Why this page exists
Most writing about AI agents for sales is speculative. This is not. Prospectio’s marketing, competitor research, site build and a good deal of its customer follow-up run through a desktop agent, and have for months. What follows is what has actually held up.
The useful mental model: it is not a smarter chatbot, it is a junior who never gets bored, has read everything you have written down, and will confidently do the wrong thing if you have not told them the rules.
The three jobs it has genuinely taken over
Verifying a list before it exists. This is the highest-value use and it is the least interesting to talk about. Give the agent forty profiles and ask it to make the case, in one sentence each, for why a specific problem sits in that person’s remit, with instructions to say “no case” when it cannot. It will drop a third of them.
That is not tidiness, it is the largest effect in our dataset. Broad job-title targeting replied at 13.0 percent; the same sequence to a list where subject-matter fit had been verified per person replied at 51.9 percent. Four times, from work that happens before anything is written. Nobody does this by hand at volume because it is tedious, and tedious is the whole point of an agent.
Inbox triage every morning. Read what came in, classify it, draft responses, send nothing. Replies answered inside 24 hours converted to a meeting at 21.4 percent; after 48 hours, 3.6 percent. A solo founder or a small team loses that window constantly, not through carelessness but because the inbox is not the thing on fire that morning. An agent that has already drafted the replies by the time you sit down changes the economics of that entirely.
Competitor teardowns. Pull a competitor’s sitemap, work out what page types they run, read a few in full, and report what they publish that you do not. We found a competitor mid-rollout of a 38-page programmatic build that was not in their sitemap index, visible only from three footer links. That is a couple of hours of careful work done in ten minutes, and it is the kind of research that never gets done otherwise.
What it is bad at
Writing that needs to be right rather than plausible. The failure is subtle: output that reads fine and drifts away from your standards across a batch. Our copy tests are the clearest illustration. A message with solution detail replied at 18.6 percent and the same message with it removed replied at 44.5 percent. Left to its own judgement a model writes the longer one, because it reads as more helpful and it is worse.
Anything requiring a relationship. It cannot do the interview, make the call, or ask the customer for the review.
Knowing when it is wrong. It will produce a confident number, a confident claim, and a confident recommendation with the same tone whether or not it has any basis. That is a supervision problem, not a capability one, but you have to actually supervise.
The two rules
Write your context down once. One file with your product in a sentence, your ICP, your buyer, your banned phrases and the proof you can actually defend. Everything improves from that single file, and without it you get output that is technically correct and contextually wrong. Ours includes a list of words that must never appear, which is more useful than any positive instruction about tone.
Keep a human on anything irreversible. Drafts, not sends. Not forever, but until you have read enough to know the constraints are holding. This applies to email, to outreach, to anything published, and to anything that touches a customer.
Where it fits with outreach specifically
A desktop agent on its own is a research and drafting tool. Connected to an outreach platform through MCP it becomes operational: create a campaign, import the people who commented on a specific post, attach a sequence, pull the stats, answer the inbox. That connection takes about five minutes and is documented here.
The boundary that matters: it drives the platform, and the platform talks to LinkedIn. Nothing should be reaching into LinkedIn directly, whatever it claims. LinkedIn prohibits unauthorised automated access and an agent with a browser is exactly what that clause describes.
The honest summary
The value is not that it writes well. It is that it does the work you were skipping.
The three biggest effects we ever measured were targeting at four times, video in the first touch at plus 40 percent, and the 24-hour reply window at six times. An agent can meaningfully help with two of those. It cannot record your video, and it cannot have the conversation.
Full method and all fifteen controlled tests are on the 2026 outbound benchmarks page.