If AI is to succeed in reshaping work, it will reshape all kinds of work. I recently had the opportunity to interview Rafsan Bhuiyan, who. While his new endeavor, OrionQ appears focused on traditional marketing and lead generation, Rahsan would rather talk about AI in construction sites, garages, and service vans.
OrionQ builds AI agents that pick up the phone when a plumber’s under a sink, or schedule jobs while a roofer’s on a ladder. He sees AI as the difference between winning a contract or losing it.
Rafsan believes trades may turn out to be the most important proving ground for AI adoption.
Rafsan Bhuiyan Interview
What makes trades work—noisy job sites, gloves, ladders—a better proving ground for AI than an office, and where does that thesis break?
Because chaos is the real test. Office AI works on clean data; field AI works in motion over noise, dust, accents, and spotty connections. If an agent can handle that, it can handle anything. The thesis breaks when AI tries to replace judgment. On-site work still runs on human instinct; we just automate the grunt work so they can focus on safety and craftsmanship.
Your agents answer phones and schedule work mid-task; how do you guarantee latency, accuracy, and accent/multi-lingual handling in real field conditions?
We run edge-optimized models trained for low-bandwidth environments and multilingual speech: English, Spanish, Bengali, Hindi, and Urdu. Every voice agent caches essential workflows locally, so even when the connection lags, it finishes the call cleanly. Accuracy comes from task-specific small language models, not generic chatbots.
What are the hard KPIs that matter to a contractor—booking rate, first-call resolution, schedule adherence, truck rolls avoided—and how do you attribute ROI without hand-waving?
We only bill when verified outcomes occur: booked jobs, answered calls, or leads converted. Each QToken ties directly to an action, not a login. That makes ROI traceable to revenue, not activity. Contractors see it right in their dashboard: fewer missed calls, faster dispatches, and less idle time.
Outcome-based pricing sounds bold; how do you meter outcomes fairly across seasonality, lead quality, and cancellations so neither side gets gamed?
Every QToken has a checksum. We log timestamps, source quality, and resolution context. If a storm kills demand or a lead cancels, it’s excluded from billing. AI adjusts expected throughput based on seasonal and regional data, so no one pays for bad weather or bad data.
When connectivity drops—or tools are covered in dust—what’s the offline or “graceful degradation” path so the business doesn’t stall?
Agents store local task queues and use SMS fallback. If the signal dies mid-call, the agent auto-texts the customer, reschedules, and syncs once back online. Nothing gets lost. Think of it as military-grade “fail operational” design, not “fail safe.”
Which systems do you integrate with out of the box (dispatch, CRM, QuickBooks, ServiceTitan/Jobber), and what data ownership and privacy guarantees do customers actually get?
We connect to ServiceTitan, Jobber, HubSpot, Salesforce, QuickBooks, and custom CRMs through secure APIs. Data stays in the customer’s cloud; we don’t resell or train on it. OrionQ is SOC II in progress, built around explainability and traceable usage logs.
Field work is safety-critical. What guardrails prevent hallucinations, double-booking, or bad directions from creating real-world risk—and how are incidents audited?
Our AI never acts without context. It confirms all scheduling with cross-checks against GPS, time windows, and crew availability. Every AI action has an audit trail, timestamped reasoning, voice logs, and rollback capability. Hallucinations get quarantined like bugs in code, with human-in-the-loop review before deployment.
How did Army and Fortune 500 experience shape your reliability model: redundancy, failover, red-teaming, and post-mortems tailored to small shops?
Army training taught me redundancy saves lives. We run mirrored clusters and auto-failover at the agent level, not just the cloud layer. After any anomaly, we hold a “digital AAR,” an after-action review. It’s the same discipline I brought to CarMax’s personalization engine and T-Mobile’s GPT rollout; mission reliability over marketing hype.
What changes in roles follow adoption—dispatchers, CSRs, apprentices—and how do you train crews so AI augments rather than alienates skilled labor and unions?
We train crews like we trained execs at T-Mobile: start small, show utility. AI handles intake and paperwork so dispatchers can manage logistics, not ring phones. Apprentices learn faster because calls, notes, and instructions are logged and searchable. We emphasize augmentation; humans make the calls that matter.
Looking three years out, what uncertainties could flip your roadmap—edge vs. cloud inference costs, regulation (TCPA/recording consent), insurer requirements, or model updates that break domain nuance—and how are you hedging now? (I’m a scenario planner).
We’re testing hybrid inference, edge for calls, and cloud for analytics, to keep latency and cost in check. Compliance is baked in; call recording and consent flow adapt by state. For model drift, we pin core logic to versioned small-language models and retrain quarterly. Regulation will evolve; discipline won’t.
Rafsan Bhuiyan interview: AI in The Trades. Serious Insights Analysis
Rafsan argues for a different center of gravity for AI: not the spreadsheet or slide deck, but the job site. The signal is a design ethos built for chaos—edge-optimized, task-specific, multilingual agents that cache workflows, degrade gracefully, and keep working through dust, dead zones, and accents. Guardrails are not theater: GPS/time-window cross-checks, immutable audits, and rollbacks align AI with the practicalities of dispatch and safety. Outcome-based tokens move the conversation from activity to revenue, making attribution legible in a way dashboards rarely do. Data residency, explainability, and SOC discipline show a governance model owners and unions can live with.
Uncertainties remain—shifting inference costs, state-by-state consent rules, insurer demands, model drift—but the mitigations are concrete: hybrid edge/cloud paths, versioned small models pinned to domain logic, consent flows that adapt by jurisdiction, and an after-action culture that treats anomalies as teachable moments. The strategic bet is that reliability and auditability become features customers select for, not footnotes. If AI can earn trust under ladders and dust, office workflows become the easy part; the harder, more valuable work is proving that augmentation improves craft without eroding judgment.
For more serious insights on AI, click here.
For more serious insights on management, click here.
Did you enjoy the Rafsan Bhuiyan interview: AI in The Trades? If so, like, share or comment. Thank you!
Leave a Reply