Every conversational build in the proof library, scanned across voice, chat, messaging, agent assist, human resources agents and screening tools. Twenty nine records qualify. Three are Hakkoda delivered. The rest carry IBM or IBM Consulting attribution and are labeled as such throughout.
Internal only
A library scan for enablement, not a client asset. Seventeen organizations appear by name and none of the names have been cleared for external use. One further name sits in section 04 as a reserved slot with no record behind it. Two figures are held for verification against the source decks. An external version carries only cleared names.
The compiled source of truth was queried for voice, speech, interactive voice response, call and contact center, chatbot, virtual agent, conversational, assistant, agent assist, natural language, service desk, omnichannel and messaging. Twenty nine of the one hundred and fifty seven records qualify once incidental matches are removed. The tiers below are ordered by how much speech sits in the actual build.
Tier
Records
What the tier proves
Voice native
2
Speech recognition is inside the build. A caller speaks and the system resolves the intent end to end.
Voice channel, text engine
8
The call arrives by phone. The model works in text behind a deflection, containment or agent assist layer.
Customer facing text
4
Web, application and messaging surfaces. Containment rates here are the highest in the library.
Workforce facing
8
Internal service desks, information technology support and employee human resources support at full workforce scale.
Talent and operations
7
Agents inside hiring workflows, clinical screening, and operational control rooms.
02
How the scan runs
Nothing in this brief was recalled or assembled by hand. The proof cards compile to a single queryable file, three Python scripts read it, and two gates stand between a generated asset and deployment. The same pipeline produced the counts in section 01 and the rows in sections 04 through 07.
Cards and the index compile to one file. Coverage reads it before a build starts, build_pager writes the rows under proof discipline, and lint_asset checks the finished page against the library. The shell lock gate applies to three-pagers, where the frozen shell must hash clean. This brief ran the query, the discipline rules and the client name check. It has no locked shell to verify.
Rule
Enforced by
Effect on this brief
Confirmed takes the number
build_pager.py
Twenty three rows carry a headline figure. Every one is confirmed in the source.
Directional goes structural
build_pager.py
Six rows appear with no figure, marked Directional, Projected or Structural.
Client exists in the library
lint_asset.py
All twenty nine names verified against the compiled file before publication.
Attribution stays with the group
lint_asset.py
Hakkoda delivery and IBM delivery never share a section.
No em dash, no en dash
lint_asset.py
This page is pure ASCII, which the linear brief template requires anyway.
03
Voice against text
The delta between a text build and a voice build is real but narrow, and it is concentrated at the edges of the turn. Four differences carry most of the work.
What a text build already carries
Seam
What voice adds on top
Input
A typed string
Clean tokens, correct on arrival, available for rereading.
Input
Speech recognition tuned to the domain
Humana trained speech on its own provider traffic and reports roughly 95 percent accuracy on trained inputs.
Output
Rendered on screen
Length is cheap. A list, a link and a table all work.
Output
Synthesis under a latency budget
Answers have to be short, ordered and speakable, with no link to fall back on.
Repair
Scrollback
The person can reread the last turn and correct the thread.
Repair
Confirmation in dialogue
Nothing persists on screen, so confirmation and correction have to be designed into the turn.
Escape
Deflection
Handoff to a page, a form or an asynchronous queue.
Escape
Live transfer
The caller is still on the line, so escalation has to hand over context in the moment.
04
What Hakkoda has delivered
Three records, plus one reserved slot. All three are natural language over governed data inside the client platform rather than a voice or contact center build. One carries a confirmed metric. Two are directional and appear as structural rows on purpose, since the source has no hard number to cite.
Mitsui USA
Healthcare, clinical operations
Clinical Trial Waiting Room. Language model pre screening and patient education across an end to end Snowflake and Azure architecture.
2x
Faster patient pre screening. Four times the industry standard build speed.
Medtronic
Medical technology
Snowflake Cortex Analyst across three connected semantic models with a custom procurement front end. Category managers ask in plain language.
Directional
Strong accuracy and user trust across an eight week build. No hard number in the source.
Under Armour
Retail, apparel
Snowflake Intelligence natural language querying for business users, plus language model product description generation.
Directional
Thousands of hours saved annually. Source language is directional with no confirmed count.
DirecTV
Media and entertainment. Reserved
Interviewing agent. Reported as a build that questions a person directly, which nothing else in the library does. No engagement record, architecture, delivering entity or metric exists in the compiled file.
Placeholder
Held for validated data. Described in conversation, not yet documented.
The DirecTV row is a reserved slot, not a proof point. The source of truth index states that DIRECTV appears in no processed source file and must not be listed as a client until a source deck is identified, so nothing can be written about the engagement, the architecture, the delivering entity or any metric until one lands, and the row is removed from any client-facing version of this page. It is held here because of what it is reported to be. An interviewing agent would be the first build in the library to question a person directly, which is the gap named in the headline above. A source deck closes both at once.
05
Voice native and the call channel
Delivered by IBM and IBM Consulting. The first two records have speech inside the build. The remainder work the phone channel with a text engine behind it, which is the shorter distance to a voice deployment than any of these numbers suggest.
Humana
Healthcare, payer
Voice agent on Watson with custom speech training. Eligibility, benefits, claims, authorization and referral sub intents.
7,000
Provider voice calls handled each business day, from 120 providers.
Major US grocery chain
Retail, grocery. Blinded
Cognitive call center virtual agent on Watson. Voice first, then chat and email, across twelve back end systems.
60%
Lower operating cost. 1.8 million interactions a month, a quarter of ten thousand daily calls contained.
E.ON UKS
Energy, utility
Amazon Connect omnichannel contact center delivered as a service. First deployment of its kind in Europe.
80%
Lower operating cost per interaction. Live in twelve months.
Elevance Health
Healthcare, health insurance
Watson Assistant with an asynchronous messaging cloud on hybrid cloud, with healthcare specific tooling.
24%
Of forty six million annual calls moved to digital in under two years. Satisfaction five points above every other channel.
Lumen
Telecommunications
Watson Assistant for self serve chat with Watson Discovery over a dozen knowledge sources. Built on the 2017 to 2020 Watson generation.
70%
Lower cost to serve. Over half of interactions close without a person.
Nuuday
Telecommunications
Agent assist on a local language model across three thousand guides, inside a center handling 4.5 million calls a year.
14%
Lower average talk time. New hire onboarding faster by twenty to thirty percent.
Bouygues Telecom
Telecommunications
Conversation analytics turning eight million annual customer and agent conversations into operating insight.
30%
Reduction in pre and post call work. The yearly cost saving is projected, not realized.
Camping World
Retail, outdoor recreation
watsonx Assistant across every web property with telephone coverage and workflow automation behind it.
33%
Better agent efficiency. Wait times down to thirty three seconds.
CEMIG
Energy, utility
watsonx Assistant with open models and a semantic search layer, unifying messaging, mobile and the web portal.
20 pts
Higher net promoter score. Branch agent handling time down twenty percent.
Marriott Vacations Worldwide
Hospitality, travel
Contact center inside a wider human resources service stack spanning talent acquisition, payroll and employee data.
42%
Lower human resources cost. Live seventeen days after signature.
06
Customer facing and workforce
The same stack pointed at two different populations. Customer facing builds carry the highest containment figures in the library. Workforce builds carry the fastest adoption, because the population is captive and the systems of record are already integrated.
Leading aviation giant
Travel, airline. Blinded
Two generative agents launched together, one for customers across web, application and messaging, one internal. Delivered by Neudesic, an IBM company, on Azure OpenAI.
93%
Automated resolution. Over twenty thousand questions a day.
Unnamed British bank
Financial services. Blinded
watsonx Assistant turning a legacy chatbot into a virtual assistant with a language model classifier, under regulatory oversight.
91%
Of queries answered correctly against a sixty to seventy five percent industry average. The saving is projected.
B2B leader in life sciences
Healthcare, life sciences. Blinded
Conversational generative assistant scoped to high value buying and support intents, iterated against user research.
90%
Lower cost per query. Most product inquiries now resolve automatically.
Availity
Healthcare, interoperability
Over fifty conversational workflows built on interoperability data for behavioral health and specialty disease management.
Structural
Fifty plus workflows in production. No outcome percentage in the source.
IBM
Internal, information technology support
Watson support assistant covering the top two hundred issues, with language auto detection and device lookup.
96%
Of conversations resolved without live support. Eighty eight thousand employees adopted it inside thirty days.
Global professional services company
Professional services. Blinded
Artificial intelligence layered onto a tiered human resources service desk, simplifying inquiry handling and transactions.
34%
Less volume requiring a person. Escalations on an employee behalf down fifty eight percent.
WINDTRE
Telecommunications
Watson intelligent automation across the information technology service desk, triaging in three modes from full resolution to routing.
10x
Faster response. Over ten thousand reports handled a month.
Air Canada
Travel, airline
Bluebuddy virtual agent for incident triage inside an operations framework, with watsonx.ai and an open weight model reasoning behind it.
$12.5M
Saved over five years. Incidents down forty five percent.
IBM Chief Information Security Office
Internal, risk and security
Self serve chatbot fronting an autonomous risk governance orchestrator across classification, due diligence and monitoring.
54%
Cycle time reduction. Four thousand suppliers monitored continuously.
North American internet service provider
Telecommunications. Blinded
Chat enabled generative assistants by role and domain across the software development lifecycle.
30%
Effort saved in testing. Twenty percent saved in analysis.
American healthcare and insurance company
Healthcare, payer. Blinded
Custom assistants built on the consulting platform for microservice creation, code and test script generation.
20 to 30%
Reduction in effort on microservice creation.
Leading US grocery retailer
Retail, grocery. Blinded
Employee virtual assistant on watsonx Assistant and Discovery, giving four hundred thousand frontline staff round the clock human resources support.
Projected
Every saving figure in the source is labeled projected. Structural until a realized number exists.
07
Talent and operations
The closest thing in the library to an interviewing agent sits here, and it does not interview. Providence guides the hiring manager rather than the candidate. Nothing in the library is an artificial interviewer, a digital persona, or an assessment agent that questions a person and scores the answers.
Providence Health
Healthcare, provider
RITA, a named agent inside Oracle Recruiting Cloud on watsonx, guiding hiring managers from requisition to selection with compliance built into the path.
$16.9M
Saved over four years. Ninety percent time savings on manager transactions.
Swiss International Air Lines
Travel, airline
HYVE, a conversational and collective learning system for the operations control center, contextualizing past cases against live inputs.
87%
Faster analysis during irregular operations. Yearly savings given as an estimated range.
Global pharmaceutical leader
Healthcare, pharmaceutical. Blinded
Conversational interface for recruiter and candidate engagement over talent platform optimizers, learning continuously from feedback.
Directional
The productivity gain in the source is labeled expected. Faster time to hire is directional.
Austrian Federal Ministry for National Defence
Government, defense
A watsonx.ai natural language interface onto a central documentation database, built to a human in the loop principle.
Structural
No numeric outcome exists in the source.
Sourced from source_of_truth_compiled.json, one hundred and fifty seven records, cross checked against the proof point cards and the source of truth index. Record counts differ from the two hundred and fifty one cited in the project instructions, which should be reconciled. Blinded clients appear as the source describes them. Directional, projected and estimated records carry no headline number by rule.