A voice agent runs the interview. What people say becomes findings that anyone can ask questions of in plain English, with no names attached. It is live now at DIRECTV. This covers what was built, the wider record behind it, and what a first build at US Foods would look like.
01
Where it was built
DIRECTV asked IBM Consulting for an assessment of how its engineering organization should run with AI in it. Six weeks, fixed fee, May to June 2026. Six artifacts at the end: a current-state map, a target operating model, role specifications, a standards handbook, a hiring plan and a management handoff plan.
All six had to rest on what the organization actually said, which meant 93 stakeholder interviews inside six weeks. Booking, running and reading 93 conversations by hand does not fit in that window. So the interview itself was built as a voice agent, and the reading of it was built as a layer underneath.
The engagement
Detail
Why it matters here
Client and scope
DIRECTV CIO organization
Audit and redesign of the engineering operating model, delivered by IBM Consulting.
Evidence gathered
93 invitations, 26 done
Process owners and senior leadership so far, with the executive tier still to come. The program is open and the reporting shows the gap rather than averaging over it.
What came out
442 findings
Each one tied to the sentence that produced it. Three signals surfaced in nearly every conversation: knowledge is hard to find, AI is already in the work, and the handoff lacks clarity.
What it fed
six deliverables
Every artifact in the engagement traces back to this evidence base, which is what made the findings defensible in front of the client.
The system described in the rest of this document is that build. Nothing here is a concept.
02
What it is
Part
Job
Built on
Alex, the voice agent
runs the interview
LiveKit handles the call, Deepgram turns speech into text, ElevenLabs gives it a voice, and Claude Sonnet 4.6 decides what to ask next.
The findings layer
reads the transcripts
Each transcript is broken into findings. Every finding keeps the exact sentence it came from. Names are removed here and go no further.
The answer surface
answers questions
A question in plain English comes back answered, with the lines it was built from. Claude Opus 4.8 writes it over a search of the findings.
All of it runs inside Snowflake. No interview data is sent to an outside service.
03
How it works
Three rules hold at the end of the line. If nothing matches, it says it does not know. If fewer than four people said it, nothing is shown. Asked who said something, it has no name to give.
04
What was built
Running
The voice agent.Alex runs the interview end to end.
Running
The interviews.93 invitations out. 26 people completed one and had it analyzed. More are booked.
Running
The evidence base.442 findings, each tied to the sentence it came from.
Running
The answer surface.Questions in plain English, answers with sources, no names. Tested against attempts to make it name someone, and it holds.
05
The wider record
29 conversational builds sit behind this one. Three were delivered by Hakkoda, 26 by IBM or IBM Consulting. Grouped by what the build actually does.
Speech in the build
Client
What it does
7,000 calls a day
Humana
Takes provider calls on coverage, claims, authorizations and referrals, start to finish. 120 provider organizations.
1.8M a month
A major US grocery chain
Customer care. Handles 25 to 30 percent of 10,000 daily calls, then extended from the phone into chat and email.
Voice channel, text engine
Client
What it does
24% of 46M calls
Elevance Health
Moved to digital in under two years. Satisfaction runs 5 points above every other channel.
80% lower
E.ON UKS
Cost per interaction, on a contact center rebuilt and live inside 12 months.
40% daily use
Nuuday
An assistant call center staff can ignore, and do not. Calls run 14 percent shorter when they use it.
8M conversations
Bouygues Telecom
Turned into something the business can act on. Work either side of the call is down 30 percent.
65% retention
CEMIG
On the messaging channel, with 1.1 million more messages a month and a 20 point gain in net promoter score.
33% better
Camping World
Agent efficiency, with waits down to 33 seconds and engagement up 40 percent.
42% lower
Marriott Vacations Worldwide
Cost of the HR function. Live 17 days after the contract was signed.
100% of chat
Lumen
Carried by the assistant, with over half of all interactions needing no person. Built on the 2017 to 2020 Watson generation.
Customer facing, text
Client
What it does
20,000 a day
A leading aviation giant
Questions answered, 93 percent of them without a person touching it.
91% correct
A stalwart British bank
Customer queries answered accurately, against a 60 to 75 percent industry average.
90% lower
A global life sciences leader
Cost per query, with 60 to 70 percent of product questions resolved automatically.
50+ workflows
Availity
Conversational workflows built for a healthcare data exchange. No outcome figure on file.
Workforce facing
Client
What it does
88,000 in 30 days
IBM internal IT support
Employees moved onto a conversational front door. 96 percent of conversations finish without a person.
58% fewer
A global professional services firm
HR partners calling on behalf of staff, and 34 percent less volume needing a human at all.
10,000 a month
WINDTRE
Reports handled, with responses coming back ten times faster than before.
45% fewer
Air Canada
Incidents, and 12.5 million dollars saved over five years on IT operations.
54% shorter
IBM security office
Risk assessment cycle time, while request volume rose 15 percent and staffing fell.
30% less
A North American internet provider
Testing effort across the delivery cycle.
20 to 30% less
An American health insurer
Effort building new services, with the assistant doing the first pass.
Structural
A leading US grocery retailer
Two agents serving 400,000 employees, 80 percent of them frontline with no regular system access. The savings in this record are projected, so no figure is shown.
Talent and operations
Client
What it does
16.9M dollars
Providence
Saved over four years by a named agent that walks hiring managers through the process in dialogue. 90 percent time saved per transaction.
87% faster
Swiss International Air Lines
Analysis during irregular operations, so crews get answers while the disruption is live.
Structural
A global pharmaceutical leader
Recruiting assistant. The productivity gain in the source is expected, not measured.
Structural
Austrian Ministry of Defence
Documentation made findable across organizations. No numeric outcome in the source.
Hakkoda, plain language over governed data
Client
What it does
2x faster
Mitsui USA
Patient pre-screening for trial feasibility, built four times faster than the industry norm.
Structural
Medtronic
Category managers ask procurement questions in plain language across three connected models. Eight week build, no hard figure in the source.
Structural
Under Armour
Plain language querying for business users. The source says thousands of hours saved, without a count.
The three Hakkoda builds are text, not speech. The voice agent described above is the first of its kind here, and it runs on the same idea: ask a question in plain language, get an answer traced back to governed data.
06
What this means for US Foods
The build described above was aimed at one subject, an engineering organization talking about how it works. The same three parts point at any subject where the evidence lives in what people say and nobody has time to collect it by hand. Frontline incident reporting is that shape.
US Foods is already a Hakkoda account. Two engagements, 28,000 employees, 250,000 customers and 60 locations served, with 2x payback on the program by year two. All confirmed. A voice build would be an expansion of that work.
The question US Foods will ask
Answer
What the record shows
Integration
Can a voice agent reach the systems we already run
The ChefS discussion.
A major US grocery chain
Connected to 12 back-end systems
Grocery, real voice, and 1.8 million interactions a month at 60 percent lower operating cost.
Vocabulary
Will it understand how crews describe equipment and incidents
Trade words, abbreviations, and the names people actually use.
Humana
Trained against five named transactions, not an industry
7,000 voice calls a business day from 120 provider organizations.
Workforce
Will people who are never at a desk use it
Drivers, warehouse crews, shift staff.
A leading US grocery retailer
400,000 employees, 80 percent frontline with no regular system access
The closest workforce shape in the library. Its savings figures are projected, so only the shape is usable.
What happens next
Does anything useful come out the other side
Safety review, and the Snowflake dashboard.
Bouygues Telecom
8 million conversations turned into something the business can act on
Work either side of the call down 30 percent. The dollar saving in that record is projected.
One question has no answer in the record. Nothing in 157 records covers speech recognition in warehouse or yard noise. Humana proves specialist vocabulary in a call center, which is a quieter environment. A scoped pilot settles it in weeks and costs a fraction of a commitment.
01
One report type
Start with the report that already gets filed late, or not at all.
02
One site
A single facility, so the acoustic question gets a real answer rather than a guess.
03
Their words
Vocabulary taken from how the crews describe the work out loud.
04
A human path
Someone to escalate to, behind every flow, from the first day.
05
30 days
Read take-up at day 30. That is early enough to tell a working build from one needing its vocabulary redone.
Build figures read from the running system, June 2026. Client records come from the Hakkoda and IBM proof library, 157 records, quoted as the source states them. Projected figures are shown as projected and never as results. DIRECTV is named here because this document is internal. Reference approval has not been requested, so any client-facing version describes the engagement as a national pay-TV provider.