How scoring decides who gets asked

The intelligence engine turns donation data into an answer to one question: who should you ask next, for how much, and by when? This is the logic behind that answer.

Together's scoring model isn't a forecast. It's an opinionated triage system: given the data on your donor base, here are the highest-value conversations to have this week. This page explains the reasoning so you can trust the recommendations.

RFM is the substrate

Recency, Frequency, Monetary value: three signals every donor record carries naturally, no enrichment required. RFM is fifty years old as a marketing technique because it works for a simple reason: a donor who gave a lot recently and often is more likely to give again than one who gave once a long time ago. Together computes RFM as quintiles within your org so the scores are comparative, not absolute.

Comparative scoring matters because absolute thresholds drift. "Recent" in a hyper-active campaign month is different from "recent" in a quiet quarter. Building thresholds into the model would fight your reality every time the campaign rhythm changed. Quintiles self-adjust.

Upgrade and lapse are predictions, sort of

Beyond RFM, Together computes two propensity signals. Both are rules-based, not learned: they use simple, defensible heuristics that you could explain to a board.

Upgrade signal asks: based on this donor's recent behaviour, are they trending up? Indicators include a recent increase in gift size, multiple gifts in a short window, conversion from one-off to recurring, and engagement with non-donation touch points (where Together has access to that data). A score above 0.6 means several indicators are firing at once.

Lapse risk asks: based on the gap since the donor's last gift, and the segment they previously sat in, are they on track to stop? A Champion who hasn't given in a year has higher lapse risk than a Lapsed donor (who's already gone). The model gives the heads-up before the cliff edge, not after.

Neither signal is a probability in the Bayesian sense. They're scores you can rank by. Treat 0.6 / 0.5 as triage thresholds, not lines on a probability axis.

Segments turn numbers into actions

Numbers are hard to act on. "Donor X has RFM=82, lapse=0.41" is a fact, not a call to action. Segments translate the numbers into named groups that map to plays your team already runs: Champions get a major-gift conversation; About to Lapse get a retention call; Promising get a welcome sequence.

The segment names follow a customer-lifecycle pattern that works for donor behaviour but isn't identical to the marketing literature. Together doesn't use "Whales" and "Big Spenders"; donors aren't a slot machine. The labels are chosen to match how fundraising teams already think.

Ask amounts are a separate layer

The score tells you whether to ask. The ask amount tells you for how much. Together computes a suggested ask per donor using per-segment rules:

• Champion + strong upgrade signal: last gift x 1.15. They're already engaged and trending up.
• Potential Loyalist: last gift x 1.10. A gentle bump as they consolidate the relationship.
• At Risk: the smaller of their last gift or their personal average. Don't ask up when they're sliding.
• Lapsed: the smaller of 0.75 x their average or 0.8 x the segment median. Re-engagement asks beat retention rates.
• Floor: all asks clamp to a minimum of $5 to avoid suggesting trivially small amounts.

The result is a single number per donor per pass, recomputed on every full pipeline run.

What the model intentionally does not do

A handful of decisions are baked in by design:

• No black-box ML. Every score can be traced to a specific rule. If your compliance lead asks why donor X got a particular ask amount, you can show the working. Black-box models lose this property.
• No cross-org learning. Your data scores your donors. Together doesn't fold other orgs' patterns into your scores. Some accuracy is lost; explainability and privacy are gained.
• No personal forecasts. Propensity scores are population-level patterns. A donor with lapse risk 0.7 isn't 70% likely to lapse; they're sitting in a group where lapsing is more common. Don't treat individuals deterministically.
• No actions taken automatically. Together computes; your team decides. The Raise tier adds automated outreach, but even there the rules are explicit and adjustable.

When the model gets it wrong

The model will sometimes be wrong about an individual donor. Two specific cases come up often:

• A donor about to give a big one-off can read as Lapsed right up until the donation lands. There's no signal in the data that distinguishes them. Trust your team's relationship knowledge over the score in these cases.
• A donor who'll never give again but hasn't gotten round to telling you can read as Champion for months before lapsing. The score reflects history, not intent; relationship cues catch what the score can't.

The right posture is: scores set the priority order; you and your team decide the actual conversations. Disagreeing with the model on a specific donor is fine; designing without it is leaving signal on the table.

What to do next

• Read the model fields in detail: The donor scoring model.
• Activate intelligence if you haven't yet: Activate intelligence.