Every workflow below follows the same underlying shape: something that used to take a person real, dull time now happens with a person checking the output rather than producing it from scratch. None of them remove the human, and that's deliberate; every one keeps a person reviewing what actually gets sent or decided. These are illustrative patterns, not case studies from named businesses.
Voice note to quote, before leaving the van
A tradesperson doing several jobs a day writes quotes from memory after getting home, and loses some of them to the end of a long day. A typical workflow: talk into a phone for a minute or two describing the job, an AI assistant turns it into a structured quote, and it goes out before leaving the customer's drive rather than being written up hours later.
Clinical notes that write themselves
A healthcare or veterinary practice loses real time every day to typing up notes after appointments. A typical workflow: ambient note capture during the consultation, producing a structured summary the practitioner reviews and approves, rather than a blank page at the end of a long list of patients.
A tender response pipeline
A services business rewrites the same set of standard questions from scratch on every tender or proposal, which takes real time and limits how many it can respond to. A typical workflow: a reusable knowledge base of past answers, with AI drafting new sections from it, often cutting response time substantially and making it possible to respond to more opportunities.
An inbox that triages itself
A growing support inbox leaves a small team drowning and response times stretching out. A typical workflow: incoming messages are automatically sorted and a reply is drafted, and a human approves or edits before it sends, generally cutting response time substantially without adding headcount.
A single source of truth across several sites
A multi-site business has no visibility across locations, with each site effectively running its own version of things and problems only surfacing at the end of a quarter. Connecting the sites into one system with a regular performance digest tends to catch cost or performance drift far earlier than periodic manual reviews would.
What these have in common
Every one of these starts as one specific, named frustration, not a general ambition to "use more AI." Working out what your own version of that frustration is actually costing you is usually the first useful step, and it's the same question in every case: which task, done how often, costing how much time.