Bee, our inventory AI, picked the wrong house and wrote a 47-item inventory for a job that needed 12 items. The customer almost got billed for it.
Here’s what happened and what we did next.
What Bee Does
Bee is MoveRight’s AI inventory assistant. It can auto-generate an inventory from three sources:
- Photos uploaded by the customer during a virtual survey
- Virtual survey video analyzed frame by frame
- Text descriptions parsed from customer notes
The intent is clear: save estimators time on data entry, reduce the gap between what the customer says they have and what actually gets quoted, and produce a more accurate starting point than a blank form.
When Bee works, it works well. A typical 3-bedroom house with standard furniture, captured in a 10-minute virtual survey video, produces a 95%-accurate inventory in under 90 seconds. The estimator reviews, confirms, adjusts, and moves on. What used to take 20 minutes of manual entry now takes 3.
What Went Wrong
Here’s the anonymized incident:
A customer submitted a virtual survey video for a 2-bedroom apartment. Bee analyzed the video and produced a 47-item inventory that included items from the wrong unit. The AI had identified the exterior of an apartment building and, unable to distinguish Unit 4B from Unit 4A, pulled inventory items from a different floor plan that matched the building’s general description.
The estimator caught it during review. The job would have been quoted for 47 items — roughly 3x the actual scope. The customer would have received a quote for a 4-bedroom move on a 2-bedroom apartment. If the estimator hadn’t caught it, the crew would have arrived expecting 47 items and 4 hours of work for a 2-hour job. The downstream effects would have been messy: a revised invoice, a confused customer, and a dispatcher rescheduling the rest of the day.
The Right Reaction Is Not “Kill the AI”
It’s “gate the AI.”
Bee’s core capability — converting visual input into structured inventory — is sound. The failure mode was specific: unsupervised write access. Bee was allowed to write inventory directly to the job record. No human in the loop. No confirmation step. No guardrail against the obvious failure pattern of “the AI is confident but wrong.”
That’s a deployment mistake, not a capability mistake. The fix isn’t to remove the AI. The fix is to change the permission model so the AI can suggest but not write without human confirmation.
The New Setting We Shipped
Effective immediately, every MoveRight account has a per-account toggle for Bee’s inventory behavior. Three modes:
Suggest (default for new accounts). Bee proposes inventory items. The estimator sees them as suggestions — highlighted, deletable, editable — but nothing is written to the job record until the estimator explicitly confirms. If Bee suggests 47 items, the estimator sees all 47, deletes the wrong ones, confirms the right ones, and the inventory is clean.
Assist. Bee writes inventory only when an agent explicitly requests it via the prompt. The agent types “Bee, generate inventory from this video” or clicks the Bee assist button. Bee produces a draft. The agent reviews. The draft replaces the existing inventory only on explicit confirmation. This is for estimators who trust the AI but want to control when it runs.
Auto (opt-in only). Bee writes inventory directly from virtual surveys without human intervention. This mode requires explicit opt-in from an account admin and is recommended only for high-volume virtual-survey operations where the cost of a wrong inventory line is low (the estimator will review before quoting anyway) and the cost of delays is high.
No existing account was switched. Current behavior is preserved. New accounts default to Suggest. The setting is in Settings → AI → Bee Permissions and can be configured per zone for multi-location operators.
Why We’re Publishing This
Most software companies hide bugs. They ship the fix silently, update the changelog with “performance improvements,” and hope nobody notices.
We think the moving industry deserves a vendor that talks about mistakes openly. Not because mistakes are good — they’re not. But because hiding them is worse. If Bee can make a mistake this significant, you deserve to know about it, understand what caused it, and see exactly how we fixed it.
AI in moving software is early. The industry is going to see more AI features from more vendors. The vendors who talk about what went wrong and how they fixed it will be more trustworthy than the ones who only announce features. We want to be in the first camp.
What This Means for AI in Moving
A short defense of “boring AI” — predictable, gated, auditable.
Predictable: the AI does the same thing given the same input. No surprises. No hallucinations in front of customers. No inventories for the wrong house.
Gated: the AI can suggest, but it can’t act without a human in the loop. This isn’t a limitation. This is how you deploy AI in production safely.
Auditable: every Bee action is now logged with the mode it ran in (Suggest, Assist, or Auto), the agent who triggered it, and the artifact it produced. If something goes wrong, you can trace exactly what happened, when, and why.
That’s boring. It’s also responsible. And it’s what we’ll keep doing.
Toggle Bee’s permissions in your account settings. If you’re not on MoveRight yet, this is what shipping AI responsibly looks like.
References:
- Voiceflow — AI agents for movers — https://www.voiceflow.com/industries/movers
- Avoca AI — https://www.avoca.ai/
- Supermove — AI Voice Agents for Moving Companies — https://www.supermove.com/blog/ai-voice-agents-for-dummies-moving-company-edition