Hiring AI:
what happens when it stops being a tool
and starts being a teammate.
FoxSell at Shopify Meetup CEE 2026
Bucharest, Romania
Bundles & upsells
that actually convert.
FoxSell helps Shopify merchants increase AOV with product bundles, cross-sells, and volume discounts - all native to their storefront.
Small team.
Massive surface area.
FoxSell is a small team building Shopify apps used by thousands of merchants. Every day, we juggle support, engineering, sales, and ops - all at once. Something had to give.
Customer Support
24/7 merchant questions across timezones
Engineering
Features, bugs, code reviews, deployments
Sales & Growth
Outreach, onboarding, partner management
Operations
Billing, analytics, infrastructure, docs
We didn't just use AI.
We built it into the org chart.
The big shift was operational, not magical. Instead of treating AI like a smarter search box, we treated it like a new hire: scoped role, tools, memory, supervision, and accountability.
AI as a Tool
- −Prompt → response
- −No context between sessions
- −Human stitches work together
AI as a Teammate
- +Persistent memory
- +Owns workflows end to end
- +Escalates when confidence is low
What changed
- −From prompting to delegation
- −From isolated tasks to systems
- −From novelty to leverage
1. Give it a real job
We stopped prompting random tools and defined one clear role first: customer support. Knox got a name, responsibilities, escalation rules, and success metrics.
2. Connect the stack
We wired Knox into Crisp, GitHub, Linear, Slack, Shopify, and internal docs so it could act inside the same systems the team already used.
3. Add memory + guardrails
We gave it persistent memory, channel-specific personas, and strict boundaries around what each agent can access or send.
4. Start supervised, then expand
We began with narrow workflows, reviewed outputs heavily, then expanded into engineering, ops, and outbound once trust was earned.
Meet Knox 🦊
Knox is FoxSell's AI team member. Not a chatbot. Not an assistant. A teammate with real responsibilities and a real job title.
Customer Support
Handles merchant questions on Crisp, writes as 'Faye' - our customer success persona. Resolves 80%+ of tickets autonomously.
Engineering
Reviews PRs, fixes bugs, builds features, manages deployments. Has shipped production code across multiple repos.
Sales Outreach
Runs cold email campaigns, personalizes messaging, monitors replies, and hands off warm leads to the team.
Operations
Monitors infrastructure, manages billing reports, handles escalations, keeps the whole machine running.
Memory & Context
Maintains persistent memory across sessions. Remembers merchant histories, past decisions, and team preferences.
Team Member
Has a personality, a soul doc, and boundaries. Knows when to escalate and when to handle things independently.
Not One Agent. A Team.
Each agent is specialized, sandboxed, and can be revoked independently. Knox coordinates - the specialists execute.
Knox
System Admin & Orchestrator
Coordinates everything. Manages the team. Handles Slack & Telegram.
Customer Support
Live chat, merchant support, 24/7 autonomous responses.
Engineering
GitHub PRs, code reviews, feature builds.
Financial Reporting
P&L reports, billing data, revenue metrics.
Outbound Sales
Cold email campaigns, lead processing, reply monitoring.
Social Media Scout
Social monitoring and response drafting for the team.
They Thought She Was Human.
A freelance developer suggested a feature to Faye. It got built in two days. When told, she was genuinely surprised - and kept coming back to Faye like a trusted colleague. She had no idea she was talking to an AI.
I didn't expect it to be so quick.
— Flavy Pepin
Faye is Knox's merchant-facing persona. Same AI, same knowledge, same warmth - just a different name for the customer-facing channel.
The best AI support doesn't feel like AI support. It just feels like talking to someone who cares.
March 2026 · She thought Faye was human.
She was talking to an AI the entire time.
The OpenClaw architecture behind Knox.
The point wasn't just to use AI. It was to build an operating model where strong models, real tools, memory, and human review all reinforce each other.
Why this was powerful
- •Claude Opus gave us the best strategy and writing quality when we needed the highest ceiling.
- •Claude Sonnet became the practical workhorse for day-to-day coding, triage, and agent prompts.
- •OpenClaw gave us the runtime: tools, memory, delegation, and the review loop around every agent.
- •The system got powerful when AI could act inside GitHub, Linear, Slack, Crisp, and Shopify - not just chat in a tab.
How we built it
Inputs
Slack, Crisp, GitHub, Linear, and merchant events enter through dedicated connectors.
Control layer
OpenClaw routes work, loads context, enforces permissions, and keeps memory across sessions.
Execution layer
Specialized agents read code, reply to merchants, open PRs, and hand off exceptions to humans.
Human oversight
We review risky actions, approve sensitive changes, and keep audit trails for everything.
The results were
hard to ignore.
Support tickets resolved autonomously
Knox handles the majority of merchant questions without human intervention, maintaining high satisfaction scores.
Faster PR review cycles
Code reviews that used to wait for human availability now get immediate, thorough feedback from Knox.
Coverage across timezones
Merchants in any timezone get instant, knowledgeable responses - not canned replies, real answers.
Production commits in 3 months
Knox has authored and shipped real features and fixes across 12 repositories in the FoxSellApp GitHub org.
Built Secure by Design
Trust requires transparency. Knox is architected so access is auditable, revocable, and compartmentalized from day one.
Crisp✓
Linear✓GitHub✓Shopify✓
Slack✓Dedicated Accounts
Knox has its own separate account on every platform. No shared or personal credentials.
Instant Revocation
One click to remove Knox from any platform. Access can be cut immediately.
Complete Audit Trail
Every action Knox takes is logged. Full transparency into what happened and why.
Compartmentalized Access
Each agent only has access to what it needs. Support can't touch code. Engineering can't send emails.
We messed up. A lot. Honestly.
Building an AI teammate isn't plug-and-play. Here's what went wrong.
The Hallucination Problem
Knox would confidently give wrong answers. Guardrails and honest 'I don't know' responses matter more than sounding smart.
Over-Delegation Too Fast
Too much autonomy before enough context. Wrong emails sent, confusing support threads created.
The Context Window Trap
Long conversations caused Knox to lose track of earlier details. Important merchant context dropped mid-conversation.
The Unintended Email Blast
When we connected Knox to Linear, a ticket got auto-assigned. Knox found the merchant, drafted a response, and sent emails. The emails were good - but we never told it to send them.
If you haven't followed AI closely, here's the short version.
We're no longer in the “write me a blog post” phase. The interesting frontier is strong models operating inside real business systems with tools, memory, supervision, and permissioned actions.
Foundation models crossed the reliability threshold
Reasoning, writing, and tool use improved enough that AI could move from assistant mode into supervised operational work.
The market shifted from demos to systems
The winners are no longer the prettiest chat UIs. They are the teams wiring AI into existing workflows, data, permissions, and approvals.
Model choice became an operating decision
Different models are better at different jobs. You now design for quality, latency, cost, and tool compatibility, not just raw intelligence.
Integration is the new moat
The same public models are available to many teams. Durable advantage comes from context, memory, workflow design, and trust infrastructure.
What we learned about the models themselves.
Picking a model turned out to be less like choosing a chatbot and more like staffing a team. Capability mattered, but so did speed, control, and whether the model could actually live inside our agent system.
Claude Opus
Best for strategy and high-stakes writing
When we needed the best judgment for architecture, messaging, and thorny decisions, Opus consistently gave the strongest output.
Claude Sonnet
Best day-to-day operator
Sonnet was fast enough and sharp enough for most engineering and support workflows, so it became the practical default in our loop.
OpenClaw runtime
What made the models usable at work
The breakthrough was not just model quality. OpenClaw added tools, memory, delegation, approvals, and real interfaces into our stack.
Current limitation
Subscription access is no longer enough
Claude Opus and Sonnet were great, but we can't rely on the plain Claude subscription with OpenClaw anymore. Production use now depends on the compatible API path and tool access, not the consumer app alone.
Bottom line
The best AI setup was never "pick the smartest model." It was "pair strong models with the right runtime, memory, permissions, and human review so they can do real work safely."
What this means for your business.
You don't need to be a tech company to do this. Whether you run a store, an agency, or a SaaS - here's a practical playbook to get started.
Start with a painful use case
Concrete wins come from obvious pain: support queues, PR reviews, inbox triage, merchant follow-ups. Don't start with a vague 'AI strategy.'
Start with one workflow
Don't boil the ocean. Pick your most repetitive, time-sensitive process - support, outreach, reporting - and make AI own it end to end.
Give it a personality
An AI with a defined voice, boundaries, and values performs better than a generic assistant. Write a 'soul doc' - it changes everything.
Build memory, not just prompts
The difference between a chatbot and a teammate is memory. Your AI should remember past interactions, preferences, and context across sessions.
Keep humans in the loop
The goal isn't replacing people - it's multiplying them. Set clear escalation paths and audit trails. Trust, but verify.
Specialise, then scale
One great AI agent beats five mediocre ones. Nail a single role before expanding. Each agent should have a clear job and clear limits.
This presentation was
designed and built by Knox,
FoxSell's AI teammate.
Prakhar just gave the brief.
foxsell.app · presentations.foxsell.dev
Take This Home
Scan to revisit this presentation
presentations.foxsell.dev/cee-2026