Across six engagements, the gap was rarely missing software. The highest-value workflows leaned on one senior person who knew which spreadsheet to open and which paragraph to paste. The exception nobody had documented, and what came next, lived in their head.
The common leak
A quoting estimator goes out for two days and three deals stall. Revenue leaks when quoting, proposal response, onboarding, reporting, or follow-up runs through one experienced operator. The system holds right up until that person is out for the day or buried under a different fire.
AI earns its place when it captures the repeatable part of that operator's judgment and drops it into the workflow the next person already opens every morning.
What fixed it
None of the fixes were generic chatbots. They were narrow tools: quoting apps, RFP draft queues, intake routers, enrichment loops, and handoff runbooks, each one wired to a bottleneck someone could measure.
Which model sat underneath, GPT or Claude, mattered far less than where the workflow drew its edges. The system that won was the one the team could explain in a sentence and then keep improving after we left.
The lesson
Start with the bottleneck that already has demand behind it. Build only the piece that changes the next decision or kills the next delay. Nothing else.
Do that and AI starts paying for itself in a services business inside the first quarter.
