AI agents for ecommerce
The queue is full of where-is-my-order. The sale is in the product question nobody answered.
VoxDonna builds custom AI agents for ecommerce teams across voice, WhatsApp and email. Each one is wired to the systems the work lands in — your store, helpdesk, CRM or ERP — and hands over to a person with the full context when it should.
Updated 31 October 2026 · Custom builds are scoped on a call.
An illustration of the flow, not a customer recording.
The short answer
It depends on what you are buying. If you need a self-serve chatbot app that sits on your store and answers common questions, an app from a store marketplace is cheaper and faster, and you do not need an agency. If you need an agent that works across phone, WhatsApp and email and connects to your helpdesk, CRM or ERP, you need a custom build, and that is the case an AI agency like VoxDonna is for.
VoxDonna is an AI agency run from Donna AI Labs Private Limited, registered in India, delivering remotely. We build custom AI agents on voice, WhatsApp and email, wired to the systems the work lands in, with human handoff that carries the conversation across.
We take one workflow end to end, such as order status or returns intake, and tell you plainly when an agent does not fit.
Why ecommerce brands need it
None of these problems is new. They break a small team in the same places, every season.
Where-is-my-order is a known ticket category in ecommerce support, and it repeats across phone, WhatsApp and email. Each contact is quick for an agent and still takes a person, a lookup and a reply. The answer sits in a store record that nobody on the team needs to read twice.
Orders are placed at night and problems appear at weekends. A queue that opens at nine means a night of silence for the customer who has a question, and a pile of messages waiting for your team in the morning.
A brand that ships across borders gets contacts in languages the desk is thinner in. Replies get slower and less sure of themselves in exactly the language where the customer most needs clarity.
A return or warranty claim needs a model, a date and a description. Collected over email, that takes several rounds. Collected in one conversation, it can go straight into the case with everything attached.
A buyer asking about sizing or compatibility is closer to a sale than most browsers. If the answer arrives the next day, they have often bought elsewhere. Nobody sees the sale that quietly did not happen.
When a chatbot gives up, the customer repeats everything to a person. The agents we build hand over with the conversation, the order details and what was already tried.
What the agents do
Each is a workflow we scope and connect to your systems. They are not shelf features, and we start with the one that hurts most.
A customer asks where their order is. The agent answers from the store record, not from a script.
The claim is opened while the customer is still talking, with the details your team needs.
Sizing, compatibility, care and availability, answered from approved catalogue material.
Some contacts are buyers still deciding. The agent tells them apart from customers with a problem.
Late, damaged or missing parcels are the contacts that need a calm first response and a clean handoff.
One agent, the customer's language, the same rules and the same escalation path.
The pattern in production for emailed orders, applied in a scoped build to the orders and requests that reach your inbox.
A conversation customers already read, used for the messages after the sale.
Channels
Voice, WhatsApp and email each suit different contacts. The rules, the data and the escalation path are the same across all three.
Voice suits the contacts that arrive by phone: order status, delivery problems, warranty claims and buyers who would rather talk. The agent picks up at any hour, answers in the caller’s language, collects what is missing and hands the exception to a person with the call attached.
Voice is the channel with the strongest published evidence on this site. Le Marquier, a premium consumer brand, runs its support line this way in French, English and Italian, with order and delivery status read from the ERP.
Honest scope: a voice agent answers from approved data and from the systems it is connected to. It does not improvise policy. Published managed plans start from €499 a month.
See AI voice agentsWhatsApp suits typed, asynchronous contacts: order questions, returns, product questions and the follow-up after a sale. The customer replies when it suits them, and the conversation stays in one thread. When a person takes over, they see the full history instead of asking the customer to start again.
VoxDonna runs agents on WhatsApp, Instagram DMs, Messenger and web chat. Self-serve plans start from $20 a month and cover AI replies. Appointment booking and follow-ups start from $99 a month, as listed on the WhatsApp agents page.
Honest scope: the self-serve plans are a starting point for replying to enquiries. Connecting WhatsApp to your store, helpdesk or ERP for order lookups is a custom build, scoped on a call.
See WhatsApp agentsEmail suits the contacts that arrive with detail attached: order requests, return forms, photos of a damaged item. The agent reads the message, checks it against your rules, takes the permitted action and replies in the same thread.
The email agent is in production reading emailed orders, validating them against ERP master data, raising the order and replying in-thread. Its published acceptance tests are written up on the order agent page.
Honest scope: the production use is order intake into SAP, Oracle and Dynamics. For a store’s support inbox, where-is-my-order and returns, it is the same pattern applied in a scoped build, not a shelf product.
See the email order agentProof
Le Marquier is a French outdoor-cooking brand, not an online store. It is the closest published deployment to ecommerce support: the same repeated questions, in several languages, connected to order data.
The agent answers in French, English and Italian on product care, warranty intake and dealer lookup, reads order and delivery status from the ERP, and qualifies leads into Zendesk. It handles 2,500 conversation minutes a month and took 8 weeks to deploy, with a soft launch at 20% of calls.
These are Voxdonna’s own published figures, not an independently audited result.
Comparison
A neutral view. The right row depends on your workflow, not on who is selling.
| Approach | Best for | Setup | Integrations | Languages | Who maintains it | Cost model |
|---|---|---|---|---|---|---|
| Self-serve chatbot app | Common store questions on a single website chat | Install and configure it yourself | Whatever the app offers for your store | Whatever the app supports | You, with the vendor’s product updates | Subscription, often by volume |
| DIY workflow tool | Teams with technical staff and time | You design and wire each step | As many as the tool connects to, built by you | As far as you build and test them | You | Tool subscription plus your own time |
| Custom AI agent build by an agency | Several channels, several systems, rules specific to you | Scoped, built and tested with you; Le Marquier took 8 weeks | Store, helpdesk, CRM, ERP, scoped case by case | Agreed and tested with your terminology | The agency builds and runs it | Scoped build cost plus running cost, agreed in writing |
If a self-serve app already fits your process, it is cheaper and we will say so.
Why VoxDonna
Every point below is something you can verify on this site.
We scope one repeated workflow end to end instead of dropping a generic widget on every page.
Zendesk and an ERP at Le Marquier, Salesforce and an RMA portal at Usha Martin, and a live Shopify store. Other systems are scoped case by case.
English, French, Italian, Hindi and other Indian languages, each tested with your terminology before launch.
Escalations carry the conversation, so your team starts where the agent stopped.
We can deploy in your own cloud region. Usha Martin runs in its own AWS India VPC.
The six cases an order agent must pass before it gets write access are written up in full, against a simulated backend, and labelled as such.
How we build agentsPricing approach
There is no single price for an ecommerce AI agent, because the price follows the channels, the systems and the languages. A WhatsApp agent answering enquiries is a different job from a voice agent reading order status from your ERP.
Where a plan is published, we show it. WhatsApp agents have self-serve plans from $20 a month. Voice agents have managed plans from €499 a month, with per-minute rates for extra usage. Custom builds across channels and systems are scoped on a call, and you receive the build cost, the running cost and the launch plan before work begins.
Before you start
It depends on the channel and on how much of your stack the agent has to touch. WhatsApp agents have self-serve plans from $20 a month. Voice agents have published managed plans from €499 a month, with per-minute rates for extra usage. A custom build across several channels and systems is scoped on a call, and you get the build cost and the running cost in writing before work starts.
Yes. VoxDonna has connected agents to a live Shopify store. Through scoped API access the agent can look up order status and product and stock details, and answer from that record instead of guessing. The exact permissions are agreed with whoever owns the store, and read access and write access are separate decisions.
Other store platforms are scoped case by case. We look at the interfaces, exports and permissions your platform exposes before agreeing what the agent can read or change, and we tell you early if a requirement does not fit.
It depends on the number of channels, systems and languages. The closest evidence we publish is Le Marquier, a premium consumer brand, where a French, English and Italian voice agent went live in 8 weeks, starting with a soft launch at 20% of calls. We do not quote a fixed number of days for custom work.
English, French, Italian, Hindi and other Indian languages. Each deployment is tested with your own product names and terminology before launch, rather than assumed from the list.
It hands over to a person with the conversation attached, so the customer does not repeat themselves. Complaints, anything outside the approved answers and anything the agent cannot verify go to your team. At Le Marquier, 2% of conversations were escalated and 98% were resolved by the agent.
Where data location matters, the agent can be deployed in your own cloud account and region. The Usha Martin deployment runs in the client's own AWS India VPC. Access to your store, helpdesk or ERP is scoped with the team that owns each system. We do not make certification claims on this page; ask for the details that matter to your security review.
It does not replace your support team. It will not invent an order status, a stock level or a policy that is not in your approved data, and it does not take actions in your systems that were not agreed in scoping. Abandoned-cart recovery and SMS are not part of what we describe here.
Yes, as a scoped build. VoxDonna's email agent is in production reading emailed orders, validating them against ERP master data, raising the order and replying in the same thread. A store's support inbox (where-is-my-order, returns) is the same pattern applied to a different workflow. It is not a shelf product, so it starts with a sample of your real emails.
Start where your customers already contact you. If most contact is typed and asynchronous, such as order questions and returns, WhatsApp is usually the faster and cheaper start. If customers phone, especially outside office hours or in several languages, voice carries more of the load. Many brands end up with both, sharing the same rules and the same escalation path.
Start where it repeats
Where-is-my-order, the return form, the sizing question. We will tell you whether an agent fits it — and when it does not.
Book an AI agent demoA short qualification form first. A discovery call if there’s a fit.
Prefer email? hello@voxdonna.com