Chief Data Officer / Japan
Decisions that cost revenue, time or pride
- Annual company revenue
- ¥23.8bn
- Year to March 2026; cable, broadband, community channel and resold mobile
- Subscribing households
- 231,000
- Of 520,000 homes passed in Nara and southern Kyoto prefectures
- Data colleagues
- 34
- Engineering, analytics and the privacy and data-governance office
Decisions & consequences
01Opt-in consent for viewing logs, at a quarter of the offered revenue
In 2023 an advertising-technology company offered ¥140 million a year, for three years, for pseudonymised viewing logs from our 118,000 capable set-top boxes. It wanted to combine them with postcode-level segments and aim advertising at those households on its other platforms. Our inherited terms mentioned statistical use but nothing a subscriber would recognise as this arrangement. The offer equalled more than five percent of our operating profit.
Our advertising director wanted to proceed on an opt-out basis. Our lawyers advised that a properly anonymised dataset could be provided with public notice rather than consent, if the processing met the legal standard. I doubted that data detailed enough for household targeting would remain anonymous, and I did not want subscribers to learn of the arrangement from a page on our website. I recommended opt-in consent, aggregate reporting with at least fifty households behind each figure, and no household-level export. One of our audit and supervisory board members asked how a subscriber would withdraw; the president supported my recommendation after that discussion.
We gave up most of the revenue. The partner restructured its offer to ¥1,250 a year for each consenting box and dropped postcode targeting. We wrote the request in large type, tested it with forty subscribers in two community halls, sent it by post and on screen, and trained contact-centre staff to answer questions without encouraging either choice. Withdrawal sits in the remote-control menu.
After nine months, 33,600 boxes had opted in, 28.5% of those eligible. First-year revenue was about ¥39 million. We received 54 complaint calls, mostly about the letter's length, and 1,900 households later withdrew. The consent field now governs every outside use of viewing data, including our own advertising products.
I would do one thing differently. I told the board the likely opt-in rate only after negotiations had begun, and the advertising team felt ambushed. A range agreed in advance would have made the trade-off a shared decision rather than my objection.
Retiring the strongest signal in a churn model I had built
In 2024 our broadband churn model, which I had built in 2020, was allocating a retention discount of ¥1,100 a month for twelve months to the highest-risk tenth of 184,000 subscribers. Planned spending for the year was ¥260 million. Monthly churn stood at 0.82%. A contact-centre supervisor asked me why several households who had complained repeatedly about evening speeds were being offered money rather than a technician.
The review was uncomfortable. The model's strongest signal was the number of fault calls in sixty days. About thirty-one percent of the top decile had made three or more, and most lived on older coaxial segments in hill towns. We were paying discounts to compensate for a fault. The same signal also scored older subscribers, who telephone us more readily, as risky for asking for help.
I proposed splitting the output. Households whose risk came mainly from repeated faults would go first to network operations; only price- and contract-driven risk would receive an offer. The marketing director disagreed, because the change would delay offers during a quarter with a churn objective. Field operations objected to roughly 2,100 additional visits. The president accepted the split, and we missed that quarter's objective: churn reached 0.86% against a planned 0.80%.
Network operations inspected the affected segments and, on their own judgement, brought forward node upgrades in seven areas at a cost of ¥120 million. Over the following twelve months, monthly churn in the repair group fell from 2.7% to 1.5%. Discount spending was ¥158 million. Company-wide churn fell to 0.76%, although I cannot attribute the whole movement to this change.
What I would do differently begins in 2020. I would have asked field technicians and contact-centre supervisors to review the model's inputs before it went live, and I would have measured whether an offer changed behaviour, rather than only whether the model ranked customers accurately.
Limiting an AI call-summary tool to drafting, one centre at a time
In 2025 our head of customer operations proposed AI-drafted call summaries across our three contact centres, which handle about 1.6 million calls a year with 210 agents. The vendor's tool would also classify each call, including whether it contained a complaint. Complaint counts feed our board reporting and our handling of consumer-protection obligations for telecommunications services. The vendor's standard contract allowed it to use recordings to improve its models, and stored data outside Japan.
I supported the summaries but not the contract or the classification. Recordings contain personal details and, at times, information about the content of subscribers' communications. I asked for a clause preventing training on our recordings and for storage in Japan, which raised the three-year cost from ¥96 million to ¥128 million. I also proposed that agents confirm every summary and continue to classify complaints themselves, and that we begin in one centre. The operations head wanted all three centres and automatic classification at once, to capture the time saving sooner. The vendor resisted the clause for two months.
We agreed a pilot. Our quality team audited 400 summaries each month against the recordings. In the pilot centre, after-call work fell from 6.1 to 4.0 minutes. The audit found material errors, such as a wrong appointment date or a request attributed to the wrong person, in 3.4% of summaries, and a complaint expression omitted in 0.7%. That second figure settled the classification question: agents kept it.
We extended the tool to the other centres after five months. Across all three, after-call work now averages 4.3 minutes, which frees roughly 48,000 agent hours a year. The audit continues, and two agent representatives now help select its samples.
Next time I would set the exit criteria before the pilot started, including the error rate that would stop the extension. We argued about the threshold after seeing results, and that made the discussion more personal than it needed to be.