Every software vendor you use has bolted an AI assistant onto their product in the last two years. Some of these genuinely save hours a week. Others produce plausible-looking garbage that costs you more time than they save, or worse, quietly makes decisions that should have been yours. The difference isn't the underlying AI. It's the shape of the task. Here's how to tell the two apart before you roll something out to your team.
Where AI assistants actually earn their keep
The wins share a pattern: the AI does the labor, a human keeps the judgment, and checking the output is faster than producing it would have been.
First drafts. A blank page is slow; editing is fast. Assistants are good at turning bullet points into a customer email, a job description, a proposal outline, or a policy document. The draft is 70% right in 30 seconds, and you spend five minutes fixing the rest instead of thirty writing from scratch. This works because you read the whole thing before it goes anywhere.
Lookups against your own documents. "What does our contract with this vendor say about termination notice?" "What did we quote this customer last spring?" Pointing an assistant at your own files and asking questions beats digging through folders, and the answer can cite the source document so you can verify it in one click. The verification step is what makes this safe.
Meeting notes and summaries. Recording a call and getting back a summary with action items is one of the most reliable wins we see. Transcription is mature, summarization of a transcript is a bounded task, and everyone who was in the meeting can spot an error immediately. Tools like the meeting assistants built into Teams and Zoom, or standalone ones like Otter, do this well today.
Triage and sorting. Routing inbound email to the right person, flagging which support tickets sound urgent, tagging expenses by category. The assistant makes a first pass, humans handle the cases that matter, and a misroute costs you a minor delay rather than real money. Low stakes per decision, high volume: ideal territory.
Where they fail, and why
The failures share a pattern too: the output looks finished, so nobody checks it, or the task required judgment the AI doesn't have.
Hallucinated facts. Language models generate plausible text. Plausible is not the same as true. Ask one for a statistic, a regulation, a case citation, or a product spec and it may confidently produce one that doesn't exist. The failure is worst exactly where it's most tempting: when you don't know the answer yourself and can't spot the fake. Rule of thumb: use assistants to process information you give them (summarize this, draft from these notes), and treat anything they assert from memory as unverified until you've checked a real source.
Unreviewed output going out the door. The technology fails less often than the workflow does. An assistant drafts a customer email, and after a few weeks of good drafts, people stop reading before hitting send. Then one draft misstates your refund policy, and now you've got a customer holding a written commitment you never made. Any AI output that reaches a customer, a vendor, a regulator, or your books needs a named human who reviews it. If nobody has time to review, you didn't automate the work, you just stopped doing it carefully.
Automating the judgment instead of the labor. This is the deepest failure mode. Which candidate to interview, whether to extend credit to a customer, how to respond to a legal threat, whether a safety complaint is serious. These tasks aren't slow because of typing; they're slow because they require weighing things. Handing them to an assistant doesn't speed up the judgment, it removes it. The assistant will always produce an answer, delivered with the same confident tone whether it's right or wrong, and that confidence is exactly what makes it dangerous in judgment calls.
A simple test before you adopt anything
Ask three questions about the task:
- Is checking the output faster than doing the work? If yes (drafts, summaries, sorting), good candidate. If verifying requires redoing the whole task, the assistant saves nothing.
- What does a wrong answer cost? A mis-tagged expense costs a minute. A hallucinated contract term costs a client. Match the stakes to the level of review.
- Is this labor or judgment? Automate the labor. Keep the judgment.
How to know you've done it right
Six months in, a good AI assistant rollout looks like this: specific tasks got measurably faster, and your team can name them. Every output that leaves the building still passes a human. Nobody can point to a decision that got made by the AI alone. And when someone catches the assistant being wrong, that's treated as normal and expected, not as a scandal, because the workflow assumed errors from day one.
If it instead looks like "we pay for five AI tools and nobody's sure what they do," you bought the demo, not the workflow. We help businesses sort one from the other: which tasks in your operation are the drafts-and-lookups kind, which are the judgment kind, and how to wire up the first kind without risking the second.
Stuck on this, or want it done for you? That's the job.
Email us →