AI for journalism that still pays for itself

Problem

Cancellations cluster in the first 90 days and in the week before a promotional price step, but circulation only finds out when the cancel form is submitted. Save offers then go out blanket, discounting subscribers who would have renewed at list price anyway.

Approach

A discrete-time hazard model on subscriber-day rows — LightGBM, or a Cox/DeepSurv variant where the tenure curve carries the signal — scored daily against days-to-renewal rather than calendar month. Billing state from Piano, Zuora or Omeda (tenure, plan, promo step and its size, payment method, prior save history, grace-period events, and the January 2026 euro redenomination price step, which is still a live churn event in every Bulgarian subscriber base) joins behaviour from Snowplow, Chartbeat or Deep.BI (RFV, sessions over trailing 7/28/90 days, topic breadth across IPTC tags, app versus mobile web, push opt-in) and newsletter click rate rather than open rate, because Apple Mail Privacy Protection broke opens in September 2021. A second, separate uplift model — two-model or causal forest — trained on historical randomised offer exposures decides who is worth an offer. Orchestration through Braze or Sailthru, with a permanent 5–10% untreated holdout.

Problem

A fixed five-article meter is simultaneously too tight on high-ad-value traffic — killing RPM on a story that will never convert anyone — and too loose on the investigation people would have paid for. Editorial and commercial argue about it every month with no shared number.

Approach

Two models feeding one decision layer. Model A estimates expected ad revenue per pageview from Google Ad Manager Data Transfer files at impression level, segmented by placement, geo, device, referrer and consent state — the TC String and Consent Mode v2 together, because a consented TC String with Consent Mode denied prices nothing like a fully consented request. Model B estimates expected subscription value as likelihood-to-subscribe times offer price times expected survival. The decision layer in Piano Composer, Zuora Zephr or Poool then picks free, meter, regwall or hard wall per request, and executes at the CDN edge — a Cloudflare Worker or Fastly Compute with a server-side entitlement check — so walled body text never enters the DOM. That stops scraping and removes the layout shift a client-side wall causes, which is a Core Web Vitals decision as much as a revenue one.

Problem

A fifth to two fifths of gross churn is a failed payment rather than a decision. Recurring rebills are merchant-initiated and sit outside SCA once the mandate is authenticated, so the PSD2 damage lands somewhere else — at initial checkout, on any amount or schedule change that forces re-authentication, and on issuer-side soft declines where an MIT exemption is not honoured. Add expired and reissued cards and issuer risk rules, and subscriptions nobody intended to cancel die silently, with the standard three-retry ladder recovering about half of them.

Approach

Enable card account updater services — Visa VAU and Mastercard ABU — through Stripe or Adyen, and move to network tokens so a reissued card keeps working. Pre-dunning at T-30 on known card expiry, which averages about a 73% open rate against anything sent after a decline. Then a retry-timing model predicting the hour and weekday with the highest authorisation probability, conditioned on issuer BIN, decline code (a soft insufficient-funds behaves nothing like a hard do-not-honour), amount and prior retry outcomes. The ladder differentiates by LTV decile — a ten-year print-and-digital subscriber gets a phone call, a one-euro trial gets two emails — and SEPA Direct Debit becomes the offered fallback for long-tenure subscribers — 3–5% returns against roughly 15% authorisation failure on cards, though the two are measured differently and SDD failures surface days later. The trade you are making is a lower failure rate against SEPA Core’s unconditional eight-week refund right, a thirteen-month window on unauthorised claims and pre-notification on every amount change, which is why we offer it to long-tenure subscribers on stable pricing and never to trial or promotional cohorts.

Problem

Returns of 15–35% of draw burn paper, ink, plates, inserts and haulage on copies that were pulped. Allocations were set years ago and adjusted by hand, wholesaler EPOS arrives up to ten days after the sale, and an outlet that sold out is recorded as having sold exactly its allocation.

Approach

Formulated as a newsvendor problem per outlet, title and day, ordering at the critical fractile cu/(cu+co) — cu the contribution lost on an unmet sale, including the advertising and insert contribution the copy carried and not only cover-price margin; co the marginal print-and-haulage cost net of wholesaler credit and recovered waste-paper value. Both are finance inputs, agreed with your circulation director in writing before the first model runs, because an underspecified fractile is confidently wrong. The optimiser then runs under a hard floor on audited circulation per title — a returns saving that costs you rate base is a loss — alongside bundle sizes, inserter capacity and van route cut-offs. Demand from gradient boosting or a Temporal Fusion Transformer on day of week, front-page topic taken from the IPTC tags on the flat plan, cover mounts and competitor promotions, school and public holidays, weather, local events and outlet type. The step almost everyone skips is censoring correction: sold-out outlets must have latent demand imputed, Tobit-style or through an EM loop, or the model learns your supply history and confidently recommends what you already print. Output lands in ppi Media or the circulation system before the plate deadline rather than after it.

Problem

Twenty years of archive tagged with free-text keywords typed by whoever filed the story. Newsletter segments are built on URL path regexes, recommendations recycle the same three sections, seller-defined audience cohorts cannot be constructed, and no AI licensing conversation can describe what the corpus actually contains.

Approach

Hierarchical classification into IPTC Media Topics — 1,200+ terms across five levels — by embedding retrieval over the term definitions followed by LLM verification of the shortlist. Never free generation, which invents terms the vocabulary does not contain. Named entities link to a publisher-maintained authority file keyed to Wikidata and ISNI, so two spellings of the same politician resolve to one node. Confidence thresholds differ by level: auto-apply at levels 1–3, queue levels 4–5 for one-click editorial confirmation, because sibling terms deep in the tree are where precision collapses. Tags write to NewsML-G2 or ninjs in the CMS and to schema.org NewsArticle about/mentions on the page. Bulgarian needs a maintained label layer — IPTC publishes Media Topics translations in over a dozen languages and Bulgarian is not among them.

Problem

Half of print volume is usually home-delivered subscription copies, where there are no returns at all and the levers are different ones: route density, carrier cost per copy, vacation holds and redirects, delivery complaints. Cost per copy rises fastest on the thinnest routes, and nobody can say which postcodes have stopped paying or what withdrawing from them would cost in cancellations — so the decision gets argued in the board meeting instead of priced.

Approach

Route-level contribution per copy built from the delivery roster rather than from a district average: carrier payments, fuel and depot cost against the subscription contribution actually carried on that route. Then a delivery-complaint-to-cancellation hazard, estimated from the complaints log joined to the cancel log, so a missed delivery carries a churn number rather than an apology. Those two feed a consolidation and withdrawal optimiser with a modelled churn cost per withdrawn address, and a mail or digital-only substitution offer priced into each option rather than assumed to convert.