Picture the transcript: “Yes, that’s fine. No, I understand. Okay, go ahead.” Every box ticks. No complaint keywords, no flagged phrases, the customer agreed to everything.
Now listen to the audio. The voice is flat, the pauses are long, and “that’s fine” comes out the way people say it when nothing is fine. That customer may be vulnerable, confused or under pressure, and a text-only system will never know.
That gap is the difference between what a Named Entity Recognition (NER) system does and what Smartbot does. They are not rivals. They answer different questions, and under FCA Consumer Duty you need both answered.
What NER does well
NER reads text and pulls out the things that matter: names, account numbers, dates, amounts, product names, firms. In a contact centre it turns a messy transcript into structured data you can search, report on and audit.
It earns its place in four ways:
- PII redaction. It finds card numbers, addresses and dates of birth so they can be masked before transcripts are stored or shared.
- Structure. It tags which product, which policy and which amount a call was about, so you can route, report and spot trends.
- Keyword and phrase evidence. It can show that a required disclosure was read, or that a customer mentioned bereavement, illness or job loss.
- Auditability. A tagged entity is easy to explain to a regulator: here is the word, here is where it was said.
It is mature, cheap to run and well understood. If you only need to know what was said, NER is a good answer.
Where NER goes quiet
NER only sees words, and only looks for entities in them. Everything else a call carries is left on the table.
- No tone. Stress, hesitation, confusion and resignation live in pitch, pace and pauses, not vocabulary. “I’m fine” tags the same whether it is said brightly or through tears.
- Vulnerability rarely announces itself. Customers in difficulty often don’t say “bereavement” or “debt”. If the keyword never appears, the entity is never found.
- Usually after the call. Most NER runs on finished transcripts. By the time a flag is raised, the customer has hung up and the outcome is set.
- Transcription errors compound. Accents, crosstalk and poor lines degrade the transcript, and NER inherits every mistake.
Consumer Duty asks firms to evidence good outcomes and to recognise vulnerable customers. A system that only extracts entities, only after the fact, misses the most important signals.
What Smartbot brings
Smartbot reads how the customer is, from two sources at once: their voice and their words. It does this live, while the call is happening.
- Tone and text, fused. Our patent-pending tonality model, built with MTU, analyses the voice in rolling three-second windows. A text model reads the live transcript alongside it. SIGNAL™, our real-time fusion of tone and text, combines the two. In benchmark testing, the fused model reached 79.9% accuracy, against about 70% for text or audio on its own.
- Real time, not post-mortem. An agent or supervisor can see distress building while there is still time to change the outcome.
- Vulnerability without keywords. A customer who never says the trigger word can still sound anxious, confused or defeated, and their words can still carry it. Smartbot reads both and can flag it.
- Explainable scoring. A calibrated scoring layer sits above the neural models, so every flag comes with thresholds and probabilities a compliance team can defend, not a black-box score.
- Your data stays in your AWS account. Smartbot is available on AWS Marketplace and deploys in your own AWS environment, with no telephony rip-and-replace.
In short: NER tells you what is in the words. Smartbot tells you how the customer is, from the words and the voice together.
What Smartbot doesn’t do
Smartbot reads the transcript, but for sentiment, not entities. We would be overselling if we said it replaces NER.
- It won’t redact your PII. Knowing a customer is anxious tells you nothing about where their card number sits in the transcript.
- It won’t tag the product, amount or account. Smartbot shows how the call went, not what it was about.
- It won’t prove a disclosure was read word for word. That is a check on specific words, and NER and keyword tools do it well.
Better together
NER and Smartbot cover different ground. Run together, NER tells you what was discussed and whether the right words were said. Smartbot tells you how the customer was coping while it was said.
- What was said? NER: extracts entities and keywords. Smartbot: reads the words for sentiment.
- How did the customer sound? NER: no. Smartbot: yes.
- Which product, amount or account? NER: yes. Smartbot: no.
- Can it redact PII? NER: yes. Smartbot: no.
- Can it spot vulnerability with no keyword? NER: no. Smartbot: yes, from tone and words together.
- When does it flag? NER: usually post-call. Smartbot: in real time, during the call.
- Is the output explainable? NER: yes, at word level. Smartbot: yes, with calibrated scores.
The bottom line
If your compliance stack only extracts entities from transcripts, it misses how the customer is. NER is excellent at what is in the words. Smartbot reads the words and the voice together, in real time, so the customer who says “that’s fine” and doesn’t mean it gets noticed before the call ends.
Whether your customers are speaking to contact centre agents or AI voice agents, Smartbot gives you live oversight of every conversation. Want to hear what your calls are really saying? Try Smartbot free on AWS Marketplace or get in touch at hello@smartbotsoftware.com.
Put live oversight on your voice AI agents
SIGNAL™ flags vulnerability and distress during the call and builds your Consumer Duty evidence trail. Deploy it from AWS Marketplace and it runs inside your own AWS account.
Not on AWS, or prefer a hosted option? Talk to us →
