What Your Team Leader Did Last Tuesday
Here's a real week, reconstructed from the kind of operation we talk to constantly. Six people on the phones, roughly nine hundred calls between them.
| Time spent | On what | What came out of it |
|---|---|---|
| 2 hours | Listening to eleven recorded calls, chosen because they were the right length and easy to find. | Eleven scorecards. Coverage: about 1.2% of the week's conversations. |
| 1 hour | Filling in the scorecards and writing feedback notes. | Documentation that will be read once. |
| 1 hour | Six short conversations about calls from two to four weeks ago. | Two people couldn't remember the call. One disputed the sample. Three nodded. |
| Four hours of your most experienced person, to review 1.2% of the work and change approximately nothing. | ||
Nobody in that story is doing a bad job. The team leader is conscientious, the scorecard is well designed, the feedback is delivered kindly. The method is the problem, and it's worth naming exactly how.
Statistically hollow
Eleven calls from nine hundred describes which calls got picked more than it describes anybody's performance. Two people of identical ability score very differently on chance alone.
Too late to be coaching
Nobody learns from feedback on a conversation they can't recall. Coaching works close to the event, and a monthly cycle guarantees it never is.
Reads as a verdict
A rare score on a rare sample feels like judgement, which is why reviews produce arguments about the sample rather than changes in behaviour.
Most quality programmes exist so the organisation can say it has one. That isn't a criticism of the people running them — it's a description of what's possible when your only instrument is a human ear and a spreadsheet.
— the VOCPhone team
Why This Is a 2026 Conversation
Two capabilities arrived close together, and their combination is what makes this newly practical rather than a 2019 pitch.
Speech-to-text got accurate and cheap enough to run on everything. Not perfect — we come back to that — but good enough that transcribing every call is now a routine platform function rather than a budget line item.
Language models became able to evaluate against a written rubric. This is the genuinely new part. Earlier systems could only search transcripts for keywords, producing the well-known absurdity of an agent scoring well for saying "I understand your frustration" in a flat monotone. A model can now be handed a description of what a good call looks like in your business and asked to assess a conversation against it — including things keyword search could never detect, such as whether the customer's question was actually answered.
| Manual review | Keyword spotting | AI evaluation | |
|---|---|---|---|
| Coverage | ~1–2% of calls | 100%, phrases only | ✓ 100% of calls |
| Latency | Days to weeks | Immediate | ✓ Immediate |
| Understands meaning | ✓ Fully | ✗ Not at all | Substantially |
| Consistent call to call | ✗ Varies by assessor and mood | ✓ Perfectly | ✓ Highly |
| Cost per call reviewed | High — senior labour | Negligible | ✓ Low |
| Trusted by the team | Depends on the leader | ✗ Openly gamed | Depends entirely on rollout |
That final row is where these projects live or die, and section six is about it. First, precision — because over-claiming is the fastest way to lose a team.
Three Tiers of What AI Can Judge
✓ Reliable — automate fully
Was it said? Greeting, recording notice, required disclosure, identity check, mandated next step, whether the offer was made at all.
Mechanics. Talk-over ratio, dead air, longest monologue, interruptions, hold time, who spoke more.
Classification. Call reason, topic, product, outcome, escalation — across every call.
~ Partly reliable — aggregate only
Sentiment and customer effort. Genuinely useful across thousands of calls and over weeks. Weak on any individual call, because tone is ambiguous — and Australians in particular will say "no worries, mate" while being thoroughly unimpressed.
Use for trends. Never quote a single call's sentiment score at a person.
✗ Not reliable — keep human
Was the answer right for this customer? Needs their account, history and circumstances.
Were they satisfied, or just polite? Different things.
Relationship context. The fourth call in a bad week reads nothing like the first.
These are the calls a manager should hear — and AI is excellent at finding them.
The rule that keeps you honest: automate the objective and observable; reserve human judgement for the interpretive. Used that way, AI doesn't replace team leaders — it stops them spending four hours on calls that were fine so they can spend it on the ones that weren't.
It's also worth knowing that transcription errors aren't randomly distributed. Accuracy drops with strong accents, background noise, poor line quality and specialist vocabulary — trade terms, drug names, part numbers, legal phrasing. Which means errors concentrate in exactly the industries and workforces most likely to be treated unfairly by a naive scoring system. That's a reason to design carefully, not a reason to avoid the technology. What drives accuracy in practice is covered in AI call transcription, summaries and CRM notes.
The Compliance Case
If you need one business case that survives a board meeting, this is it — because it replaces a statistical argument with a complete record.
Plenty of Australian businesses have obligations attached to what gets said on a call: a disclosure before payment, a notice that the call is recorded, identifying yourself in a prescribed way, a defined process for customers in vulnerable circumstances, a required warning before a product is sold. Traditionally you assured that by sampling and hoping the sample was representative.
1.2%
Coverage a manual programme achieved in the week above
100%
Coverage an automated check achieves
Day 2
When a new starter's missing disclosure surfaces, instead of a quarterly audit
Two consequences. Systemic problems surface almost immediately — the new starter who has never given a required disclosure shows up on day two, when it's cheap to correct, rather than at a quarterly review with three months of calls behind it. And you can demonstrate coverage rather than describe a methodology, which is a materially stronger position in front of a regulator, an auditor or an insurer.
Obligations of this shape are also expanding — see the transparency rules and, for consumer-facing operations, the growing set of scam-related duties. Every addition makes complete coverage worth more than it was last year.
Two things that cut the other way
First, recording and monitoring calls has its own rules. In Australia there's no single national rule — call recording is governed by state and territory listening-devices and surveillance-devices legislation, which varies, on top of your Privacy Act 1988 and Australian Privacy Principles obligations. The safe posture almost everyone adopts is simply to tell people: a clear notice at the start of the call plus a documented internal policy. Get that right before you build scoring on top of it. Second, recordings and transcripts are personal information, so where they're stored and processed is part of your obligation — and AI features are the most common way that data quietly leaves the country. See follow the money: the four companies between you and your calls.
Month One, Month Three, Month Six
The scoring is the dull half. The interesting half is what complete coverage lets a good team leader do — and it compounds.
| When | What becomes possible |
|---|---|
| Month 1 | Nobody types call notes any more. Summaries and action items appear on the deal automatically. The team's first experience of the technology is relief, and you've bought a month of goodwill for free. |
| Month 2 | Coaching gets specific: "here are the four calls this week where the customer went quiet right after you quoted price — let's listen to two." Recent, evidenced, and impossible to dispute as an unrepresentative sample. |
| Month 3 | Patterns appear that nobody could see before. Tuesday's dominant call reason. The question 40% of callers ask that isn't answered anywhere on your website. The step in your process that generates the most repeat calls. |
| Month 4 | New starters stop learning by absorption. They get the five best real calls for the enquiry type they'll face tomorrow, chosen from data rather than from whoever was free to shadow. Ramp time is usually the single biggest win. |
| Month 6 | Best practice is identified from calls that actually resolved and taught deliberately — instead of being whatever the loudest senior person happens to do. And your team leader has their four hours a week back. |
Notice the pattern: every row is specific, recent and evidenced. That's the difference between coaching that changes behaviour and coaching that produces a nod.
The Trust Problem Is the Real Risk
Everything above is achievable. The reason most of these projects underperform has nothing to do with the technology and everything to do with how it arrives.
Put yourself on the other side of it. You take calls all day. On Monday someone announces that an AI will now listen to every one and give you a score. However carefully that's worded, the first thought isn't "excellent, better coaching". It's "they're looking for a reason".
- Tell them before you switch it on. Finding out afterwards causes damage that takes a year to repair, and it's entirely avoidable.
- Show each person their own data first. Let the team see their own scores before any manager acts on them. This one step converts more scepticism than any explanation you could write.
- Put it in writing that scores alone never lead to discipline. A human listens to the actual call before any consequence. Scoring a conversation is not the same as understanding it.
- Start with something that helps rather than judges. Automatic call notes are ideal: the AI does the paperwork nobody wanted, and the team's first experience of it is relief rather than exposure.
- Check your own data for unfairness deliberately. Because transcription accuracy varies with accent and line quality, compare scores across teams and individuals and ask whether differences track performance or track something else. If you find a pattern, fix the process rather than defend the tool.
Point four is worth more than it sounds, and it costs nothing to sequence it that way. Teams whose first encounter with this technology is it removing admin accept scoring far more readily than teams who meet it as a new way of being marked.
Measuring What Matters, Not What's Easy
The second failure mode is subtler and hits well-run operations hardest: anything you measure becomes a target, and anything that becomes a target gets optimised at the expense of the thing you actually wanted.
| Tempting to score | What you'll actually get | Score this instead |
|---|---|---|
| Script adherence | Scripts followed, problems unsolved | Was the customer's actual question answered? |
| Average call duration | Fast calls, and callbacks tomorrow | Resolved without a follow-up contact |
| Positive language count | Performative cheerfulness in situations needing straight talk | Customer effort — how hard was this for them? |
| Calls handled per hour | Rushed calls and people leaving in six months | Outcomes per hour, quality included |
| Sentiment per individual call | Arguments about individual calls | Sentiment trend across a team over weeks |
The defence is to score outcomes rather than behaviours wherever you can, and to revisit the rubric every quarter asking one question: what is this measure quietly encouraging? Being able to score everything makes rubric design far more consequential than when you scored eleven calls a week — because now the incentive reaches every single conversation.
Your AI Agent Needs This More Than Your Team Does
Here's the argument almost nobody makes, and it may matter most.
If you've deployed an AI Phone Agent to answer calls, you've created a worker who handles high volume with no supervisor, no team leader walking past, and no colleague overhearing something wrong. Consider the asymmetry:
Human error is self-limiting
One person misunderstands a policy and gets it wrong a handful of times before somebody corrects them. The damage is bounded by one shift and one caseload.
AI error is systematic
An AI agent with a wrong understanding applies it identically on every call until somebody notices. Four hundred calls, four hundred identical errors, and no variation to trigger anyone's suspicion.
So the same pipeline reviewing your team should review your AI, against its own questions:
- Did it answer accurately, or confidently invent something?
- Did it stay inside its defined scope, or drift into territory it should have escalated?
- Did it hand over to a human at the right moment — and inside the same conversation, without making the customer start again?
- Did it make a commitment on your behalf that it had no business making?
- What did callers ask that it couldn't handle? This is the most valuable output of the lot: your product and process roadmap, written by your customers.
A fair test for any AI answering vendor
"Show me how I audit what the AI said." If the answer is a dashboard of call counts and containment rates, that's metrics, not assurance. You want transcripts, the ability to search them, and scoring against your own criteria. A provider selling AI answering with no means to audit it is selling half a product — and it's the risky half they kept.
Which calls to hand an AI in the first place is a separate and equally important decision: which calls to automate and which to keep human.
Six Steps, In This Order
- Write down what a good call is. Two pages, plain language, agreed by the people who actually manage the team. Descriptive, not aspirational. AI cannot score a standard you've never articulated — and most organisations discover here that they disagree internally, which is itself worth finding out.
- Start with notes, not scores. Turn on transcription and automatic summaries first. Buys goodwill, proves accuracy on your actual calls and accents, and costs nothing in trust.
- Calibrate against humans. Have your best assessor score fifty calls independently, then compare against the AI and investigate every disagreement. You'll find both AI errors and rubric ambiguity — fixing them now is far cheaper than arguing later.
- Tell the team, then show them their own data. Explain what's measured, what it's for, and what it will never be used for. Then give each person their own scores before any manager sees them in a review. This is the step that decides the outcome. Don't compress it.
- Coach only, for a full quarter. No scorecards in reviews, no league tables, no consequences. Lets the data prove itself as help rather than judgement, and gives you a baseline to measure improvement against.
- Review the rubric, then extend. Ask what the measures are quietly encouraging. Adjust. Only then extend into compliance alerting and formal reporting — and keep revisiting the rubric quarterly, permanently.
Most failed implementations jumped straight to step six. The sequence is the intervention.
Buying Criteria
Built in, not bolted on
Recording, transcription, AI summaries and reporting as part of the platform rather than three vendors and an integration project. VOCPhone includes them in transparent per-user pricing.
Your rubric, in your words
A fixed vendor scorecard describes a generic contact centre. You don't run one. Insist on defining the criteria yourself.
Australian data handling — including AI
Ask specifically where audio and transcripts are processed for AI features, whether a third party retains them, and whether they train models. VOCPhone keeps all of it on Australian infrastructure by default.
Searchable transcripts, not just dashboards
You need to reach the actual conversation behind any number. A dashboard you can't drill into is decoration.
Pushes into the systems you use
Summaries and outcomes landing in your CRM automatically — 1,000+ integrations and open APIs. A transcript nobody sees is a transcript nobody uses.
Every channel in one view
Quality isn't a voice-only question. Calls, SMS and chat should be assessable together — see one inbox for calls, SMS and chat.
For the wider platform decision this sits inside, the best contact centre software in Australia covers the full evaluation and why VOCPhone leads Australian CCaaS covers where we sit in it.
Frequently Asked Questions
What is actually wrong with sampling a few calls per agent?
Four things, and they compound into a programme that consumes real money while producing very little. It is statistically hollow: reviewing a handful of calls from several hundred tells you more about which calls got picked than about how somebody performs, and two people of identical ability can score very differently on chance alone. It arrives too late to be coaching, because nobody learns from feedback on a conversation they cannot remember. It reads as a verdict rather than help, which is why quality reviews produce defensiveness about the sample instead of changes in behaviour. And it is expensive in the wrong currency, because it consumes the most leveraged labour in the room - a team leader spending hours listening to calls instead of leading. The uncomfortable summary is that most quality programmes are performed rather than useful, and everybody involved half knows it.
What can AI judge reliably, and what can it not?
Three tiers, and any vendor unwilling to draw them is selling you a future argument. Reliable: whether specific things were said - a greeting, a recording notice, a required disclosure, a mandated next step; conversation mechanics such as talk-over ratio, dead air, longest monologue and interruptions; and classification of call reason, topic and outcome across every call. Partially reliable: sentiment and customer effort, which are genuinely useful in aggregate and across time but weak on any single call, not least because Australians will say no worries while being thoroughly unimpressed. Not reliable: whether the answer given was actually right for that customer's circumstances, whether they were satisfied rather than merely polite, and anything needing the context of a long relationship. The rule that keeps you honest is to automate the objective and observable and keep human judgement for the interpretive.
How do we stop the team seeing this as surveillance?
By treating it as a trust problem rather than a technology problem, because that is what it is, and it is the reason most of these projects underperform. Four commitments make the difference. Tell people before you switch it on - finding out afterwards is what causes lasting damage and it is entirely avoidable. Show each person their own data first and let them see it before any manager acts on it, which converts more scepticism than any amount of explanation. Put in writing that a score alone never leads to discipline and that a human listens to the actual call before any consequence, because scoring a conversation is not the same as understanding it. And start with something that helps rather than judges: automatic call notes are ideal, because the technology's first act is to remove paperwork nobody wanted to do. Teams whose first experience of this is relief accept scoring far more readily than teams who meet it as a new way of being marked.
Is the compliance case strong enough to justify it on its own?
For many Australian businesses, yes, because it replaces a statistical argument with a complete record. If you are required to make a disclosure, state that a call is recorded, identify yourself in a particular way, follow a defined process for customers in vulnerable circumstances, or complete a scripted step before taking payment, then automated checking covers every call instead of two of them. Two things follow. Systemic problems surface almost immediately - a new starter who has never given a required disclosure shows up on day two rather than in a quarterly audit, which is when it is cheap to fix. And you can demonstrate coverage rather than describe a sampling methodology, which is a materially stronger position in front of a regulator, an auditor or an insurer. Australian obligations of this shape are also expanding, so the value of complete coverage rises rather than falls.
Can we use AI scores in performance reviews or pay decisions?
With real care, and never as the only input. Three practical reasons. Speech recognition is imperfect and its errors are not randomly distributed - accuracy drops with strong accents, background noise, poor lines and specialist vocabulary, which means errors concentrate in exactly the workforces most likely to be treated unfairly by a naive system. Anything measured becomes a target, so scoring script adherence reliably produces people who follow the script while helping the customer less. And a decision with consequences for someone's employment needs a human in the loop, both because it is fair and because you may have to justify it later. The workable position is to use AI for coverage, trend detection and coaching, and to require a manager to listen to the actual call before any decision that affects someone's standing or pay.
If we run an AI Phone Agent, does it need reviewing too?
More than your people do, and this is the point most businesses miss entirely. An AI agent handling calls is a worker with no supervisor, no team leader walking past and no colleague overhearing something wrong. Human error is naturally self-limiting: one person misunderstands something and gets it wrong a handful of times before somebody corrects them. An AI agent with a wrong understanding applies it identically to every single call until somebody notices - four hundred calls, four hundred identical errors, and no variation to trigger anyone's suspicion. So the same review pipeline should cover it, against questions specific to it: did it answer accurately or confidently invent something, did it stay inside its scope, did it hand over to a human at the right moment and inside the same conversation, did it make a commitment on your behalf it had no business making. And crucially, what did callers ask that it could not handle - which is your product and process roadmap, written by your customers.
What do we need before we can start?
Less technically than most people expect, and more organisationally. Technically: call recording enabled and retained long enough to be useful, transcription, and the ability to evaluate transcripts against criteria you define. VOCPhone includes call recording, transcription and AI summaries as part of the platform rather than as separate purchases. Organisationally, three things matter more than the tooling. A written description of what a good call looks like in your business, because AI cannot score a standard you have never articulated - and most organisations discover at this step that they disagree internally, which is itself worth finding out. A decision about who sees which scores, made before launch rather than after an argument. And clarity on where recordings and transcripts are stored and processed, because both are personal information under the Privacy Act 1988. VOCPhone keeps that data, including the AI processing, on Australian infrastructure by default.
