Most AI-for-business advice is written for companies with a transformation office and a year to spend. If you run a mid-market company ($50M to $500M ARR), you need something narrower and more honest: which one or two workflows AI actually pays back in, what they cost to run, and how to ship them before the opportunity ages out. Here is the version that survives contact with a real P&L.
Not everywhere, and not the flashy places. The workflows that return are usually high-volume, rule-heavy, and text-shaped, where a small time saving multiplied by thousands of repetitions becomes real money.
| Pays off | Usually does not (yet) |
|---|---|
| Support triage, first-draft replies, ticket routing | Anything where a subtle error is expensive and unrecoverable |
| Document extraction, summarizing, data entry from text | Judgment calls that are legally or financially your responsibility |
| Sales and marketing drafting, research, personalization | Work with no clean data source to read from |
| Internal search across your own scattered knowledge | A workflow no one will actually change their behavior to use |
The honest cost of AI in a business is not the subscription. It is inference at volume, the engineering to integrate it into the systems people already use, the human review that does not go away, and the maintenance as models and data change. A workflow that looks free at the demo has a real total cost of ownership, and the ones worth doing are the ones whose return clears that bar, not just the sticker price. Buying a pile of tools is not a strategy. Naming the workflow, the return, and the owner is.
Between 80 and 95 percent of AI pilots never reach production, and almost never because of the model. They die on the org side: no owner whose job depends on shipping it, data that was clean only for the demo, and output that lands in a dashboard nobody opens. The failure is organizational, not technical, which is good news, because organizational problems are visible in advance and cheap to check before you spend.
Do not start with a platform or a strategy deck. Start with a two-week diagnostic that ranks your candidate workflows by return net of running cost, checks whether the data is ready, and gives you an honest go or no-go on each. Then build the top one to production with an engineer who owns the outcome. One shipped workflow that pays back beats ten pilots that impressed a steering committee.
That diagnostic is the AI Readiness Assessment: two weeks, $2,500, a ranked plan with ROI, TCO, and a build path, credited in full toward the build. Related: enterprise AI strategy for the mid-market and build the first one in house or with a team.