The provider silently updates the model under your prompts
Same prompt, different distribution. Accuracy slides 4 to 8 points in a quarter. The eval suite would catch it, if anyone were running it.
Eval suites rot. Prompts drift. Vendors deprecate the model under you. Training material goes stale the week after rollout. 2 to 3 days a month covers the monthly health-check, eval re-runs, prompt regression triage, vendor escalations, and a Sev-1 SLA inside one business day. The continuity tier below the active engagements.
A monthly health-check artifact, a Sev-1 SLA, and a system that doesn't quietly decay between board meetings.
Production AI is a system with seven failure modes. They don't announce themselves. They surface as a Sev-1, a board question you can't answer, or a slow-bleed cost line nobody traced back to the source.
You don't need another build. You need someone watching what you already shipped, on a named cadence, against a baseline.
Same prompt, different distribution. Accuracy slides 4 to 8 points in a quarter. The eval suite would catch it, if anyone were running it.
Six engineers tweak the same prompt over six months. Nobody owns the regression suite. The prompt ships drift as a feature.
90-day deprecation window. No migration plan. The replacement model has different tool-calling behavior and your evals fail on day one.
A CRM field gets renamed. A pricing table gets reshaped. The agent keeps running, returns plausible answers, and silently corrupts the next step in the workflow.
No owner. No cadence. By month four, the eval suite is a folder in the repo nobody opens. Regressions ship to production unflagged.
Latency spikes, cost overruns, retry storms, runaway loops. The dashboards exist. There's no human in the loop reviewing them on a cadence.
Training material captured at launch goes stale the first time a workflow changes. RBAC roles drift as the org reshuffles, and the AI inherits access nobody meant to grant.
Each deliverable below has a frequency, an artifact, and an owner. 2 to 3 days a month, scope-locked. No creep into new builds.
Accuracy, latency, cost, error rate, and adoption against the locked baseline. Delivered by the 5th business day each month. Forwardable to the board without a redraft.
Same golden set, same scoring rubric, same cadence. Drift flagged with a delta number, not a feeling. The suite stays warm instead of rotting in the repo.
Anthropic, OpenAI, inference provider sunset calendars tracked. Migration plan drafted before the 90-day window opens, not after. You hear about deprecations from me, not from a Twitter thread.
When a prompt breaks, I reproduce in the eval suite, ship the fix, re-baseline. No tickets to your engineering team. Logged in the monthly health-check.
Quota issues, billing surprises, rate-limit changes, tool-calling regressions, security questionnaires. I negotiate. Your team stays out of vendor inboxes.
Production AI broken in a user-visible way is the SLA event. I'm on it inside one business day. If not, the next month's retainer is waived.
The AI-ready knowledge base your team and your AI both work from gets a quarterly pass. New workflows captured, retired ones removed, ownership reassigned. Onboarding stays sub-week.
When a model version, prompt, or workflow changes, the team-facing training gets a targeted update. No annual rip-and-replace. The Leadership-Lab-Crowd structure stays current.
As the team reshapes, RBAC drifts. Quarterly review of who and what the AI can touch. Decommissioned roles pruned. Audit-ready.
A single new prompt or workflow, sized to the day allowance. Same eval discipline as the original build. Anything larger flows to Scoped Build.
Async-first, calendar-confirmed. The monthly health-check is the agenda.
Three months of shipped, broke, next, in board-ready language. The CFO can defend the line item without a translator.
Scope-locked on purpose. Work grows, engagement graduates. Not the retainer.
One small addition fits. A new multi-agent workflow does not. That's Scoped Build: 60 to 90 days.
Helping you hire a full-time AI lead is Fractional work. Different altitude.
Board-ready roadmaps and Three-Horizon classification are their own engagements. Continuity keeps what you shipped running.
Leadership-Lab-Crowd training is a 4 to 8 week engagement. Not a maintenance line item.
Scope-locked at 2 to 3 days per month. Price scopes against system complexity, eval suite size, and vendor count. Quoted in writing within 48 hours of the discovery call.
Scope-locked at 2 to 3 days per month. Quoted in writing within 48 hours of the discovery call.
First month is reading the system. Month two catches the first drift. Month three is the baseline. Cancel anytime after with 30 days' notice.
Anything shorter is paying for ramp without the payoff. The first drift report only means something against month one's reading.
Past the cap, the work isn't Continuity. It's Scoped Build or Fractional. The retainer doesn't quietly expand. The price is honest.
Continuity insurance for production AI you already shipped. 2 to 3 days per month covering: a monthly health-check artifact, eval suite re-runs against the locked baseline, prompt regression triage, model-deprecation early warning, vendor escalations, quarterly tribal-knowledge inventory refresh, and workforce retraining touch-ups when models or workflows change. One small scope addition per month fits inside the day budget. Sev-1 incidents are acknowledged within 1 business day. Built for the gap between shipping and hiring a full-time AI lead.
3-month minimum. Monthly renewal after. Scope-locked at 2 to 3 days per month. Price scopes against system complexity, eval suite size, and vendor count. Quoted in writing within 48 hours of the discovery call.
Clients who shipped a Scoped Build or finished a Fractional engagement and need someone keeping the system running. Production AI is in place. A full-time AI lead at senior-exec comp is the wrong shape. AI System Continuity is the in-between.
Fractional is Director-altitude leadership: org redesign, hiring, board work, new builds, vendor strategy. AI System Continuity is operational maintenance only: eval monitoring, prompt triage, vendor escalations. Different altitude, different scope.
Then you don't need AI System Continuity. You need Scoped Build for a new use case, or Fractional for ongoing leadership. I'll tell you in the intro call which one fits. The retainer doesn't quietly expand into the higher tiers.
Two guarantees. SKU guarantee: Sev-1 incidents (production AI broken in a user-visible way) acknowledged within 1 business day, triaged within 3. If not, next month's retainer is waived. Umbrella guarantee on every engagement: a 14-day mutual exit. Either party can end the engagement inside the first 14 days. Days worked are billed at the agreed day rate. No further commitment.
Below the active engagements. Not competitive with them. The continuity tier most clients graduate into after their Scoped Build or Fractional ends.
The AI gets shipped. Eval suite stood up. Team trained. Engagement closes.
2 to 3 days a month. Eval monitoring, prompt triage, vendor escalations. The system keeps running.
Production AI breaks loudly. If I'm not on Sev-1 inside one business day, the retainer doesn't get paid that month. That removes the "ghost vendor" risk from the buyer's pile.
Sev-1 incidents (production AI broken in a user-visible way) acknowledged within 1 business day, triaged within 3. If not, next month's retainer is waived.
Every engagement carries a 14-day mutual exit. Either party can end the engagement inside the first 14 days. Days worked are billed at the agreed day rate. No further commitment.
Book the call. 30 minutes to map your stack, your eval surface, and your vendor exposure. Quoted against the work in writing within 48 hours. 3-month minimum.