Service levels The operated subscription, in writing
What I commit to in writing, and what I don’t.
I commit to response windows I control, in business hours, Pacific time. I do not commit to uptime, because the vendors under the system own that and publish their own figures. The exact windows are set per system in your proposal, because the vendor’s published service level bounds them. This page is the structure every proposal follows.
01 Severity and response
Three severities, three windows, one fallback.
Business hours are Monday to Friday, Pacific time, excluding US federal holidays. Outside them, the fallback is on your side and it is written in your fix-it guide: who puts the system into safe mode, and how.
| Severity | What it looks like | What happens |
|---|---|---|
| S1 · Down or wrong | The system is unreachable, or it is giving customers wrong answers, misrouting calls or writing bad records. Callers notice. | Acknowledged inside the S1 window. Safe mode first: the fallback in your fix-it guide puts the system into a state that cannot do harm, then the fix. |
| S2 · Degraded | Slower, partly failing, answering from stale context, or a vendor incident upstream. Work continues with friction. | Acknowledged inside the S2 window. A workaround or a fix target, in writing, before the end of the next business day. |
| S3 · Cosmetic or request | A wording change, a new question for the knowledge base, a report tweak, a question from your team. | Acknowledged inside the S3 window. Scheduled into the monthly tuning pass, or sooner if it is five minutes. |
02 The monthly report
The report is the deliverable. Same metrics every month.
Generated from telemetry by my own automation, read and signed by me, sent on the first business day of the month. If only one thing from me gets read each month, it is this page.
| Metric | Applies to | What it tells you |
|---|---|---|
| Answer and containment rate | Voice and chat intake | How many conversations the agent handled to the end, how many it handed to a person, and why. |
| Routing accuracy and latency | Intake and automations | Did the record land in the right place, with the right fields, and how long did it take. |
| Grounding score | Document and knowledge assistants | The golden-question set, run against this month’s context. The score, the misses, and what was fixed. |
| Incidents by severity | Every system | Each one with the time to acknowledge, the cause, and the change made so it does not repeat. |
| Context changes and their evaluation deltas | Every system | What changed in the knowledge base or the prompts, and what the evaluation score did when it changed. |
| Spend | Every system | Model, telephony and platform spend, by system, next to last month. |
| Upstream status | Every system | The vendor incidents that touched you, against the vendor’s own published service level. |
03 Incidents
Named, classified, written up.
An AI incident is not only an outage. Incidents are classified with the taxonomy of the CoSAI AI Incident Response Framework V1.0, OASIS, Oct 2025: data, model, deployment, infrastructure and user-interaction incidents. The handling phases follow NIST SP 800-61 Rev. 3, Apr 2025: prepare, detect and analyse, contain and recover, learn.
Notice
You hear from me inside the severity window, by the channel named in the proposal, with what is known and what is being done. Not a status page you have to go and find.
Safe mode first
Every operated system has a state that cannot do harm: the agent hands every call to a person, the assistant answers only from the verified set, the automation pauses and queues. Getting there comes before the fix.
The write-up
Cause, what was touched, what changed, and the evaluation that now covers it. In the monthly report, and sooner for an S1.
Hours on incidents
Counted and reported. If incidents eat more than a fifth of the month, the report says so and proposes what to change: the system, the vendor, or the window.
04 What backs it
One operator is a single point of failure. Here is what is done about that.
- Every commitment is backed onto a published one. Each operated system lists its upstream vendors and their published service levels, and the proposal’s windows sit inside them. Retell AI reliability documentation, for example, states a >99.9% uptime guarantee; a system on that platform inherits that figure, and my window is written for the part I control.
- The fallback is on your side, in writing. One named person at your company knows how to put each system into safe mode. The fix-it guide tells them how, in plain language, and it is tested at handoff.
- The routine runs without me. Evaluations, drift checks, consent sampling and spend telemetry run unattended on my own automation and page on exception. A quiet week is a week in which nothing paged.
- A named peer for continuity. For illness or travel, a named peer operator covers the response windows under a written agreement with me. They appear in your access table as their own named user, only after your written OK, and they never hold your accounts.
- The exit is always available. If the arrangement stops working, you leave with a handoff and a fix-it guide, and the system keeps running, because it was never on my accounts.
05 Kept current
The rules that move, and what is kept current for you.
Boundaries, not legal advice. Your counsel owns the conclusions; the operation keeps the evidence they will ask for. Read from the linked sources on 2026-10-23.
| Rule | What it requires | What the operation keeps |
|---|---|---|
| TCPA and FCC Declaratory Ruling 24-17, Feb 2024 | The FCC confirmed that AI-generated voices are “artificial or prerecorded” under the TCPA. Outbound AI calls need the called party’s prior express consent; marketing calls need it in writing. | No outbound AI call without consent evidence in the ledger. The ledger is yours. |
| CIPA, Cal. Penal Code §§631, 632, 637.2 (In re Otter.AI Privacy Litigation, N.D. Cal., Aug 2026) | All-party consent for recording and for a third party listening in. Statutory damages of $5,000 per violation. In August 2026 a federal court let CIPA claims against an AI notetaker vendor proceed. | An upfront disclosure script on every operated voice agent and notetaker, and the consent evidence kept with the call record. |
| Cal. Bus. & Prof. Code §17941 | A bot used to sell or persuade in California must disclose that it is a bot, clearly and conspicuously. | Every operated bot discloses. It costs nothing and removes the question. |
| 10DLC registration (The Campaign Registry, 10DLC, via Tychron) | US carriers have blocked unregistered application-to-person text traffic on local numbers since 1 February 2025. | Campaigns registered under your brand, on your account. The registration is yours to keep. |
| CPPA ADMT regulations, approved Jul 2025 | California’s rules on automated decision-making technology apply from 1 January 2027, with risk assessments and consumer notices. | A technical evidence pack for your counsel: what the system decides, on what inputs, with what human review. The assessment itself stays with counsel. |
| Bulk-sender authentication | Gmail, Yahoo and Microsoft require SPF, DKIM and DMARC and a working unsubscribe for bulk senders. | DMARC and sender reputation on your domain, monitored, with the report in the monthly pack. |
Ask for the proposal template.
The windows, the metric set and the fallback are the same structure in every proposal. Thirty minutes, no prep.