Embedded insurance revenue is forecastable from five variables you mostly already track. How to build the model, where each input comes from, and where forecasts usually break.
Embedded insurance revenue looks exotic to a finance team the first time it appears in planning — commission on a regulated product, priced by a third party, with refund mechanics attached. It is actually one of the more forecastable revenue lines a platform can add, because it is a chain of five multiplied variables, and the platform already tracks the biggest one: its own transaction volume. This guide sets out the model, where each input should come from, and the three places forecasts most often break.
One framing note before the mechanics: forecast your commission revenue, never gross written premium. GWP is the insurer's number. Building a platform forecast on GWP guarantees a later, awkward conversion to the money you actually keep.
The five-variable chain
Per product line, per period:
- Transaction volume: the platform's own forecast of relevant transactions or journeys.
- Eligibility rate: the share of those transactions where an offer can be shown — right customer segment, right product, quote service returning a price.
- Attach rate: the share of eligible transactions that convert to a bound, paid policy.
- Average premium: the mean price of the policies actually sold — a mix-weighted figure, not the cheapest product on the shelf.
- Commission rate: the platform's contracted share of premium.
Multiply the five and you have gross commission; deduct givebacks (cancellations, cooling-off refunds and their clawbacks) to reach net commission, which is the line that belongs in the plan. A deliberately illustrative example, with invented numbers chosen for round arithmetic: 100,000 monthly transactions, 60% eligible, 5% attach, an average premium of 400 and a 15% commission rate gives 3,000 policies and 180,000 in gross commission a month; a 5% giveback rate takes net commission to 171,000. None of those values are benchmarks — the structure is the point.
Where each input should come from
- Volume is your number. Take it from the existing business plan so insurance inherits the same seasonality and growth assumptions as the core business — a separate volume assumption inside the insurance model is a consistency bug waiting to be found in review.
- Eligibility comes from engineering and the infrastructure partner: scope rules plus observed quote success rates. New programmes routinely overestimate it; real-world data gaps and quote declines bite harder than scoping documents admit.
- Attach rate is the assumption that deserves the most scrutiny. Before launch, take ranges from your infrastructure partner's comparable journeys and treat them as hypotheses. After launch, your own cohort data replaces everything — and early attach usually shifts as placement and pre-fill are tuned, so re-forecast quarterly in year one.
- Average premium belongs to the insurer's rating, so source it from quoted-price distributions, not brochure prices, and assume it moves: panels reprice, and product mix drifts toward or away from cheaper covers as the journey evolves. A flat premium assumption over a multi-year model is a silent error.
- Commission rate is contractual — the easy input. Model any volume tiers or profit-linked components explicitly rather than as a blended guess.
Recognition, timing and the renewal layer
Three timing questions to settle with your accountants and your partner early, because they change the shape of the line even when the totals agree.
First, recognition. Commission on a single-payment annual policy is typically recognised at or near binding, but treatment varies with the contract's cancellation and clawback mechanics — get a ruling from your auditors rather than assuming. Second, cash. Commission statements and settlement commonly run on a lag; the model should carry both a revenue line and a cash line, and the working-capital gap between them. Third, renewals. From year two, revenue splits into a new-business layer and a renewal layer with its own retention rate — and the renewal layer compounds. Model them separately from the start: a single blended growth rate hides the fact that renewal revenue is cheaper, steadier and entirely dependent on a retention assumption you should be testing.
Scenarios, not a point estimate
A five-variable multiplication is exactly the structure where small per-variable optimism compounds into a fantasy. Two defences work. Build low, base and high cases by flexing the two genuinely uncertain variables — attach and eligibility — while holding contractual ones firm. And run the model backwards as a sanity check: given the revenue the plan wants, what attach rate does it imply, and would anyone defend that number in the open? An implied-attach test kills more bad insurance forecasts than any amount of cell-by-cell review.
Treat repricing as a scenario too. The premium input is set by insurers responding to claims experience; a case where average premium falls while attach holds is worth a line in the model, because falling premiums cut commission even in a quarter where every operational metric improved.
Model-review checklist
- Forecast is net commission, not GWP, with givebacks modelled explicitly.
- Volume inherited from the corporate plan, not re-assumed.
- Eligibility grounded in observed quote success, with a haircut in year one.
- Attach sourced from comparable journeys pre-launch and replaced by cohort actuals on a set cadence.
- Average premium taken from quote distributions, with a repricing scenario.
- Commission tiers and clawback mechanics modelled per the contract.
- Revenue and cash shown separately, with the settlement lag stated.
- New business and renewals as separate layers, each with named assumptions.
- Implied attach rate of the high case reviewed and defended.
- Per-policy reporting from the partner reconciled monthly against forecast actuals.
The limits of the model are worth stating plainly. It forecasts a distribution commission, so it cannot see insurer-side shocks — a carrier exiting a line, a regulatory change to commission structures, a repricing cycle after a bad claims year — except as scenarios. And before launch, every attach assumption is borrowed. The honest pre-launch forecast is a range with a review date, and the honest year-two forecast is built on your own cohorts. If those two sentences survive into the board version of the model, it is a good model.