All work
B2B Services · B2B services company (anonymized)

AI workflow that replaced three roles

A multi-step AI workflow now handles work that previously required three full-time staff — freeing the team for higher-value work.

  • AI / LLM
  • Automation
  • Internal tooling
−66%Staff cost on the function
~10×Throughput increase
< 6 wksTime to production

Timeline: ~6 weeks · Discovery → shadow run → cutover

The problem

Three full-time staff spent the majority of every week on the same loop: receive document → read → classify → route → log.

Volumes were growing faster than the team could scale, but hiring more people for repetitive work was not the answer the leadership team wanted.

The work was not strictly low-skill — context mattered — but the patterns were learnable from existing examples.

Our approach

  • Mined the team's historical decisions to build a labeled dataset from their past routing choices.
  • Designed a multi-step LLM pipeline: extract → classify → route → human-review-when-uncertain. Confidence threshold tuned with the team so edge cases still escalate to a human.
  • Wired the pipeline into existing tools (no new dashboards to learn) and shipped behind a feature flag for a two-week shadow run before going live.
  • Built a small review console for the team to spot-check decisions and feed corrections back into the system.

The result

The pipeline now handles the bulk of the volume autonomously, with a small human-review queue for edge cases.

The team was redeployed to client-facing work — the headcount cost on the function dropped sharply without anyone losing their job.

New document types onboard in days, not months, because the model picks up patterns from a handful of examples.

They built an AI workflow that now does what three people used to do manually. We re-deployed the team to higher-value work and our monthly headcount cost on that function is a fraction of what it was.
Head of Operations, B2B services company (anonymized)

Stack

  • Python
  • OpenAI / Claude
  • PostgreSQL
  • AWS Lambda
  • Next.js (review console)

Like to talk about something similar?

We only take on a small number of new projects each month. If your team has a problem in this shape — or something completely different — start with a free 30-minute call.

More work