The clearest early result was that the repricing task disappeared within the first week for most sellers using the repricing agent. That confirmed part of our original view, but the cohort also showed us that independent sellers need more than task removal. They need useful controls, better category context, and a clear record of what happened while they were away.
This is our debrief from the first early-access cohort. We put independent Shopify and Etsy sellers through several weeks with early versions of the Seltova agents. The product was rough and not the version we would offer new signups today. We wanted candid input from people running real shops, and we received it. Some findings supported our thesis. Others sent us back to the whiteboard.
Better than expected
The repricing agent delivered results almost at once. Most sellers had been changing prices manually, with some doing it every morning before handling anything else. After connecting the agent, most said that task had essentially left their daily routine within the first week. They still reviewed pricing summaries, but the hands-on work was gone.
The Reply Agent also performed well with a defined group of buyer questions. It reliably handled routine inquiries such as "when will my order ship," "do you offer custom sizes," and "what is your return policy" without seller involvement. The biggest gains came for sellers with clear written shop policies. The agent could answer well only when the store documentation contained the answer. When it did, most sellers no longer treated the inbox as a first-thing-in-the-morning task.
Overnight processing produced the shift we hoped to see. Most sellers reported fewer evening logins and said they were checking in once in the morning rather than multiple times a day. The exact time saved varied by seller, but the change in their work pattern was consistent.
What fell short
The Listing Agent created the most friction. Our first version attempted complete listing drafts from very little input, usually a photo and short description. That was adequate for simple products. Handmade, one-of-a-kind, and highly specific items were different: the drafts often left out the vocabulary that buyers in those categories care about.
Sellers making hand-thrown ceramics told us that buyers look for details such as kiln type, glaze chemistry, cone temperature, and whether an item is food-safe. Our first Listing Agent did not ask for those details or reliably add them to the copy. The descriptions were accurate, but beside an experienced seller's work, they sounded generic.
We changed the Listing Agent's intake to request more detail before drafting and added category detection to guide its vocabulary. The work took longer than planned. Sellers testing the revised version said the drafts were substantially more useful as starting points.
Where sellers needed more control
Sellers repeatedly asked for visibility into the agents' work. After an overnight cycle, they wanted to see which listings changed price, by how much, and why. They also wanted to know which buyer messages were answered, what was sent, and which messages were sent for human review. Early in the cohort, our activity log was not good enough.
This mattered more than we anticipated. Sellers generally trusted the outputs after the first week, but they still wanted to inspect the work, especially at the start. The log served a practical purpose too. If a repriced listing performed poorly, the seller needed to know what changed and when so the issue could be investigated.
We shipped a basic activity summary within the first month of the cohort. Sellers consistently called it one of the most useful additions. It did not alter the agent's actions, but it kept sellers informed without requiring constant checks.
Which sellers gained the most
A clear pattern emerged. Sellers with some baseline ops hygiene before joining gained value sooner. Their shop policies were written, their product costs were tracked at least roughly, and their listings already had fairly complete tags and attributes.
Sellers starting from a more chaotic position found that the agents revealed gaps before they could fill them. Without a written return policy, the Reply Agent cannot answer questions about returns. Without a known cost per unit, the repricing floor cannot be set properly.
We are not saying a store must be perfectly organized before Seltova can help. Still, sellers who had the basics in order moved further, faster. We now put an ops-readiness checklist into onboarding because of this pattern.
What we are still building
Multi-channel inventory sync remains under development. Several cohort sellers operated on both Shopify and Etsy, making aligned inventory counts a recurring pain point. We tested a rough cross-channel reconciliation flow, but it was not reliable enough for broad release. We are treating inventory sync as a near-term priority.
Edge-case buyer messages are another open area. Routine inquiries work well, but complaints, disputes, and requests outside standard policy categories still require a human decision. The agent flags them for review, which is appropriate, yet sellers wanted help deciding how to respond, not only a flag. We are building a suggested-response layer for these escalations.
What the cohort taught us
Running a cohort before opening access more broadly was the right decision, even though it slowed us down. Without direct feedback, we might have released a Listing Agent that served handmade sellers poorly. We also might have launched without an activity log because we underestimated how much sellers wanted to see the work completed overnight.
Our thesis still holds: independent sellers carry real ops work that an agent-based system can handle, and removing some of it changes how they use their days. Making each agent genuinely useful for its category required far more iteration than we expected. That is the honest picture of where we stand.
For anyone on the waitlist or considering early access, we can say that the product is meaningfully better than the version tested by that first cohort. It is not finished, and each cohort will shape further improvements. We are building incrementally, learning from real operators instead of polishing something that misses the actual problem.