Replacing the After-Hours Phone Room: 38 AI Agents Covering Nights, Weekends, and Holidays

The most boring job in trucking is also the one you can't ignore

The after-hours phone room is one of those structural costs every mid-size carrier carries because nobody has a better answer. Loads fall out at 11 PM. Drivers lock themselves out of trucks at 3 AM. A storm rolls into Wyoming on Saturday at 4 AM and fourteen drivers are exposed. Customer service tickets stack up over the weekend that need someone — anyone — to look at them before Monday.

Most carriers handle this with a person on the phone, paid $20–$28/hr, working nights and weekends. The fully-loaded cost runs $78K to $94K a year per coverage seat. They're handling about 12–18 events per shift, of which 80% follow exactly the same workflow: get the facts, take a small action, log the outcome, move on.

That's the kind of work AI agents are actually good at. Not the creative problem-solving — humans still do that. The repetitive, well-bounded workflows where the right action is determinable from the situation, and the cost of an action is small if reversible.

Centrix runs 38 distinct AI agent profiles, each with a curated set of tools, a defined risk tier per action, and a permission gate. They handle the after-hours load alongside the human on-call — usually with the human only seeing the events that escalated past Tier 0 or Tier 1.

How the agent system actually works

Three architectural pieces matter here, because the failure modes of "AI doing operational work" are real and most platforms hand-wave them:

Tier-based action gating

Every action an agent can take is classified into one of three tiers:

• Tier 0 — read-only or trivially reversible (look up a load, check HOS,

send a message). Agents take these without supervision.

• Tier 1 — operational, logged and traced (open a comm thread, dispatch a

weather call, flag a load). Agents take these with full audit trails.

• Tier 2 — destructive, financial, or driver-affecting (cancel a load,

modify a settlement, terminate a dispatch). These require explicit human approval — the agent prepares the action, presents it for review, and waits.

The classification lives in `policy/commands.yaml` and is enforced by the security kernel. CI blocks any new agent tool that doesn't have a tier assignment, so this can't drift.

Curated tool sets per agent

Agents don't have access to everything. Each agent profile (defined in `agent_profiles.py`) gets a specific list of tools matched to its job. The `after_hours` agent gets dispatch tools, safety tools, fleet tools — but not finance tools. The `fuel_optimization` agent gets fuel tools and read-only fleet tools — but not driver tools. The `claims` agent gets claims-only tools under a granular `safety:claims` permission.

This bounds the blast radius. An agent that doesn't have access to a tool literally can't call it, even if a prompt tries to manipulate it into doing so.

DLP scanning on every outbound

Anything an agent sends — Telegram message, SMS, email, voice transcript — runs through outbound DLP scanning with a 30-layer policy. PII is redacted before it leaves the system. Sensitive financial data is blocked. The result is that an agent can't accidentally text a driver's SSN or share a customer's bank information, even if the underlying data is in scope.

What the after-hours agents actually do

Load fall-out coordination

A load goes "Available" at 11 PM because the assigned driver hit HOS. The after-hours agent:

• Notices the status change in Alvys (within the 5-min sync cycle)
• Checks for HOS-compliant trucks within range
• Surfaces options to the dispatcher's after-hours channel via Telegram
• If no options inside the cutoff window, escalates to the human on-call

Human time consumed: 30 seconds, vs 12–20 minutes pre-Centrix.

Driver lockout / breakdown coordination

Driver calls in at 2 AM with a lockout. The agent:

• Locates the truck via Samsara
• Identifies the nearest 24-hour service provider in the carrier's preferred

vendor network

• Sends the vendor contact to the driver via Telegram in their language
• Logs the event in the comm hub for the morning brief
• Escalates to the human on-call if the truck is in a region without coverage,

or if the driver reports a safety concern

Weather safety auto-call

NOAA pushes a winter storm warning at 4 AM. The agent:

• Identifies all active drivers whose route intersects the warning area in

the next 12 hours

• Initiates Twilio TTS calls in the driver's preferred language
• Logs the call result (acknowledged, no-answer, escalated)
• For severe-weather + black-ice index above threshold, recommends a 4-hour

delay

• All of this happens before the on-call human's alarm goes off

Customer service triage

Customer emails and portal tickets accumulate over the weekend. The agent:

• Parses the ticket category (rate dispute, status request, document request,

appointment reschedule)

• Resolves Tier 0 items directly (status updates from Alvys, document pulls

from the load record)

• Escalates Tier 1+ items with a draft response for human review

By Monday morning, the customer-service queue has 70%+ already resolved or draft-ready, instead of 100% needing fresh attention.

The human role doesn't disappear — it changes

The after-hours human goes from 12–18 events per shift to 2–4. They're the exceptions: judgment calls, edge cases, escalations. They're also the high-value events — the kind that benefit from a human's experience.

Most carriers don't eliminate the after-hours role; they reduce its hours and have the same person handle exception escalations across two carriers or use the freed time for daytime work. Either way, the labor cost drops 60–80%.

What it costs

Anthropic API costs run $2,500–$5,000/month on a 100-truck fleet using the agents heavily — including the morning brief, the comm hub assistance, and the after-hours coverage. That's $30K–$60K a year of AI cost.

Compared to:

• $78K–$94K per after-hours coverage seat eliminated/reduced
• $20K–$30K of dispatcher time saved on routine load-board hunting
• $40K–$60K of safety/fleet manager time saved on morning-brief preparation

Net savings on a 100-truck fleet using the agents heavily: $80K–$130K/yr, plus the harder-to-quantify benefits of faster response times and consistent follow-through.

The Telegram channel

All agent activity routes through Telegram, which most carriers already use informally for driver communication. Centrix turns it into a structured channel:

• Department channels (dispatch, fleet, safety, accounting, HR)
• Direct messages (driver-dispatcher, dispatcher-fleet, etc.)
• Bot channels (after-hours, weather, fuel optimization)
• All messages logged, searchable, DLP-scanned, and tied back to the

underlying load/truck/driver record

The 30-layer Telegram security policy (`policy/telegram.yaml`) enforces rate limits, content protection, input sanitization, and chat-id whitelisting so the bot ecosystem can't be abused.

Where to start

If you're running a 50+ truck carrier with after-hours coverage:

• The weather auto-call is the highest single-event ROI — one avoided severe

weather incident pays for the year.

• The customer-service triage is the highest weekly ROI — 70% of tickets

resolving themselves over the weekend frees Monday morning.

• The load fall-out workflow is the highest dispatcher-satisfaction ROI —

most dispatchers list "after-hours coverage anxiety" as their single biggest stress source, and removing it changes their week.

See the agent layer — bring two weeks of after-hours incident logs and we'll show you exactly which workflows would have run agent-led.