AI Writes the Call Notes. Agents Rewrite Them Anyway.

The dashboard says the AI call summaries are a hit. Nearly every call gets one, and they read beautifully. Then you look at after-call work at an Australian utility three months in, and it has barely moved. The model writes the note in seconds. The agent spends the next minute rewriting it.

Sit next to an agent for twenty minutes and you see why. The summary appears, the agent reads it, deletes half of it, retypes the account details into the right fields, picks the disposition code by hand, and adds the one line that actually matters for the next person who picks up the case. The AI did write the note. The agent still did the work.

This is one of the most common AI use cases in Australian enterprises right now, and one of the most quietly disappointing. Not because the models are bad at summarising, but because a summary is not the same thing as a usable record.

Fluent, Tidy and Rewritten Anyway

Most rollouts measure the wrong thing. They check whether a summary was produced and whether it reads fluently. Agents and team leaders judge it on a different test: can the next person act on this note without listening to the recording?

When summaries fail that test, it is usually for a handful of predictable reasons:

• The summary is prose, but the CRM wants fields. The model writes a tidy paragraph. The case record needs a reason code, a product, a promised callback date and a hardship flag. Agents end up copying details out of the paragraph and into the form.
• It captures the conversation, not the commitment. A good note records what was promised: a credit applied, a technician booked, a complaint escalated. Generic summaries spend three sentences on the customer's frustration and half a sentence on what the business agreed to do.
• It gets the details slightly wrong. A transposed digit in a reference number or a misheard suburb is a small error in a summary and a serious one in a customer record. Agents learn quickly that they need to check every specific, which removes most of the time saving.
• It ignores the house style. Every team has shorthand that works for them. A summary that does not match it feels foreign, and people rewrite things that feel foreign.
• Vulnerability and complaints need care. Notes about financial hardship, family violence or a formal complaint carry obligations. Agents are rightly cautious about trusting a model to word them, so they rewrite those notes entirely.

Count What Survives, Not What Ships

The single most useful metric for this use case is how much of the AI summary survives into the saved record. If you capture the draft the model produced and the note the agent finally saved, you can compare them. Track the share of summaries saved with no changes, light changes and heavy rewrites, and break it down by call type, team and queue.

That one measure tells you far more than a satisfaction survey. You will usually find the model is genuinely helpful on some call types, like simple billing questions or address changes, and close to useless on others, like complaints or multi-issue calls. That is not a failure. It is the map of where to invest next.

Pair it with the numbers the business actually cares about: after-call work per call type, repeat contact rate, and how often a second agent has to replay a recording because the note did not help. If the summary is working, those move. If they do not move, a high generation rate means nothing.

Fix the Note Before the Prompt

The teams that get real value tend to start in the same place. Before touching the model, they sit down with team leaders and quality assurance and agree what a good note contains for each major call type. Then they build the AI step around that.

• Generate structure, not just prose. Ask the model for the specific fields the CRM needs, such as reason, outcome, commitments and follow-up date, and write them straight into those fields as suggestions the agent confirms.
• Put commitments first. The opening line of every note should be what the business agreed to do and by when. Everything else is supporting detail.
• Verify the specifics against systems. Reference numbers, account identifiers and appointment times should be checked against the source system rather than trusted from the transcript. Where they do not match, highlight them for the agent instead of silently guessing.
• Route sensitive calls differently. Calls flagged for hardship, vulnerability or complaints can get a more conservative template, a mandatory review, or no auto-summary at all. Being explicit about this builds agent trust in the rest.
• Use the edits to improve. The rewritten notes are a free, ongoing source of examples. Review a sample every fortnight, find the patterns agents keep fixing, and update the template and instructions.

The Boring Bits Bite Back

Call summaries sit on top of recordings and transcripts, which means they inherit every obligation those already carry. Check that your call recording notice and privacy policy cover transcription and AI processing, that transcripts and drafts are retained and deleted on the same schedule as the recordings, and that you know where the transcription and summarisation actually run. A summary that ends up in a vendor's logs in another region is a data handling question your privacy team will want answered before go-live, not after.

Quality assurance teams also need a view. If notes become partly machine written, QA should be able to see which parts the agent confirmed and which were accepted as suggested. That matters when a complaint escalates and someone asks who wrote what.

Five Steps, Three Call Types

1. Pick three high-volume, low-risk call types and agree a note standard for each with team leaders and QA.
2. Capture draft and final notes from day one so you can measure survival rate and see what gets changed.
3. Write to CRM fields as suggestions, with the agent confirming each one, rather than dropping a paragraph into a free text box.
4. Set a target that matters, such as a reduction in after-call work for those call types, and review it after six weeks.
5. Expand call type by call type only when the survival rate and the business metric both justify it.

How TruFyre Helps

At TruFyre we help Australian organisations move AI use cases like call summarisation from a promising demo to something agents genuinely rely on. That means designing the note around the workflow, integrating with the CRM properly, putting measurement in place from the start, and making sure privacy, retention and QA requirements are covered before the rollout grows.

If your contact centre has AI summaries switched on and after-call work has not budged, talk to us. The fix is rarely a better model. It is usually a better definition of what a good note looks like.