Where AI actually helps a small team (and where it quietly hurts)

AI · 2026-01-29 · 10 min read

By Arjun Mehta, Engineering & product

After a year of shipping AI features, the pattern is clear: models are excellent at compression and terrible at commitment.

Language models are compression engines. Point one at a long thread and it will hand back the shape of the argument faster than any human skim. Point it at a decision with real consequences and it will produce something confident, fluent and unaccountable.

The trick is to notice which of those two things you are actually asking for. Compression — shorten this, sort this, rephrase this — plays to the model's strengths. Commitment — decide this, guarantee this, be right about this — is where the fluent answer becomes a liability.

The tasks that consistently pay off are summarising, classifying, translating tone, and drafting the first version of something a human will edit anyway. Each has a cheap verification step: you can tell within seconds whether the summary matches the source.

That cheap-verification test is the whole filter. If checking the output takes less time than producing it would have, the model is a genuine multiplier. If checking it takes longer, you have not saved work — you have moved it downstream and disguised it as a draft.

The tasks that quietly hurt are the ones where verification costs more than doing the work. Generated financial reasoning, invented citations and synthesised customer insight all read well and cost hours to check.

These are dangerous precisely because they fail gracefully. A wrong number in a confident paragraph looks exactly like a right one, so nobody double-checks until it has already been forwarded, quoted or acted on. The cost lands weeks later, far from where it was created.

So we design AI features with a visible source. Every summary in our toolkit links back to the paragraph it came from, and every rewrite keeps the original one keystroke away. That single constraint removed most of the trust problem.

The effect on how people use the tools was immediate. When the source is one click away, users actually check it, catch the occasional miss, and come to trust the feature more, not less — because it never asks them to take its word for anything that matters.

The second constraint is that no model output is stored as a fact. Outputs are drafts until a person accepts them, and the interface makes that boundary obvious rather than blurring it for the sake of a smoother demo.

Put together, the rule we keep coming back to is simple: let the model do the typing, never the deciding. Used that way it genuinely helps a small team punch above its size. Used the other way it just generates confident work that someone still has to quietly redo.

Tags: ai, workflow, writing

Arjun Mehta — Arjun writes about web performance, developer tooling and applied AI for ESPYCRUX. He has spent the last decade building and maintaining production web apps, and most of his articles start from something that broke in one of them.