That figure is IBM's own internal support model, adopted by 88,000 employees inside a month, with 96 percent of conversations resolved without a person. Frontline incident reporting has the same shape and a harder constraint, because someone has to stop and fill in a form, usually after the shift that produced the incident. Speech removes the stop. The deployments below were measured on whether people came back to them past the first week, and four design choices run through all of them.
29 conversational builds in the library26 IBM Consulting delivered2 with speech in the build1 voice agent live now0 Hakkoda voice builds
Internal review before external release
Three items are open. The pay-TV engagement in section 06 is described without the client name pending reference approval. The Elevance Health figures are sourced from a public post rather than a case study and need verification. The Humana source language on recognition accuracy is self contradictory and is quoted nowhere here. Nothing else on this page is held.
01
What the adopted builds share
Four choices show up in every deployment on this page that reached sustained use. Each one is documented in the record.
Design choice
In practice
Where it shows in the record
Vocabulary taken from the work
named transactions
Humana trained recognition against five named provider transactions: eligibility, benefits, claims, authorization and referral status. Scoping to a transaction list means the words the system expects are the words already in use on the floor.
A feedback loop from week one
low confidence queries
IBM internal IT support improved through continuous analysis of the queries that came back below confidence. Each one became scope for the next release.
A channel the workforce can reach mid-shift
speech first
One US grocery retailer employs more than 400,000 people, roughly 80 percent of them frontline without regular system access. Speech reaches that population while the work is happening.
Output that survives being quoted
traceable figures
Every figure carries its source, and the system holds a claim when the source is missing. Section 05 shows the chain.
02
Adoption, in the numbers
Each record below is selected for a usage figure: how many people took the system up, how fast, and whether they came back to it.
IBM internal IT support
Cross-industry / Internal IBM
Employee-facing support across the whole company, measured on take-up rather than deflection.
88,000
Employees adopted the new support model within 30 days. 96 percent of conversations resolved without live support.
Nuuday
Telecommunications / Denmark
Agent assistant inside a call center, adopted voluntarily by the agents themselves.
40%+
Of employees use the system every day. Average talk time falls 14 percent when the guides are used.
Unnamed major US grocery chain
Retail / Grocery and health
Customer care virtual agent across 12 back end systems, later extended from speech into chat and email on the same build.
1.8M
Interactions processed each month. The agent handles 25 to 30 percent of 10,000 daily calls.
Elevance Health
Healthcare / Payer
Channel migration at national scale, tracked against the phone line it replaced.
24%
Of 46 million annual phone calls moved to digital in under two years. Satisfaction runs more than 5 points above any other channel.
Humana
Healthcare / Payer
Voice native. Provider administrative staff call in and the agent resolves the request end to end against live coverage and claims data.
7,000
Voice calls handled each business day, from 120 provider organizations.
Unnamed leading aviation giant
Travel and transportation
Customer facing virtual agent carrying the full front line of enquiry volume.
20,000
Questions handled every day, at a 93 percent automated resolution rate.
CEMIG
Energy / Utility
Messaging channel for a national utility customer base, measured on whether people came back to it.
65%
Retention on the channel, alongside a 1.1 million increase in monthly messages and a 20 point gain in net promoter score.
Unnamed global professional services company
Professional services
Employee-facing assistant, measured on whether the workaround behavior stopped.
58%
Fewer human business partners calling on behalf of employees. The workaround habit stopped, which is the rarer of the two results.
Unnamed leading US grocery retailer
Retail / Grocery
Two named agents serving more than 400,000 employees, roughly 80 percent of them frontline without regular system access. The closest structural match to a distribution workforce anywhere in the library.
Structural
The savings figures in this record are labeled projected and are therefore not shown as results.
Every record above was delivered by IBM Consulting or by IBM. Hakkoda holds no voice or contact center build of its own, and nothing here claims one.
03
Quality against the human baseline
Take-up holds when quality holds. Two records put the system and the staffed alternative on the same scale, measured by the client.
The human baseline
Measured
What the build returned
Unnamed major US grocery chain
A staffed call center with an established satisfaction score
The client ran both and read the monthly numbers.
Client sponsor, on the record
The agent outscored the client's own call center agents on satisfaction, month after month
The sponsor describes the result as a surprise, which is worth reading as written. The expectation going in was that customers would tolerate it.
Unnamed large US retailer
Live agents answering customer email at 70 to 75 percent accuracy
Response time could stretch past two days.
Generative build, IBM Consulting delivered
91 percent total accuracy, with response time under a minute
Accuracy above the human baseline and latency below it, on the same task and the same queue.
04
Where it fits
The record supports a narrower claim than a vendor would usually make. The limits on the right are as much a part of the evidence as the results above.
Supported by the record
High volume transactions with a bounded vocabulary that can be named in advance
Workers who spend the shift away from a terminal and cannot stop to open a form
Intake where the realistic alternative is a form that does not get filled in
A first channel that later extends into chat and messaging on the same build, as the grocery chain did
Any flow where a human path remains available behind the agent
Not supported by the record
Warehouse and industrial acoustics. The library holds no proof. Humana proves domain tuned recognition in a call center, which is a different acoustic problem entirely
Open ended judgement work with no defined transaction behind it
Anything that must be right on the first attempt with no escalation path
A signature or a legal record captured by speech alone
Any claim that Hakkoda has delivered a voice build before. The spine here is IBM
05
Why the numbers hold
What decides whether a leadership team acts on a report is the chain between a spoken sentence and a figure on a slide. Every step below is deterministic except the one marked generative, which is why the same question asked twice returns the same answer.
The chain from a spoken sentence to a number a leadership team can repeat. One generative step, four deterministic ones, and a gate that fails closed.
Two behaviors follow. The system declines to answer when the evidence is thin, and the cohort floor means a single outspoken person cannot be reconstructed from the reporting. Both cost coverage, and both are what make the output safe to read out in a room.
06
A voice agent running today
The architecture above is in production on a current engagement at a national pay-TV provider, delivered by IBM Consulting, where a voice agent conducts stakeholder interviews and the findings feed an operating model redesign.
In production
The interviewing agent.Conducts the conversation by speech, transcribes it, and extracts structured findings tied to the exact words spoken. 93 invitations were issued across the in-scope population and 23 conversations have been completed and analyzed. The program remains open.
In production
The answer surface.Leadership asks the evidence base a question in plain language and receives an answer drawn only from de-identified findings, cited to what was said, naming nobody.
In production
Privacy by construction.Individual identifiers never reach the layer the answer is drawn from, so the system has no name available to disclose even when asked directly for one.
In motion
Coverage.One stakeholder tier is still to be interviewed, and the reporting shows that gap on the face of it.
Held
The client name.Described here without attribution pending reference approval. The engagement, the architecture and the figures are accurate as stated.
07
What carries over to a distribution workforce
Five things transfer directly from the record to a frontline intake build.
01
Vocabulary
Recognition trained against named transactions taken from the work, in the words the crews already use for equipment and process.
02
Channel
Speech first for people away from a terminal, with chat and messaging extending from the same build.
03
Grounding
Every reported figure traced to a source record. The gate holds the claim when the trace is missing.
04
Floor
Nothing identifying an individual reaches a display surface, which is what makes candid safety reporting possible at all.
05
Loop
Low confidence conversations are collected and become the scope of the next release.
Open question
Acoustics. No warehouse or yard proof exists in the library. A scoped pilot should precede any commitment on a noisy environment.
By design
The agent abstains when the evidence is absent. Coverage is traded for the ability to defend every line it does produce.
Always
A human path stays behind every flow. Nothing in the record supports removing it, at any volume.
Run one transaction, in one facility, for 30 days.
Choose the report that already gets filed late. Take the vocabulary from how the crews say it out loud. Keep a human path behind it from the first day. Then read take-up at day 30, which is early enough to tell a build that is working from one that needs its vocabulary redone.
Acoustics is the open question.
It is the one claim on this page with nothing behind it. Thirty days on one site settles it, and settles it before anyone has committed to a program.
Attribution. Every client record on this page was delivered by IBM Consulting or by IBM. The proof library holds no Hakkoda-delivered voice or contact center build, and no claim to one appears here.
Proof discipline. Confirmed records carry a headline figure. Projected, estimated and directional records appear as structural rows without one. The unnamed leading US grocery retailer is shown structurally for this reason: its savings figures are labeled projected in the source.
Excluded. Records whose headline figure is projected and whose confirmed portion adds nothing to an adoption argument. Watson-generation deployments dated 2017 to 2020, where technical currency matters to the reader. Any record the source anonymizes and that cannot be described without naming the client.
Pending verification. The Elevance Health figures derive from a public post rather than a case study. The Humana source describes recognition accuracy in self contradictory terms and no accuracy percentage is quoted here. Both should be settled before external release.
Sourced from the Hakkoda and IBM proof library, 157 records, cross checked against the proof point cards and the source of truth index. Figures are quoted as the source states them.