How Much Does Ecommerce Fraud and Chargeback Software Cost in 2026?

The bands a fraud and disputes build falls into

Two bands, and the split is whether the system decides and files, or whether it also learns and polices behaviour.

• $65,000 to $140,000, twelve to sixteen weeks. A first release: a decision engine trained on your own order history, an editable rules layer your analysts can change without a deployment, a review queue with proper case handling, and automated representment packs for your top dispute reason codes on one acquirer.
• $160,000 to $400,000, six to twelve months. The full platform: retraining pipelines with correct handling of delayed labelling, policy abuse detection with a graduated response ladder, issuer specific evidence templates, a second and third acquirer, European traffic under its different liability picture, network monitoring alerts and a merchant analytics view.

Volume is a poor predictor of price here. Payment provider count, region count and label quality set the number.

What drives a fraud build up

Label quality is the first driver, as above. Without chargeback outcomes linked to originating orders you have nothing to train on, and reconstructing that linkage from dispute files, settlement reports and order exports is a real piece of work that has to happen before modelling starts.

Acquirer and payment provider count is the second. Adyen, Stripe, Braintree, Worldpay and Chase Paymentech expose disputes differently, with different evidence field limits, different submission windows and webhooks that disagree about state. Experience with one does not transfer, so each is its own integration rather than a variant.

Multi region is the third. European traffic under strong customer authentication carries a different liability picture from card absent traffic elsewhere, which means different logic rather than a different threshold. If you sell into both, that is two decision paths, not one with a country flag.

Marketplace or multi seller structures are the fourth and they are a design problem before they are a cost. Liability allocation between platform and seller has to be modelled explicitly, and getting it wrong means disputes land on the wrong balance sheet.

Explainability is the fifth and it is not optional. Every decline needs a human readable reason, both for your support team when a customer calls and for the analyst tuning rules. A model that outputs a score and nothing else gets overridden into uselessness within a quarter, which means you paid for a model and are running on overrides.

What keeps the number down

Start with one region, one payment provider and your two highest volume dispute reason codes. Those cover the majority of disputed value, and they prove the model before you widen the surface. Everything after that is repetition of a pattern rather than new design.

Fix your labels yourself if you can. Writing chargeback outcomes back to originating orders is a data engineering task, and if your own team can run it against your warehouse it costs you internal time rather than agency rate. That single step can move a first release from the top of its band back into the middle.

Run the decision engine in shadow mode alongside your existing vendor before you switch. It scores every order, files nothing, and after a month you can compare its decisions against the incumbent's on real traffic. That month costs no development and it is the strongest evidence you will get either way.

Defer policy abuse detection. It is the second most expensive component and it reads from a feature store that phase one builds. It also needs a graduated response ladder agreed with your customer service and merchandising teams, which is a policy conversation rather than an engineering one and takes its own calendar.

A worked example that adds up

A retailer processing roughly $180M online across two regions and two acquirers, currently on a guarantee vendor, with three years of order history but no linkage between disputes and the orders that produced them.

• Discovery, label audit and dispute to order linkage assessment: $10,000
• Label reconstruction: writing historical chargeback outcomes back to originating orders: $22,000
• Feature store over order, return, delivery and support history: $22,000
• Decision engine with expected value thresholds by category, fulfilment method and customer tenure: $34,000
• Editable rules layer with human readable decline reasons: $20,000
• Review queue with case handling and mandatory written override reasons: $14,000
• First acquirer dispute integration: $18,000
• Automated representment packs for the top two reason codes: $22,000
• Second acquirer dispute integration: $16,000
• Issuer specific evidence templates held as editable configuration: $22,000
• Retraining pipeline handling delayed labelling and above threshold approval sampling: $34,000
• Policy abuse detection with a graduated response ladder: $38,000
• European traffic handling under strong customer authentication liability rules: $24,000
• Merchant analytics: approval rate by segment published to the commercial team: $16,000
• Card network monitoring alerts: $12,000

That totals $324,000 across nine months. The first release, meaning lines one through eight, came to $162,000, which is above the $65,000 to $140,000 band, and the entire overage is the $22,000 label reconstruction line. A retailer whose dispute outcomes were already written back to orders would have paid $140,000 for exactly the same release and started modelling three weeks earlier. That is the whole point about labels made concrete.

How the spend phases

Discovery first and separately, and its main output is an honest answer about your data. Can you link a chargeback to the order that produced it? Do you have returns, refunds and delivery outcomes joined to the customer? Do you have twelve to twenty four months of it? If any answer is no, that is the first phase and it should be priced as its own piece of work rather than absorbed into a modelling estimate.

Then the feature store, decision engine and rules layer together, running in shadow mode against live traffic while your existing vendor still decides. A month of shadow running produces a direct comparison on your own orders, which is worth more than any vendor benchmark.

Representment goes live next and independently, because it does not depend on the decision engine at all. Automating disputes for your top two reason codes on one acquirer delivers recovery immediately and is the easiest part of the programme to justify.

Retraining, policy abuse and additional regions follow once the first engine has produced several months of its own outcomes. Hold back ten to fifteen percent for the period after you take the decision away from the incumbent, because that transition is where the surprises live.

The ongoing costs nobody quotes

Analyst time is the largest ongoing cost and it does not go away. Somebody works the review queue, tunes rules, reviews overrides and investigates the policy abuse cases the system surfaces. Good software makes that person far more effective. It does not remove them, and a budget that assumes it does will be wrong within a quarter.

The line unique to this category is the deliberate cost of correcting selection bias. Orders you decline have no outcome, so your training set only ever contains orders you approved, which biases every future model toward your existing policy. The standard correction is to approve a small random sample above your threshold. Some of those orders will be fraud, and that loss is a real, recurring, budgeted cost of keeping the model honest. Nobody puts it in a proposal. Put it in yours.

Beyond that, a maintenance retainer at fifteen to twenty percent of build cost per year, covering acquirer interface changes, card network evidence rule changes, and model serving infrastructure. Retraining has a cadence and a cost each time it runs.

And the balance sheet change that is not a cost but belongs in the model: once you stop paying a guarantee, you carry the fraud loss yourself. That is the point of the exercise, but it moves a predictable fee into a variable loss line and your finance team should agree to that before the project starts, not after the first bad month.

Comparing a build against your current renewal

This is the one category where the comparison is genuinely straightforward, because the guarantee fee is a percentage of approved volume and you can read it off a statement.

Project five years of guarantee fees at your expected growth rate. Against that put the build cost, five years of retainer and hosting, the analyst headcount you need either way, the fraud loss you will now absorb, and the bias correction sample loss described above. If the guarantee fee line is larger, you have an arithmetic case rather than an argument.

Then add the line that usually decides it and that nobody measures: false declines. A good customer refused at checkout does not complain, they buy elsewhere and quietly stop being your customer, so the cost appears as absent revenue rather than as a charge. You can make it visible without building anything. Report approval rate by category, by customer tenure and by fulfilment method alongside fraud loss for one quarter. If your highest margin categories are being declined at the same rate as your most resale friendly ones, you have found the money, because a single global threshold applied to a catalogue with wildly different economics is leaving margin on both sides of the decision.

When buying beats building

Buy if your annual online volume is under roughly $20M, if your chargeback rate is comfortably under 0.3 percent, or if you genuinely value the balance sheet certainty a guarantee provides more than the margin it costs. There is nothing wrong with paying somebody to carry a risk you do not want to carry, and saying so is not a weakness in the argument.

Signifyd, Riskified and Forter are all credible, and they hold one advantage you can never replicate: their models see patterns across many merchants, which is a genuine strength against organised card testing that hits several retailers at once. Do not pretend otherwise when building your case internally.

Keep a chargeback specialist such as Chargebacks911 for the long tail even if you automate. Their value is process and template knowledge on unusual dispute types, unfamiliar regions and arbitration cases where your own volume is too low to build expertise economically. Automating your high volume reason codes and outsourcing the rest is a sensible steady state rather than a compromise.

Build when two or more of these are true. Your guarantee fees now exceed what a small risk engineering team costs. Your decline rate is unknown or unmanaged and merchandising has started asking about it. Your margin varies widely across the catalogue and one global threshold is visibly wrong. Your losses are increasingly policy abuse rather than stolen cards, which a transaction scoring vendor structurally cannot see. Or you are approaching a card network monitoring programme threshold and need to move your rate deliberately rather than by asking a vendor to tighten. The tipping point is not technical. At scale the risk threshold is a pricing decision, and pricing decisions should not sit outside your business.

If you want a second opinion before signing anything, Digital Heroes starts every engagement with a signed specification covering the data model, permissions and acceptance criteria, which is what keeps a fixed price fixed. Nothing about that commits you to the build.