AI in DirtFleet
AI that earns its keep. No demo-ware.
Shipped today
What DirtFleet's AI actually does in production right now.
Three concrete features. All included in the standard plan, none as a paid add-on.
Photo-of-meter OCR
Driver snaps the dash. On-device parsing reads the digits when it can; if confidence is low, a server-side Claude call covers the gap. The driver always confirms or edits before save — we trust eyes over models.
VIN camera + NHTSA decode
On the add-asset screen, point the camera at the door jamb. Tesseract extracts the 17-character VIN; NHTSA's free vPIC API fills in year, make, model, and trim. Zero typing for a fleet of pickups.
Hour-based PM auto-flagging
Every meter reading runs against the asset's service intervals. When an asset crosses the threshold (say 250 hours since last oil change), the system creates a yellow flag in the mechanic queue — no human bookkeeping required.
On the roadmap
What we're collecting data toward — and the bar each feature has to clear.
We won't ship these until they actually work on real fleet data. Here's the honest state.
Predictive maintenance
Need months of complete meter + repair-log histories across each asset class to model failure curves. We'll ship this when we can predict a specific failure with a positive likelihood ratio over baseline — not before.
Anomaly detection on hours
Spotting an asset whose daily-hours pattern drifts (e.g. consistent 9-hour days suddenly going to 14) is high-signal for theft, unauthorized use, or a stuck meter. Lower data bar than predictive maintenance — likely shipping in the second half of beta.
Conversational fleet assistant
Ask "which assets are over 80% utilized this month?" and get a real answer over your data. Plumbing exists (admin metrics, exports); the work is making the LLM ground its responses in your tenant's rows without leaking across orgs.
Route optimization, dashcam AI
High-cost features that fit OTR trucking better than yellow-iron fleets. We'll add them only when our customer mix shifts to fleets where the math works — and likely via integration with existing vendors, not as our own product.
How we use AI
Boring rules that protect your data.
No model training on your data.
Claude and Gemini calls disable training-data retention. We don't fine-tune shared models on customer photos or logs — every tenant's rows are isolated at the database layer and that boundary holds for inference too.
On-device first.
VIN OCR runs in the browser. Server-side calls only happen when the on-device path returns low confidence — meaning fewer photos leave your phone, lower latency, and graceful offline degradation.
Data & AI
Production data stays product data.
Tenant fleet data is ordinary production app state—stored and protected like the rest of the platform (see Privacy). The Claude-calling flows shipped today (OCR fallback for meter photos, the “Highlight insights” narrative on the manager dashboard, and support-ticket triage) each send only the snippets that flow needs: one meter image for OCR fallback, fleet summary JSON for “Highlight insights,” and ticket + thread context for support triage — not bulk exports of your org for model training. The user-facing AI features highlighted above (Photo OCR, VIN decode, and hour-based PM auto-flagging) cover what drivers and mechanics actually see in the product. We do not sell customer payloads to third parties for training; if that ever changes, it would require an explicit policy update and in-product notice, not a silent shift.
FAQ
AI questions, answered honestly
Is DirtFleet an AI fleet management platform?
We're honest about what AI does and doesn't do here. Today, AI handles narrow but real jobs: reading meter values from photos when typing's awkward, decoding VINs from a phone camera, and auto-flagging assets whose hours cross service thresholds. We don't claim predictive maintenance or route optimization — those need months of accumulated fleet data from real operations.
What AI features are shipped in the product right now?
Three: (1) Server-side OCR via Anthropic's Claude reads meter values from a photo of the dashboard when on-device parsing struggles. (2) On-device VIN scanning via Tesseract.js extracts the 17-digit code from a camera frame, then NHTSA's free decoder fills in year/make/model/trim. (3) Hour-based preventative maintenance compares real meter readings against per-asset intervals (e.g. "every 250 hours") and auto-creates yellow flags for the mechanic queue.
Why no predictive maintenance yet?
Predictive models need months of clean meter data per asset class to stop guessing. Predicting an alternator failure on a CAT 320 6 weeks out requires hundreds of CAT 320s with complete hours histories, repair records, and failure outcomes — none of that exists day-one. We ship a predictive layer only when field data justifies the claim, not before.
What about route optimization?
Route optimization is genuinely valuable for last-mile delivery fleets, but most DirtFleet customers run yellow iron and mixed equipment that doesn't move on a fixed-route schedule. We'll add it when our customer mix shifts to fleets where it pays back, not as a vanity feature.
Do you use AI on driver behavior or in-cab cameras?
No. We don't ship dashcams, in-cab AI, or computer-vision distraction monitoring. Those products work for OTR trucking and against driver privacy in ways that don't fit the construction/ag/mixed-fleet customer profile. If a customer needs them, we'll integrate with the existing dashcam vendor's data feed — not build our own.
Will AI cost extra on top of the subscription?
Currently no. The OCR and auto-flagging are part of the standard subscription (from $99/mo, capped at $299/mo org-wide, fair use applies). If we add a heavy LLM-driven layer later (e.g. a conversational fleet assistant), it'll be a clearly-priced add-on so the base plan stays predictable.
Where does my fleet data go when AI features run?
On-device OCR (VIN scanning) never leaves the phone — Tesseract runs in the browser. Server-side OCR for meter values calls Anthropic's Claude with the photo only when local parsing returns low confidence; the photo is processed and discarded with no training-data retention requested. We never train shared models on customer data.
Does DirtFleet use AI for manager analytics?
Fleet managers get a dedicated analytics roll-up (hours, distance, repair spend, optional fuel estimates). Separately, you can click “Highlight insights” to run Claude on that same summary JSON — useful for spotting trends, not for replacing accounting or fuel-card systems.
Get started
30 days free — then simple flat pricing for the fleet.
Hands-on onboarding, help importing historic data when you want it, and a direct line to the team during trial. Cancel anytime; export and go if it's not a fit.