AI that actually saves time (and the kind that doesn't)
Most AI features added to internal tools save zero time. A few save hours per person per week. Here's how to tell them apart before you commission either.
Every internal tool brief now contains the phrase "with AI". Sometimes that's the right answer. Often it's a hammer looking for a nail. The pattern that works is unglamorous.
We've now shipped AI features into about thirty internal tools. The ones that genuinely save time look nothing like the ones that demo well. Here's the split.
What doesn't work
A chatbot bolted to a dashboard. Nobody wants to type a question when they could click a filter. Chat is great when the input is unstructured (long text, an email, a document). It's worse than a dropdown for structured data.
"AI insights" widgets. Vague summaries generated from the same data the user is already looking at. They read well in a demo. In production, people stop reading them within a week , partly because the insights repeat, partly because nobody trusts a sentence that's also occasionally wrong.
Generation-for-generation's-sake. Auto-drafting emails that humans always rewrite anyway. Net time saved: negative , because reading and editing a wrong-shape draft takes longer than starting from scratch.
"Ask anything" search over the whole database. Works in a demo with three records. Falls apart on a real dataset, because the model can't tell which of the six tables called `users` is the one you meant.
What works
Classification at the edge of the workflow. Routing an inbound email to the right team. Tagging a support ticket. Detecting whether an invoice matches a PO. Boring, fast, measurable. Usually a small model, often deterministic enough that you can show your working in the UI.
Extraction from unstructured documents. Pulling fields out of PDFs, contracts, screenshots. Hours per person per week, week in week out. This is the single highest-ROI AI feature we ship.
Summaries of long inputs. Call transcripts, support threads, multi-page reports. Read time drops 80% and the human still makes the call. Show the source alongside the summary so people can verify in two clicks.
Search across the company's own knowledge. "What's our policy on X?" answered in two seconds instead of three Slack messages. Scoped tightly , a single knowledge base, not the whole intranet , the precision is high enough to be trusted.
Drafting on a clear template. Not "write the email". "Fill in this template with the customer's data and tone." The model is constrained, the output is predictable, the human still owns the send.
Why the difference
The features that work share a shape: the AI replaces a specific, repetitive, measurable human task, the user sees the input the model saw, and a wrong answer is obvious. The features that don't work try to generate something the user can't easily verify , so trust collapses the first time the model is wrong, and the feature is dead.
The test
Before adding an AI feature, ask: what specific thing is a person doing today that this replaces? Measure the time it takes them. If you can't, don't build the feature.
Then ask a second question: when this is wrong, will the user notice? If the answer is "only if they read the source carefully", you're building a confidence trap, not a tool.
A rule of thumb
Spend AI budget on the boring 80% of the workflow , the bit where humans are doing repetitive, structured-enough work on unstructured-enough inputs. Leave the headline-grabbing demo features to the SaaS vendors. The hours come back from extraction, classification, summarisation and search , not from a chat icon in the corner.