Hey there,
This issue focuses on what AI agents are actually used for in real life, based on behavioural data from a browser-based agent, and what that means for rolling out agents safely in a small business.
By the end, you’ll:
understand where agents earn their keep fast (drafting, summarising, routing, research packs), and where mistakes get expensive (sending, spending, compliance)
have a simple rollout rule: agents for throughput, humans for commitments
get a practical “Tool of the Week” for creators: Meta’s Segment Anything (Audio), a promptable “sound picker” that helps isolate things like barking, typing, or traffic so podcast clean-up and clip-making is faster
Research (Adoption and Usage of AI Agents)
A new paper (The Adoption and Usage of AI Agents: Early Evidence from Perplexity) is out looking at how people use a browser-based AI agent in real life, using behavioural data across hundreds of millions of anonymised interactions.
What it means in practice
An “agent” is basically a loop:
It reads what’s in front of it (web pages, emails, docs, forms).
It decides what matters and what should happen next (classify, prioritise, plan).
It acts inside a tool (draft, fill fields, click through steps, update records).
It has limits (accuracy, permissions, privacy, cost, and human trust).
The headline from this dataset: most agent use clusters around everyday knowledge work, not flashy autonomy. In the study, Productivity & Workflow + Learning & Research make up 57% of agent queries. Usage is also concentrated, with the top 10 tasks (out of 90) making up 55% of queries. That’s useful for businesses because it means a small number of patterns are worth standardising.
For workflows like
Support triage and reply drafting
Lead and account research before outreach
Document editing, rewriting, and summarising
Routine account/admin steps in web tools
The research suggests that agents are strongest at
High-volume, information-heavy work where the output is text or structured fields. The biggest topic share is Productivity (36%), followed by Learning (21%), Media (16%), and Shopping (10%). That mix is a good signal of where agents naturally earn their keep.
And weakest / riskiest at
Any action where a mistake is costly to undo.
This agent can take actions like sending emails, making purchases, booking flights, and editing documents. Powerful, yes. It also means your main failure mode is not “it can’t do the steps”, it’s “it did the wrong step with the right permissions”.
How a small business can use this
Use the “heavy head” insight as your rollout plan.
Start by using agents for:
Drafting (emails, summaries, internal updates)
Extracting fields (turn messy requests into a draft CRM/ticket record)
Routing (tags, priority, suggested next step)
Research packs (sources and a brief for a human)
Keep humans in charge of:
anything that sends externally without review
spending money
compliance-sensitive changes
high-emotion customer situations
If you do nothing else
Agents for throughput (read, summarise, draft, route). Humans for commitments (send, spend, approve).
Tool of the Week: Meta Segment Anything (Audio), a “Sound Picker” for Podcasters
Meta’s Segment Anything (Meta AI Demos) editor is best thought of as a sound picker. You tell it what sound you mean, and it tries to split your recording into two parts:
Target sound (the thing you asked for)
Everything else (the rest of the mix)
Under the hood it’s built on SAM Audio (sam-audio), a promptable audio separation model that can isolate sounds using text prompts, time-span prompts, and (when you have video) visual prompts.
The podcast problem it solves
Podcast editing is mostly three primitives:
Information: a long waveform with speech + noise + overlaps
Decisions: keep, reduce, or remove particular sounds
Actions: cut, clean up, rebalance levels, export clips
Most tools make you work backwards: you tweak knobs, then listen, then tweak again.
This tool flips it: you specify intent first (“dog barking”, “traffic noise”, “keyboard typing”, “laughing”), then it attempts the separation so you can act on it.
How it works (mechanism, not magic)
You provide an audio (or video) clip, then “prompt” the sound you want to isolate:
Text prompt: describe the sound in plain language
Span prompt: mark the time range where it occurs
Visual prompt (with video): point to the on-screen source tied to the sound
The output is effectively a separated “stem” you can turn down, mute, or enhance, plus the residual audio. (Think “give me the barking as its own track”.)
Where it helps podcasters right away
Rescue moments you’d otherwise bin: a clean quote underneath a sudden noise (dog, siren, mic bump). Meta explicitly calls out removing things like a dog barking from a podcast recording.
Faster clean-up passes: isolate a recurring noise (AC hum, typing) and reduce it more directly than manual clip-by-clip editing.
Better promo clips: pull out “foreground” moments (a laugh, a punchy line) and rebalance to make short clips feel clearer.
Trade-offs (the bill you pay)
Artefacts happen. Source separation can sound watery or phasey, especially when sounds overlap (two people talking over each other, music under speech).
Prompt sensitivity. Small wording changes can change results, so you still need a quick human listen before publishing.
Privacy and workflow. If you upload client recordings, check your data handling expectations before using any web demo.
A simple starter workflow
Find a problem section (10–30 seconds)
Prompt the unwanted sound (“dog barking”, “traffic”, “keyboard typing”)
If it misses, add a time span around the noise and try again
Bring the separated result into your DAW/editor and dip it rather than hard mute (often sounds more natural)
Bottom line: for podcasters, this is a practical new option for “surgical” clean-up and clip-making, as long as you treat it like an assistant that produces a first pass, not a final master.
Before You Go
Everything in this issue was aimed at one thing:
Helping you better understand and actually use agentic AI automation in your business.
If you only take one step from this issue:
👉 Pick one “throughput” task and standardise it this week.
Example: take 5 real support emails or quote requests, and have an agent summarise + extract key fields + draft a reply, then you do the final send.
Hit reply and tell me:
what kind of business you run, and
one repetitive process you’d love an agent to quietly handle (inbox triage, lead research, weekly reporting, document tidy-up, something else).
That’s how we keep future issues tightly focused on real small-business problems, not abstract AI talk.