For online retailers we automate the repetitive operations behind every order: exception triage for failed, fraudulent or stuck orders, WISMO and returns answered by an AI agent, supplier and invoice intake, and storefront-to-ERP sync. On our e-commerce project order cancellations fell from 15% to 5% and around 70% of customer contacts closed without a person.
Address errors, stock-outs, and failed payments stall orders in manual queues, especially overnight and at peak.
“Where is my order?” and returns questions flood support and bury the tickets that actually need a human.
Product feeds, invoices, and supplier sheets arrive in every format and get rekeyed by hand.
Multi-step returns logic runs on memory and spreadsheets, with no audit trail when something goes wrong.
Detect failed, fraudulent, or stuck orders and route each to the right fix or owner automatically.
AI agents answer “where is my order”, process returns, and escalate only true exceptions.
Read supplier sheets and invoices, normalise them, and sync clean data to your catalog and ERP.
Keep inventory, pricing, and orders consistent across storefront, ERP, and 3PL — with retries and alerts.
of contacts closed by AI
Client under NDA. The figures are the client’s own, comparing the periods before and after launch.
This is the project behind the case above, described the way it happened rather than as a method. A retailer of short shelf-life goods selling through two marketplaces, its own storefront, an offline till and ad feeds. No single system was broken — the damage was all between them. A new item was keyed in five times, stock was reconciled by an export once a day, and in the meantime a channel would sell what the warehouse no longer had. One order in seven was cancelled after the fact.
An item, its price and its stock are created once and go out to every channel. Reconciliation moved from once a day to near real time, which is where most of the cancellations lived.
Labels, manifests and the courier handover are generated from the order instead of being assembled by hand in the evening.
Receipt is scanned against the purchase order, discrepancies are raised at the dock, and shelf life is recorded at the moment the goods arrive.
Where-is-my-order, cancellations, returns and the routine part of pre-sale questions are answered automatically. Around 70% of contacts close without a person; the rest are handed over with the history attached.
Demand is projected per item and per channel, and the purchase order is drafted from it. A person still approves it.
Two months in total, and each loop went live on its own and started paying before the next one began — the catalogue first, because that is where the cancellations were. Cancellations went from 15% to 5%. What deliberately stayed with people: counting goods physically at receipt, the final approval of a purchase order, price conflicts between channels, and the roughly 30% of contacts the AI hands over.
One to two weeks. Fixed, and credited in full against the build.
A fixed quote against the scope the audit produced.
On this project. Two to four months is the range across our cases.
Model and infrastructure usage plus maintenance. Moves with volume.
In the EU and the UK a customer has a statutory withdrawal period, and an automated reply that contradicts it is a liability. Refund decisions stay with a person; the agent prepares the case.
Card and payment details are excluded at ingestion. What a flow needs about a payment comes from the payment provider by reference, not by copying the data.
Queues, rate limits and fallbacks are sized for Black Friday rather than for an average Tuesday, and we load-test against last year traffic.
Amazon, eBay and the other marketplaces cap what an automated seller message may say and how often. Flows respect those limits instead of discovering them through a suspension.
The single largest share of retail support volume, answerable entirely from data you already hold. Usually the fastest payback in the sector.
Collecting the reason, the photos and the order reference, checking eligibility against policy, and handing a complete case to a person.
Normalising incoming catalogues, prices and stock into the shape the storefront expects, without anyone retyping them.
Model-driven dynamic pricing is not something we build: the legal and reputational downside dominates the upside.
Below roughly 500 orders a month, most of these flows will not repay a build within the year.
If your product data is not consistent, nothing built on top of it will be. That is a data project first.
Each page states the workflow, the systems it integrates with, what it costs and when it is the wrong choice.
Classifies and routes inbound tickets and drafts the reply, with a confidence threshold that sends the rest to a person.
Reads incoming invoices, checks totals, tax and supplier against the ERP, and posts only what clears the rules.
Sorts mixed incoming documents by type and routes each one to the queue, folder or system that owns it.
No. We sit alongside your storefront and read order events — we don’t modify checkout or hold card data. Payment exceptions are flagged for your team, not actioned blindly.
We don’t propose an off-the-shelf product either. The audit is there precisely to see where your process differs from the ordinary one, and the build is shaped around that difference. And if it turns out a standard tool covers you well enough, we’ll say so plainly — the process map stays with you in any case.
It’s a fair thing to ask, and responsibility matters more here than accuracy. A person answers for it, and the system is built so that they can: a doubtful case is never posted quietly, it goes to a review queue. Low confidence isn’t an error for us, it’s a route. Around 70% clears straight through and a person looks at the rest — and where exactly that line sits is yours to decide.
That’s a common requirement, and a reasonable one. We fit the setup to your data rules: if nothing may leave, the model runs locally inside your own perimeter, and the documents never cross it.
That’s not unusual, and it’s fine. Integrating without an API is the most underestimated line in a quote, which is why integration surface comes first among the four cost drivers, and why we don’t name a fixed price before the audit. Working without an API is perfectly possible — files, exports, email — it simply costs more, and it’s better to know that early. Which is why, fairly often, we build the API for you.
It happens often, and usually for one reason: the pilot was measured on the happy path and the exceptions were left for later — though the exceptions are where most of the work turns out to be. So we look at the share that clears without a person rather than at extraction accuracy, and we agree in advance who handles the rest, and how.
No, that isn’t what this is about. What goes is the retyping, not the people: decisions stay with a person — the disputed document, the non-standard transaction, the conversation with the client, the signature under the reporting. On one accounting project, closing a client month went from three days to four hours, not because anyone was let go, but because a qualified specialist stopped keying in details by hand. The time that frees up most often goes into growth: more clients with the same team, without the costs rising alongside.
It’s a fair question, and sometimes doing it yourselves really is the right answer. We say so when the volume doesn’t justify a build: below roughly 300 documents a month the arithmetic usually doesn’t work. The calculator on this site includes running costs, so you can weigh that up before you ever talk to us.
They do change, and that’s exactly what we build for: the model is a replaceable part here, not the foundation. The source of truth is our own database — the document, the counterparty, the entry. The model is attached to the side, and we update it whenever something better appears; for you that’s planned maintenance, not a rebuild.
We use cookies for analytics — to see which pages bring enquiries. Nothing else, and nothing before you agree. Cookie Policy