ImageRestoreAI Support AutomationI automated support drafts
with human review.
A real customer support workload became a deployed workflow: classify the request, gather the facts, prepare a reply and let the support lead review it.
Support needed to take less time without losing customer trust.
ImageRestoreAI is an AI photo restoration service. With two founders and roughly ten support tickets a day, the same inbox contained routine access questions, valuable custom requests and sensitive billing issues.
I owned customer support. Treating every ticket the same way meant spending time on repetitive answers while still needing careful judgment on refunds, credits and high-value requests.
My responsibility
I translated the support workload into a taxonomy, selected the first automation scope, evaluated tooling and built the workflow with AI development assistance. I also remained the human reviewer for customer replies.
I analyzed 515 email threads to decide what to automate.
I combined my operating experience with an AI-assisted analysis of 515 support email threads, then reviewed the resulting categories. The analysis grouped requests into routine work, revenue opportunities and risk-sensitive cases.
Automate preparation
- Classify the ticket and identify the customer
- Retrieve order, account and policy context
- Prepare routine reply drafts
Retain human judgment
- Review replies before sending
- Handle billing disputes and sensitive cases
- Shape high-value sales conversations
The working taxonomy estimated 60% routine requests, 25% revenue opportunities and 15% risk-sensitive cases. These categories guided the scope; they were not a claim that all routine tickets could be safely resolved without review.
Each ticket becomes a draft grounded in customer context.
A support email starts the workflow.
Chatwoot sends an event to the backend. A queue separates receipt of the message from the longer AI processing work.
The system gathers relevant facts.
Customer and order context comes from Supabase, Stripe and read-only Gmail history, alongside support policy.
Three AI roles prepare the answer.
A classifier identifies the request, a drafter writes a proposed reply, and a verifier checks it against the available facts and policy.
The support lead reviews the result.
The workflow returns a private note and labels to the support interface. Cases that need escalation remain visible to a human reviewer.
Automatic refunds, credit changes and fully autonomous replies were follow-up possibilities in the case study, rather than outcomes established by this deployment.
I combined an existing support desk with a custom AI layer.
Buy the inbox; build the workflow.
I compared support platforms and automation options. Chatwoot supplied the support interface while the custom layer handled our customer context and policies.
Reuse the operating stack.
Supabase, Stripe and the existing email history already contained the facts needed to answer many requests. Reusing them limited new infrastructure and context duplication.
Keep review before sending.
A draft can save preparation time while leaving the support lead accountable for tone, correctness and customer impact. Risk-sensitive requests have an escalation path.
The case study reports a 50% reduction in my support time.
The May 2026 project retrospective reports one week from discovery to production, with support falling from roughly 30% to 15% of my working week.
These are reported results from my project presentation, not an independently audited or continuously refreshed performance dashboard. The measurement window and underlying logs are not included here.
The next step is to evaluate quality as well as speed.
Further work includes better product knowledge retrieval, tracking edits to drafts and evaluating which narrowly defined actions can safely receive more autonomy.
Let’s discuss your team’s deployment needs.
I’m seeking an AI Deployment Strategist role in San Francisco, where I can combine strategy consulting, customer discovery and practical AI implementation.