Launch day gets the screenshots. Day two determines whether an AI agent becomes part of the business or another demo people stop trusting. At Capstone Plumbing, the useful work after launch has been turning staff corrections, production misses, and changing ServiceTitan records into a system that gets better without becoming fragile.
The first version is a starting line
Capstone's production agent archives customer calls, transcribes trade-specific language, and keeps an AI summary pinned to each customer record in ServiceTitan. It gives technicians durable customer memory in the system they already open before a job. That is the shipped capability. It is not the end of the work.
Real usage immediately produces evidence a test set cannot invent. A technician recognizes that “recirc” was transcribed as an ordinary word. An office employee spots two customer records that refer to the same property. A summary includes a true detail that is no longer relevant. Each correction reveals a different layer of the system that may need to change.
Corrections need somewhere durable to go
If a staff correction only fixes one summary, the team will make the same correction again. If it only gets pasted into a longer prompt, the prompt eventually becomes a junk drawer. The goal is to turn each useful correction into the right kind of durable asset.
Trade vocabulary belongs in transcription cleanup and terminology dictionaries. A customer matching mistake belongs in identity and reconciliation rules. A preferred summary format may belong in the presentation layer. A newly discovered failure becomes an evaluation case so the same bug cannot quietly return after the next change.
A correction becomes valuable when it improves the next thousand runs, not just the one record someone happened to notice.
Review misses on an operating rhythm
Production improvement cannot depend on someone remembering to check the agent. The useful rhythm is a short, recurring review of corrections, unresolved exceptions, cost changes, and representative traces. Group the misses by pattern rather than treating every example as a separate emergency.
That review asks practical questions. Is the same term being misheard? Are summaries stale because a source watermark changed? Did an API field disappear? Are employees editing a section because the content is wrong, or because the format slows them down? Which misses carry business risk, and which are cosmetic?
The answers create a ranked change list. High-frequency, high-consequence patterns move first. Each fix ships with a replay against known examples and a check against adjacent behavior. Improvement becomes controlled software delivery, not live prompt experimentation.
Put the fix in the right layer
Not every miss is a model problem. Prompts are good for instructions, priorities, and output structure. Tools are better for retrieving an invoice, checking a membership term, or writing a marked section back to ServiceTitan. Data pipelines should handle duplicate events, changed records, and source-of-truth reconciliation. Deterministic business rules should own exact dollar figures and permission boundaries.
Sometimes the business process itself is the issue. If two teams use the same status to mean different things, no prompt can make that ambiguity reliable. Day-two work often uncovers those hidden process disagreements. Fixing the source workflow gives both people and the agent a better system.
Detect drift before trust erodes
A production agent can degrade even when its code does not change. ServiceTitan fields and APIs evolve. Capstone's services, pricing, membership terms, and vocabulary change. Call patterns shift with seasons. A model provider can update behavior behind an endpoint.
That is why the agent needs more than uptime monitoring. Track queue age, source freshness, empty or malformed outputs, correction rate, escalation rate, and evaluation performance. Keep replayable traces so a sudden change can be tied to the input, retrieved records, model output, and final write. A green server is not enough if the workflow is quietly producing worse customer memory.
Protect the trust already earned
Capstone's agent marks the sections it owns and never overwrites what a human wrote. That boundary matters as much after launch as it did on day one. Improvements should not widen permissions by accident or trade traceability for convenience. Every new capability has to preserve idempotent writes, source-backed facts, and a clear route for human exceptions.
This is how autonomy expands safely. The agent proves it can handle a bounded workflow, its misses become tests, the tests support a better version, and reliable performance earns the next piece of scope. Trust grows from evidence instead of optimism.
The compounding value is operational
After launch, Capstone does not just own a transcript archive and a summary generator. It owns a growing vocabulary, customer-matching logic, production traces, evaluation cases, and a repeatable way to turn staff judgment into system behavior. Those assets stay useful even as models change.
That is the difference between maintaining an AI feature and operating an agent. Maintenance keeps the first version alive. Operations make the system more accurate, more economical, and more aligned with the way the business actually works.
Read how the original Capstone deployment was built, or talk to Foundation about a managed agent that improves after launch instead of aging in place.