HSBC: An AI Voice Insights platform projected to save 1.5M hours a year

Yes, the engineering transfers even though the product differs. The HSBC Voice Insights work applied natural language processing, sentiment analysis and entity recognition to enterprise voice data inside a regulated bank, with innovation, operations and risk in the room from the start. Those are disciplines agentic systems still stand on: extracting reliable structure from messy language, keeping privacy and accuracy intact at scale, and building alongside the people who own the controls. What does not transfer is production evidence. The platform reached proof of concept, so treat the study as engineering pedigree in a regulated environment, not as a claim that this system runs anywhere today.

It is a projection, and the study labels it as one. The proof of concept demonstrated that natural language processing, sentiment analysis and entity recognition could remove manual administration, note-taking and summarising across millions of hours of recorded calls and meetings, and 1.5M+ hours a year is what that removal was projected to be worth at HSBC's scale. Nobody has banked it. The platform reached proof of concept and stopped short of production, so it is potential shown, not a saving audited. A projection labelled as a projection is worth more to you than a banked-sounding number nobody can stand behind.

Working prototypes, not slideware. The build combined natural language processing, sentiment analysis and entity recognition over enterprise voice data, and shipped prototypes that auto-generated meeting summaries, detected emerging themes across conversations, and surfaced real-time dashboards. The point was to turn millions of hours of captured but unanalysed calls and meetings into usable intelligence, with privacy, accuracy and scale treated as requirements rather than afterthoughts. The study's own label for it is a proof of concept shipped, not theorised.

No. It reached proof of concept, and nobody has run it at scale. The work did prove that voice data could become a new source of business intelligence inside a regulated bank, which is why it still earned a place in HSBC's thinking on AI platforms. Tenhaw writes every case study this way. Where an engagement ended at a pilot or a proof of concept, the study says so, because a prototype described honestly tells you more about how a consultancy will report to you than a rollout implied and never delivered.

Privacy was in the task definition from the start, not retrofitted. The brief was to turn recorded calls and meetings into usable intelligence with privacy, accuracy and scale intact, so the constraint sat alongside the capability and never arrived as a late objection. In practice that meant coordinating innovation, operations and risk throughout the build, so the people who would have to approve adoption were shaping the work while it was being made. The detailed controls are HSBC's own. The operating pattern is the part that travels. In a regulated estate, adoption readiness gets built with the risk function, not presented to it.

Customer insight, operational inefficiency and risk signals, in HSBC's case buried across millions of hours of voice data that was captured but never analysed. Every call and meeting held information about what customers wanted, where processes were failing and where risk was emerging, and none of it was reachable because the only extraction method was a person taking notes. That is the shape of the problem in most large organisations: the data already exists and is already being recorded, so the cost is not collection but analysis. The HSBC proof of concept showed the analysis could be automated with natural language processing, sentiment analysis and entity recognition.

Because a prototype that risk sees for the first time when it is finished is a prototype that stalls. On the HSBC Voice Insights work, innovation, operations and risk were coordinated throughout the build, so the platform was moving towards adoption readiness while it was still being made rather than queuing for approval afterwards. In a regulated bank that is the difference between a demonstration and a candidate for deployment. Tenhaw carries the same pattern into agentic work now. The people who own the controls are in the room from the start, because retrofitting their requirements is slower than building to them.

Yes, and the HSBC proof of concept demonstrated it inside a major regulated bank. Applying natural language processing, sentiment analysis and entity recognition to enterprise voice data produced auto-generated meeting summaries, emerging-theme detection and real-time dashboards, turning a passive archive of recorded calls and meetings into intelligence the business could act on. At HSBC's scale the projection was 1.5M+ hours a year of manual administration removed. If your calls and meetings are already recorded then the raw material exists. The analysis layer that reads it can be built and shipped, and three months at HSBC produced one.

Three months, from James joining HSBC's Innovation Portfolio to a shipped proof of concept with working prototypes over enterprise voice data. For comparison, Tenhaw's current Agentic Proof of Concept is a fixed £20k–£55k and runs two to four weeks, a pace today's models and tooling make possible. What has not changed is the shape of the outcome. Both end in a working system you can judge with your own eyes, not a paper recommending one.

It did at HSBC. The Voice Insights proof of concept proved that voice data could become a new source of business intelligence, and that evidence shaped the bank's strategic direction on AI platforms. It carried that weight because it was shipped, not theorised. The case rested on working prototypes running over real enterprise voice data, and innovation, operations and risk had been coordinated throughout the build, so the result arrived with adoption readiness attached instead of a queue of unanswered objections. A working prototype changes a strategy conversation in a way no deck can. That is why Tenhaw's own AI Readiness Audit ends in working prototypes.

For a single meeting, use it. The HSBC AI Voice Insights proof of concept sat a layer above that, applying natural language processing, sentiment analysis and entity recognition across an estate where millions of hours of calls and meetings were captured but never analysed, and shipping prototypes that auto-generated summaries, detected emerging themes and surfaced real-time dashboards. A built-in summariser tells you what was said in the room you were already in. Reading the whole archive tells you what customers keep raising, where processes are failing and where risk is building. The study describes that as the insight buried in voice data. Both are uses of personal data, so the smaller tool does not remove the privacy work.

Because the conversation happens first, and almost none of it reaches you. At HSBC, millions of hours of calls and meetings were captured but never analysed, with insight about customers, inefficiency and risk sitting inside them, unreachable because the only extraction method was a person taking notes. A survey gives you the people who chose to answer, after the event, in your words. The recorded estate is everyone who called, in theirs. What the proof of concept established is that voice data could become a new source of business intelligence, with natural language processing, sentiment analysis and entity recognition doing the reading no team could do by hand.

It turns the specifics of a conversation into things you can count. Natural language processing gives you structure and sentiment analysis gives you tone, but entity recognition is what pulls out the names, products, amounts and references a call contains, so a stretch of speech becomes something you can group, filter and trend. At HSBC the three were used together over enterprise voice data. That combination is what made emerging-theme detection and real-time dashboards possible. Without it you have a searchable pile of words. The same extraction lets a write-up be generated instead of typed, and that is where the projected 1.5M+ hours a year of manual administration sat.

Summaries, usually, because that is where the removable administration sits and the people whose calls are analysed feel it immediately. The HSBC figure of 1.5M+ hours a year came from manual note-taking and summarising, and it remains a projection. That platform reached proof of concept and went no further. Themes and dashboards are worth more over a longer run, since they tell you what is rising across the whole estate and not just what happened in one conversation, but they only pay if someone in operations has already decided what they will do with the answer. At HSBC all three shipped inside a three-month proof of concept, so sequencing is rarely the real constraint.