Technology & AI
Most AI projects that fail commercially did not fail technically. They were aimed at work that was never expensive enough to automate.
The question that separates a useful AI project from an impressive one is not what the model can do. It is which recurring task currently consumes paid hours, happens often enough to measure, and produces an output someone can check. If a task fails that description, automating it saves nothing that shows up in a management account.
Four conditions worth checking first
- The work repeats. Something done fifty times a week can be measured before and after; something done twice a year cannot.
- There is a cost attached today. Salaried hours, an outsourced team, a per-ticket fee - a number that already exists in a budget line.
- The inputs exist in a system. Email, a CRM, a document store, a database. If the information lives only in someone's head, the first project is capturing it, not automating it.
- The output is checkable. A draft reply, an extracted field, a classification - something a person can accept or reject, which is also what produces the training signal for the next iteration.
If nobody can name the cost the automation removes, there is no project - only a demo.
The work that usually qualifies
In the businesses we build for, four categories come up repeatedly: first-line support triage, where volume is high and answers are bounded; document and invoice extraction, where the input is structured enough to validate; sales follow-up and qualification, where speed of response drives conversion more than wording does; and internal reporting, where the same numbers are assembled by hand every week.
None of those are glamorous. All of them have a cost line that can be shown to shrink, which is the only argument that survives a second budget cycle.
Build the measurement before the model
Before anything is automated, the current state needs a number: handling time, cost per ticket, error rate, hours spent. Without it, the project cannot be evaluated, only defended. With it, the decision to extend or stop is arithmetic.
The same discipline applies to the technology choice. An AI agent that needs a human to approve its output is not a weaker system than an autonomous one; in most commercial processes it is the correct design, because the approval step is where liability sits.
Where this leads is unglamorous and profitable: a handful of narrow, instrumented automations inside processes that already cost money, each one paying for itself before the next is started.
Related
Written by the team in New York - engineers, strategists and marketers who build and grow the companies they write about.
