AI Operations Findings and Bot Drafting

AI operations is the Botlit concept for a copilot that watches a whole fleet of bots and agents, tells an operator what changed, and helps them build the next one. The illustration shows an operations view for a sample agency workspace, with its entry in the left navigation carrying a count badge that matches the number of findings, so waiting work shows in the navigation itself.

The problem is scale. A team running assistants for several clients cannot read every execution record, so regressions surface late, usually through a complaint. The findings feed inverts that. Each card states a finding in a short headline, carries a severity badge, and backs the claim with an evidence block written in numbers: a failure rate that climbed after a new version was published, with the baseline for comparison; an average cost per run that tripled over a day after a model change; a grounding problem where a large share of runs now retrieve nothing where few did before; and a knowledge base stuck indexing with none of its documents embedded. Every card also carries a specific recommendation, such as rolling back to the previous version, downshifting to a smaller model, lowering a similarity threshold or rebuilding an index, and a pair of accept and dismiss actions so each finding ends in an explicit decision. Filters above the feed narrow it by severity or by theme, covering cost, grounding and reliability.

The right side of the illustration handles creation rather than repair. A describe-a-bot panel takes a plain-language description, in the sample an assistant that answers policy questions from a handbook and escalates payroll issues to a person, and a draft action turns it into a proposed build. What comes back is a checklist rather than a finished bot: a bot app and two specialist agents in a draft state, a routing link between them, an existing knowledge base to attach, and a model already registered in the workspace, each with its own approve control.

The draft heading and a footer note make the same point: no drafted item is live or published until it is approved individually. That keeps the copilot advisory. Its findings use the vocabulary found elsewhere in the workspace, from versions, models and runs to retrieved chunks and knowledge base indexing, its draft reuses an existing knowledge base and a registered model, and the operator keeps the final decision.