AI is an accelerator, not a strategy: where AI actually pays off in a 50-person company

At around fifty employees, almost every founder gets asked some version of the same question — by a board member, an investor, a peer at a conference — “what’s your AI strategy?” It’s a fair question asked at the wrong altitude. AI isn’t a strategy any more than a spreadsheet is a strategy. It’s a tool that makes an existing process faster, cheaper, or more consistent. Whether it pays off has almost nothing to do with which model you pick, and almost everything to do with whether there’s a standardized process underneath it worth accelerating.

Where it actually pays off

In a fifty-person company, the highest-value AI applications are rarely glamorous. They’re the document-heavy, repetitive, well-defined tasks that eat a real chunk of someone’s week without requiring much judgment: summarizing intake forms, drafting first-pass responses to common support questions, extracting data from invoices or contracts into a structured format, reconciling two systems that don’t talk to each other. These are jobs that are boring precisely because they’re standardized — the same steps, the same inputs, over and over — which is exactly what makes them good candidates for automation. A standardized process is legible to a model. A chaotic one isn’t.

The pattern holds in reverse too. Decision support — helping a person make a judgment call faster, with better information — tends to work well when the underlying data is trustworthy and the decision criteria are reasonably clear. It tends to work badly when it’s asked to paper over messy, contradictory source data, because it will confidently produce a wrong answer instead of an honest “we don’t have good data for this.”

Where it quietly makes things worse

Applied to a process that isn’t standardized yet, AI doesn’t fix the chaos — it automates it, faster and with more apparent confidence. A reporting process where three people define “revenue” three different ways doesn’t get more trustworthy because a model is now generating the summary; it gets a more polished-looking wrong answer, delivered faster. A hiring process with no consistent bar for “good” doesn’t get better because a model is screening résumés against vague criteria; it gets the same inconsistency, scaled up and harder to spot because it now looks systematic.

This is the trap in the board-level question. “What’s your AI strategy” implicitly assumes AI is the bottleneck. Often the actual bottleneck is one step behind that: the process isn’t standardized, or the data isn’t trustworthy, or nobody’s actually agreed on what the output should look like. Applying AI at that point doesn’t remove the bottleneck. It just makes the output of the bottleneck arrive faster.

The three questions worth asking before any AI project

Before scoping an AI implementation, it’s worth answering three things honestly. Is the process this would touch already standardized — the same steps, done the same way, most of the time — or does it still vary by who’s doing it? Is the underlying data trustworthy enough that accelerating it produces a better answer rather than a faster wrong one? And is there a clear, specific business outcome this is meant to change — hours saved, error rate down, turnaround time down — or is the goal simply “use AI,” which isn’t an outcome at all.

If the honest answer to any of those is no, the higher-leverage move is almost always to fix the process or the data first — which is often a reporting or operations fix, not a technology one — and revisit AI once there’s something standardized worth accelerating.

We start every applied-AI engagement from the bottleneck, not the model, for exactly this reason — see Technology for how that scoping works. And because so many “AI problems” turn out to be reporting or process problems in disguise, it’s worth reading how we diagnose that upstream, on How We Work.