of your programs can never use a cloud scoring tool, and it is usually a shorter list than people assume — plenty of programs are sampled at 0.4% purely out of habit, not because anything forbids more.

Then pick one team of twelve agents and score one week of their calls locally, on the same form the client already approved. Compare what the machine flagged against what your analyst found in her sample that week, and bring both to the next quarterly business review. The conversation about changing the QA methodology exhibit goes differently when you arrive with the comparison already done.

Frequently asked questions

Does this get me out of the client's security review?

No, and do not present it that way. You will still complete the questionnaire, still map it in your SOC 2 Type II, still show it at the annual on-site audit. What changes is the answer to the hardest question on the form. "We send recordings to a named cloud vendor in a region we will disclose" becomes "the audio is processed on hardware inside the same secure area as the agents, and no client data leaves." Reviewers approve the second answer far faster than the first.

Who runs the hardware, and does it need a data center person?

In practice it lands with whoever already owns your recording platform — the CCaaS administrator who manages Genesys Cloud CX, NICE CXone or Five9 and the Verint or Calabrio recording estate. It is closer to running a recording server than to running a research lab. If your site is a leased floor with no comms room of your own, this is harder, and that is a real constraint worth checking before you promise anything.

Will the client accept machine scoring in the QA methodology?

Usually, if you propose it as a pre-screen rather than a replacement. The wording that gets signed is some version of: every interaction is screened automatically, flagged interactions are reviewed by a certified quality analyst, and the analyst's score is the score of record. That keeps a person accountable for every number that reaches the scorecard, which is what the client's compliance lead actually cares about.

We are nearshore in Bogotá and Manila. Does this apply?

It applies more. Data residency clauses are the most common reason a US client refuses to move a program offshore. Audio processed on the same floor it was recorded on — never crossing a border, never touching a third party — answers the objection that has been costing you seats.

A note from CallSphere

CallSphere builds AI voice and chat agents that answer phone lines and web chat, book appointments and capture leads around the clock. We do not score your client programs and we do not do QA. Where operators like you tend to use us is on your own inbound line — the recruiting calls during a Q4 ramp, the RFP inquiries, the after-hours calls to your corporate number that currently roll to voicemail while every seat on the floor is busy earning revenue for somebody else.