You don’t need an LLM everywhere, you need it where it adds value
Bolting an LLM onto every screen is easy and mostly worthless. The real question is narrower: where is your team losing time, and does a model actually help in that exact workflow?
There is enormous pressure right now to “add AI” to everything. A chat box in the corner of every screen, a summarise button on every page, a copilot for every tool. Most of it is theatre. It demos well, it ships in a sprint, and it changes almost nothing about how much work actually gets done.
The mistake is starting from the technology. An LLM is a capability looking for a problem; sprayed across a product uniformly, it lands mostly where it isn’t needed and misses the few places it would have been transformative. The value is never evenly distributed, so your effort shouldn’t be either.
Start from where time is lost
The right first question is not “where could we put an LLM?” but “where is my team actually losing time?” Watch the real workflow. Find the steps that are slow, repetitive, and cognitively cheap but time-expensive, the reading, the re-typing, the reconciling, the drafting, the looking-things-up. That is your shortlist. Everything else is a distraction.
Then, for each candidate step, ask a harder question: what is the value of an LLM here, specifically? Not in general, here, in this step, for this person, measured against the time or error it removes. If you can’t articulate the value in a sentence with a number in it, it probably isn’t there.
The value of AI is never evenly distributed across a product, so your effort shouldn’t be either.
Augment the workflow, don’t invent a new one
The most common failure is using AI as an excuse to design a brand-new workflow that nobody asked for. People already have a way of working; it carries context, habits and trust. The highest-return move is almost always to drop the model into the existing workflow, at the exact step that hurts, so it quietly removes friction the user already feels.
A model that saves a person ninety seconds on a task they do forty times a day will be loved and used. A shiny new AI workspace that asks them to abandon how they work will be opened once and never again. Augmentation beats reinvention almost every time.
- Map the real workflow and find where time is actually lost.
- For each step, state the specific value a model would add, with a number.
- Embed the model into the existing step; don’t invent a new workflow around it.
- Ignore the steps where AI is merely possible but not valuable.
- Measure the before and after, if the time saved isn’t real, remove it.
Fewer, deeper, measured
Ten shallow AI features that each save nobody any real time are worse than one that removes a genuine bottleneck, they add maintenance, confusion and risk for no return. Do fewer things, deeper, and measure them. The same discipline we apply to shipping agents responsibly, and to deciding what to build at all, applies here in miniature: intelligence is only worth adding where it is grounded, useful and measured.
None of this is anti-AI. It is pro-value. Used where it belongs, the slow, repetitive, high-volume steps in a workflow people already trust, an LLM is a genuine multiplier. Used everywhere, it is mostly noise. The work is in telling the difference, honestly, before you build.
