Chargeback agent for service businesses
Dispute Desk reads each chargeback, pulls evidence from the systems a business already uses, decides whether it's worth fighting, and drafts the response. The owner approves, or lets the agent submit the strong ones on its own.
Why it matters
Merchants who contest a chargeback win 44.6% of the time, yet recover only 10.7% of chargeback dollars overall. Most disputes never get answered.
Chargebacks911 2026 field report, via Redo
The average US chargeback, plus what retailers say each one costs them in internal work and outside fees.
Javelin research for Mastercard, via Chargeback Gurus
How long Square gives a seller to respond, counted from the notice email. Proof that isn't uploaded in that window never reaches the bank.
Chargebacks projected worldwide in 2026, growing 37% from 2025 to 2029.
The market today
Builds an evidence packet and submits it just before the deadline. Its fee applies only on wins.
Limit: it works from data Stripe holds. It can't see a booking app, a signed form or a text thread. Stripe docs
Reads the reason code, pulls records, writes the cover letter and submits on time for salons and spas.
Limit: it lives inside Zenoti's own platform. A business has to run on Zenoti to get it. Zenoti
Dispute automation for online stores, typically priced as a share of what's recovered.
Limit: the integrations Chargeflow lists are payment processors, e-commerce and subscription platforms. No booking or field-service software is on the list. Chargeflow help
The business gathers the proof and uploads it by hand. Housecall Pro's own guide suggests pasting screenshots into a Canva template and exporting a PDF.
Sources: Housecall Pro, ServiceTitan, Boulevard
Card payments in one place, bookings in another, signed forms in a third, texts on a phone. Nobody's tool reaches across all of it, so the owner does it by hand or doesn't do it. Dispute Desk connects what the business already has. That's the Strategia model applied to one painful job.
How it works
Translate the reason code into plain English and start the deadline clock.
Pull the evidence that answers this exact claim from every connected system. Flag what's missing.
Weigh evidence strength against the amount and the processor's fee. Accept the ones we owe.
Write the rebuttal and assemble the packet. The owner approves, or the agent submits strong cases itself.
File before the deadline, track wins and dollars, and tune the weights from every outcome.
What it's worth
Every number here is an assumption you can change. The defaults come from the published figures above, not from a promise.
How it's tested
Sample disputes from two industries, covering every common reason. The right call for each was written down first. The engine has to match all of them.
See resultsCreate test disputes in Stripe's test mode, build and submit the evidence through Stripe's API, and read back the outcome. Real pipeline, no real money.
Run five to ten past chargebacks, personal details removed, and compare what Dispute Desk would have sent with what actually happened.
Customer workspace
The Strategia overview describes a workspace where the customer sees an AI workforce, not a technical dashboard. This is that view for one agent, built from the live demo's current state.
Agent Factory blueprint
The eleven things an implementation agent has to know before anything gets built, filled in for chargeback response.
A dispute opens at the processor. On Stripe that's the charge.dispute.created event. On Square it's dispute.created.
Card processor. Booking or field-service app. Forms and e-signature. Front desk tablet or dispatch GPS. Text messaging. Photo storage.
The processor for the amount, reason code and deadline. The booking or job record for who, what and when. If two systems disagree on a time, the one that recorded the event wins.
Reason code, amount, card network and deadline. The customer and the service. Every evidence item that answers that specific reason.
Fight or accept, from evidence strength, the amount and the processor's fee. Whether Visa Compelling Evidence 3.0 applies. Whether to text the customer first.
Pull evidence. Assemble the packet. Draft the rebuttal. Submit through the processor. Notify the owner. Draft customer messages. Log the outcome.
Read access to bookings, forms, check-in or GPS records, texts and photos. Write access to disputes at the processor only. Never full card numbers.
Key evidence missing. A system that was offline during the visit. A refund already promised. A true duplicate. An amount under the minimum.
Level 3: the owner approves every response. Level 4: the agent submits at 85% strength with all key evidence, and the owner gets accepts, middle-range cases, missing evidence and contact-first cases.
An evidence packet and rebuttal filed before the deadline. An owner notification. An activity log. A monthly report of win rate and dollars recovered.
The answer-key replay passes in full. The sandbox run submits evidence end to end. On real history, each call is reviewed with the owner. No missed deadlines in the pilot.
Provider neutral
The agent is defined by what it does, not by one vendor's recipe. Each processor is a thin adapter. Switch businesses in the live demo and the same engine answers through a different one.
Agent Factory
The brow studio came first. The plumbing company was the second implementation. This is what carried over untouched and what had to be mapped.
Where this sits
A defined spec, a working engine and a test suite. This prototype.
Connected to one real business. A person checks every packet.
New businesses on a known stack deploy from the template, with review.
Deployed from the spec. People handle exceptions only.
Test 1
The expected call for each sample dispute was written by hand, from the facts of the case, before the engine ran. This runs in your browser each time the page loads, always with the default rules.
Test 2
Test 3
Waiting on five to ten past chargebacks with names and card details removed. Each one runs through the same engine, and the result is compared with what actually happened: won, lost, or never answered. This is the test that shows whether it would have recovered real money.