160,000 Letters Sent: What Our Response-Prediction Model Learned About UK Property Owner Outreach

Direct-to-owner outreach is a numbers game. Industry-typical response rates are low single digits. A 5% lift in response rate is the difference between a profitable campaign and a break-even one.

Three years of campaigns through the GalimAI pipeline have generated over 160,000 sent letters with response outcomes captured at the individual send level. That dataset is the training set for our response-prediction model.

What the model predicts

For any planned send (a defined director profile, a defined letter template, a defined send month), the model returns:

• A predicted response probability.
• The expected response type (callback, reply card return, email).
• A confidence band that widens for unusual combinations not well-represented in training.

The campaign planner uses the model to pre-allocate send volume. High-confidence segments get prioritised. Low-confidence segments either get a smaller pilot send or get held back until the model has more data.

The training data

Every letter sent through the pipeline since campaign tracking began is in the training set. For each send, the captured fields include:

• Recipient features: director age band, ownership structure, region, portfolio size, sector mix, charge profile, hold period, late-filing status, Gazette signals.
• Letter features: template ID, copy variant, envelope type, paper stock, signature style, personalisation level.
• Timing features: send month, day-of-week, days since prior contact, days since prior known transaction.
• Outcome: response within a defined window (28 and 90 days), response type, downstream conversion.

That gives the model a rich feature space for each send, and a binary outcome to learn against. Standard classification problem, large training set, well-suited to gradient-boosted trees.

What we learned, qualitatively

We do not publish specific response rates publicly (campaign economics are a competitive matter), but the directional patterns the model surfaces are worth describing because they cut against industry intuition.

Letter copy

Industry intuition: long, friendly, emotionally warm letters perform best. The model says no. Shorter, more direct letters that reference a specific public signal ("we noticed your property at [postcode] has been held for X years...") consistently outperform long, generic warmth-based templates. The lift is largest for older director cohorts.

Send timing

Industry intuition: avoid summer, avoid Christmas, weekdays only. The model partially agrees. Christmas week and August work poorly across most segments. But Tuesday and Wednesday sends do not outperform Friday sends as much as conventional wisdom suggests, and certain regional cohorts respond better to month-end sends than month-start.

Frequency of contact

Industry intuition: one letter then leave alone for 12 months. The model says it depends. Some director profiles respond better to a single well-crafted letter; others respond materially better to a two-letter sequence with 6 to 8 weeks between sends. The model has learned which is which.

Personalisation

Industry intuition: more personalisation = better response. The model says yes, but specifically. Personalisation on property details (address, hold period) consistently lifts response. Personalisation on director details (name, age) lifts response for some segments and lowers it for others (some directors find age-referencing letters intrusive). The model has learned the segment-specific calibration.

The bandit layer

Sitting on top of the response-prediction model is a multi-armed bandit that handles exploration vs exploitation. The bandit reserves a small fraction of every campaign's send volume for less-confident template-segment combinations, so the model continuously learns about under-explored cells of the matrix.

This matters because the matrix of (recipient profile x letter template x timing) has hundreds of cells and not all are well-explored. Without active exploration, the model would converge on the cells it already knows about and miss better-performing combinations the data has not yet revealed.

The bandit is conservative. Exploration sends are a single-digit percentage of any campaign, never the bulk. The bulk of every campaign goes to high-confidence, high-predicted-response combinations.

Where the model genuinely changes the economics

Three places the response-prediction layer materially changes the unit economics of UK property direct-mail outreach:

1. Volume allocation. Instead of sending 10,000 letters evenly across a list, the model concentrates 70%+ of volume on the segments predicted to respond best. Same total spend, materially better total response.
2. Template selection per segment. One letter template fits no segment. The model selects from a library of templates per send, matching template to recipient. Lift is largest where conventional wisdom about copy is wrong (see above).
3. Send-window calendaring. Campaigns get sequenced to land in the highest-response windows for each segment, with the bandit reserving smaller volumes for less-tested windows.

What this is not

The response-prediction model is not a replacement for a good letter. Bad copy will tank response no matter how well-targeted the send. The model assumes a library of competently-written letters and selects between them. Building that library is a creative-and-copywriting problem, not a machine-learning problem.

The model is also not a magic-response generator. Industry-typical response rates for cold direct mail to UK property owners are low single digits. The model meaningfully lifts that, but it does not turn 2% into 20%. It turns 2% into a defensibly higher number, and it does so with the same per-letter cost.

How GalimAI customers use it

Customers running outreach campaigns through the portal see two model outputs:

• Per-send predicted response probability, so they can prioritise their list before sending.
• Per-segment recommended template, so they pick the highest-predicted-response copy for each part of their list.

Customers running their own outreach (not through our pipeline) can still pull the prioritised list and supply their own copy. They get the targeting lift but not the template-selection lift.