- 11% → 64%
- Same-day customer resolution
- -71%
- Customer service email inbox reduction
- 14 days
- Time-to-deploy
- -83%
- Cost per customer interaction
Anonymized representative case based on patterns we see in this segment. Not tied to a specific client.
Background
The operator runs 12 stores in three countries, selling a specialty product category with a heavy mix of in-store, click-and-collect, and home-delivery orders. The stack is mature but old: the POS dates from 2011, the e-commerce platform is bespoke, and the ERP is a vertical-specific package with limited public API.
A two-person customer service team handled email and a phone line. Volume had doubled in three years; the team had not. Same-day resolution was below 15% and average response time was over 36 hours.
Challenge
The blockers were not the team's effort. They were structural.
- No clean API surface. The POS exposed almost nothing programmatically. Returns required a manager at the original store to look up the receipt manually. Stock checks required a phone call to each store.
- Channel fragmentation. Customer queries arrived by email, web form, Instagram DM, and phone. The team triaged across four inboxes.
- No system of record for conversations. Each query was a new ticket; context from previous interactions was effectively lost.
The CFO was clear on the budget: no POS replacement, no ERP migration, no new hires.
What Nordix BIOS did
Deployment used the computer-use fallback for the legacy POS — the agent operates the POS's web interface the way a human would, in a sandboxed environment, with explicit scopes on which screens and actions are permitted.
Days 1-5: BIOS was wired into the e-commerce platform's order API (clean), the ERP's stock endpoint (clean), and the legacy POS via a computer-use agent (with read-only scope for the first week, write scope added under policy after).
Days 6-10: Shadow week. The agent drafted replies on a single channel (WhatsApp) while the customer service team approved sends. Policy was tuned: returns over a defined threshold escalate to a human, stock-check answers cache for 15 minutes to avoid hammering the ERP, refunds always escalate.
Days 11-14: Go-live on WhatsApp. Email and Instagram DM were routed into the same agent in week three. The phone line stayed human-only by choice — the operator wanted a human voice for higher-value disputes.
Outcomes
After 90 days:
- Same-day resolution: 11% → 64%. Most queries (order status, stock checks, return eligibility) were resolved by the agent in under two minutes. The two-person team handled the residual queue.
- Customer service email inbox: -71%. Customers migrated to WhatsApp once they realized they got answers there in seconds.
- Time-to-deploy: 14 days. Three days longer than the typical Nordix deployment because of the computer-use integration on the legacy POS.
- Cost per customer interaction: -83%. Calculated against the fully loaded cost of the customer service team plus tooling.
A secondary win: the conversational history became a structured data source. The merchandising team mined the most-asked stock-check questions to identify which SKUs were under-stocked in which stores — a use of the data that the email inbox had never made possible.
Lessons
Three takeaways relevant to any retailer in a similar position:
- Old POS systems are not a blocker. Computer use as a fallback means an AI agent can take real actions against software that has no API. It is slower and more expensive per action than API calls, but for low-volume operations like returns it is more than adequate.
- Channel consolidation is a side effect, not the goal. The operator did not set out to unify email, Instagram, and WhatsApp. It happened because once the agent was good, customers chose the fastest channel.
- Keep the phone human, by choice. Not every channel should be automated. The operator kept the phone line human for high-value disputes and saw NPS rise on those calls, partly because the agents on the phone now had time to actually listen.