How do you choose the first AI use case?
Score candidate use cases on two axes: volume (how often the work happens) and risk (what goes wrong if the AI gets it wrong). The first pilot belongs in the high-volume, low-risk corner, where a small improvement repeats thousands of times and a mistake is easy to catch and reverse.
| Use case | Volume | Risk if wrong | Good first pilot? |
|---|---|---|---|
| Outage and slow-speed contacts | High, spikes during incidents | Low: a person can follow up | Yes |
| Appointment confirmation and rescheduling | High | Low: a missed confirmation is visible | Yes |
| Payment reminders with the operator’s own payment link | High, every cycle | Low to medium: tone and timing matter | Yes, with clear rules |
| Plan changes and upgrades | Medium | Medium: pricing must be exact | Second phase |
| Retention offers | Medium | Medium: offers cost margin | Second phase |
| Credits, disputes, legal requests | Low | High | Keep with people |
What data does AI need before it can do real work?
AI can only act on what it can see. If balance lives in the billing system, the device in a provisioning tool and the contact history in a help desk, the agent either can’t resolve anything or needs brittle integrations to three places for every case.
| Check before the pilot | Why it matters |
|---|---|
Is there one subscriber record with plan, balance, service location and equipment? The agent needs the full picture to take the right action. | The agent needs the full picture to take the right action. |
Can service state be read and changed through an API? Suspend, reconnect and device actions are where time is saved. | Suspend, reconnect and device actions are where time is saved. |
Are contacts (calls, texts, tickets) linked to the subscriber? Context for escalation, and the data to measure outcomes. | Context for escalation, and the data to measure outcomes. |
Are outages tracked by service area? Lets the agent confirm a known issue instead of opening a ticket. | Lets the agent confirm a known issue instead of opening a ticket. |
Are your policies written down (verification, credits, dunning rules)? The agent follows rules; unwritten rules can’t be followed. | The agent follows rules; unwritten rules can’t be followed. |
Which guardrails should be in place from day one?
- Authorization limits
- A written list of what the agent may change on an account, and above which amount or risk it must hand over.
- Authentication
- The same subscriber verification your team uses, before any account action.
- Audit trail
- Every action the agent takes, logged against the subscriber record, so any decision can be reviewed.
- Human handoff
- A clear route to a person at any point, carrying the context the agent already gathered.
- Consent and messaging rules
- Outbound calls and texts follow the consent and registration rules that apply to your business messaging.
Write the guardrails before choosing a vendor. They become your evaluation checklist, and any tool that can’t enforce them is out.
What does a 90-day AI rollout look like?
| Days | Focus | Output |
|---|---|---|
| 1–15 | Pick one use case, write the policies and guardrails, confirm data access | A one-page scope with success metrics |
| 16–30 | Configure the agent on real account data, test on internal accounts | Test log of resolved and escalated cases |
| 31–60 | Pilot on a share of live traffic or one channel, review escalations weekly | Weekly resolution, FCR and CSAT versus the human baseline |
| 61–75 | Fix gaps found in review, widen the traffic share | Updated policies and handoff rules |
| 76–90 | Decide: expand to full traffic, add a channel, or pick the second use case | A go or no-go with numbers behind it |
| ID | Event | Detail | Status | Value |
|---|---|---|---|---|
| #48201 | Auto-suspend | Past due · 4 days · policy DUNN-04 | Suspended | −1 service |
| #48202 | Auto-reconnect | ACH payment cleared · $74.00 | Reconnected | 2 min |
| #48203 | Cycle invoicing | Taxes, USF and E911 applied · emailed | Issued | 2,418 docs |
| #48204 | Payment reconciled | Paid via reminder link · matched to invoice | Cleared | +$74.00 |
How do you know when to expand?
Compare the pilot against the same contact type handled by people over the same period. Expand when resolution holds up, repeat contacts don’t rise and satisfaction is at least level. If one of those slips, fix the cause before adding volume; scaling a pilot that half works only scales the rework.
- Resolution rate for the use case, checked against repeat contacts within seven days.
- Escalation quality: how often the person picking up the case had to ask the subscriber again.
- Satisfaction for AI-handled versus human-handled contacts of the same type.
- Cost per resolved contact, including the human time spent on escalations.
What are the honest risks?
AI makes mistakes, and it makes them confidently. The risk is manageable when mistakes are caught early (audit trail, weekly review), limited in impact (authorization limits) and easy to escape (a visible route to a person). It becomes a real problem when an agent is launched across every channel at once, with broad permissions and no one assigned to review what it does.
Where does Inerxia fit?
Inerxia runs the operating system and the AI agents over one subscriber record, so the data-readiness problem is solved by design. Agents cover Tier 1 support, retention, billing and collections, field operations and upselling, each with subscriber authentication, limits on what it can change, escalation with full context and a full audit trail.
Deployment follows the same logic as this roadmap: four weeks to your first full billing cycle, run in parallel before cutover, then agents added one use case at a time. Agents start at $500 each; see pricing.