Processes we've automated
Four cases, each tested against real or synthetic data before it went live.
From inbox triage to auto-drafted replies
A support team was manually reading and routing every inbound email. Every message, no matter how routine, went through a person first. We built a classifier that separates the clear cases from the ambiguous ones, auto-drafting replies for the former and routing only the latter to a human.
Challenge
- Every inbound email required manual reading before it could be routed or answered
- Response time depended entirely on staff availability
- Repetitive, low-ambiguity requests consumed the same attention as genuinely hard ones
- No consistent record of how routing decisions were made
Solution
- A classifier trained on the team's own historical email volume
- Auto-drafted replies for clear-cut cases, held for ambiguous ones
- Ambiguous cases routed straight to a human, with the system's reasoning attached
- Validated by comparing the classifier's routing and drafts against what the team actually did on historical emails, before going live
Results
- From 20 hours a day of team-wide triage to 2 hours a day
- 85% of auto-drafted replies sent without edits
- Average response time down from 3 days to under 24 hours
Timeline & scope. 8 weeks, deployed on-premise.
Automated supplier outreach, still human-approved
A services company was requesting quotes from suppliers, tracking every response, and building budgets by hand. We built a pipeline that drafts the outreach, tracks supplier responses, and generates a consolidated quote for review.
Challenge
- Requesting and following up with suppliers was done manually, one email at a time
- Tracking which suppliers had responded, and with what, lived in someone's head or an inbox
- Budgets were assembled by hand from scattered supplier replies
- Too few historical tenders existed to validate a new system with a simple pass or fail test
Solution
- A pipeline that drafts supplier outreach and tracks responses as they come in
- Automatic generation of a consolidated quote from supplier replies
- Validated using synthetic negotiation scenarios built with known correct answers, since real historical tenders weren't sufficient on their own
- A human reviews every quote before it goes out, the system doesn't send anything unchecked
Results
- Quoting cycle time down from 12 hours to 4 hours
- Nearly double the suppliers handled per cycle with the same team
Timeline & scope. 12 weeks, deployed on-premise.
A self-improving loop for ad copy
A marketing team was A/B testing ad copy one variant at a time, waiting on each result before trying the next. We built a self-improving loop that generates, tests, and refines copy against real performance data.
Challenge
- Copy testing happened one variant at a time, serially
- Each testing cycle required manual setup and manual review of results
- Winning patterns from past tests weren't systematically fed into the next round
Solution
- A loop that generates copy variants, tests them, and refines based on real performance
- Each cycle's results feed directly into the next round of generation
- Performance measured against real campaign data, not a synthetic proxy
Results
- Testing cycle time down from 2 weeks to 1 week
- 18% improvement in click-through rate over the previous manual process
Timeline & scope. 26 weeks, deployed on-premise.
Evolving hand-tuned heuristics against real production data
An industrial line relied on heuristics that had been hand-tuned over time. We built a self-improving search system that evolves those heuristics against real production data, tracing every step so a genuine improvement can be told apart from a lucky run.
Challenge
- Heuristics had been tuned manually and infrequently, based on intuition rather than systematic testing
- No structured way to tell whether a change was a real improvement or noise
- Improving the heuristics further required deep, slow manual analysis
Solution
- A search system that evolves the existing heuristics against real production data
- Full tracing of every step the system takes, not just a final score
- A structured way to distinguish a real improvement from a lucky run before adopting a change
Results
- 5% measured improvement over the original heuristics
- No manual tuning process existed before, this is the first systematic one
Timeline & scope. 13 weeks, deployed on-premise.