A study published on June 8, 2026 reports that AI agents cut task completion time by 87% and cost by 94%. The study was co-authored by Perplexity using its own data, compared only Perplexity’s own two products, and lists three Perplexity employees among its four authors. For B2B marketers weighing AI agent tools, the finding is a number shaped by a company with a direct stake in the result.
What the Perplexity-authored study claims
The paper, “How AI Agents Reshape Knowledge Work,” reports large gains for autonomous agents. It says an agent ran 26 minutes of work in a session against 33 seconds for a basic assistant, and that on matched tasks the agent cut average completion time from 269 minutes to 36. That is the source of the 87% time cut and a 94% cost cut, the cost figure based on US wage data. The study also reports a 55% drop in user dissatisfaction with the agent, and says agent users worked across more knowledge areas than assistant users. For marketers, one detail stands out: marketing and sales was the third-busiest use of the agent in the sample, at 7.6%. The method has a real strength worth noting. The authors matched 10,000 near-identical opening questions that the same people sent to both products, drawn from 8,357 users, which controls for the fact that people send easy questions to a chatbot and hard jobs to an agent.
Why the study’s source matters
The data are Perplexity’s own. The only two products compared are Perplexity’s basic answer engine and its paid agent, so the headline result is one Perplexity product beating another, not a test against rival tools. Three of the four authors work at Perplexity, alongside a lead author from Harvard. Two of the claims also rest on softer ground than they first appear. The “quality” gain is measured by whether a user complained, re-asked, or corrected the result in the next step, not by whether the work was actually accurate. The cost savings are modeled, not observed, using assumptions about how long each task would take by hand. The authors themselves note the 90-day window skews toward power users and paying subscribers.
Who reported it and what independent research says
The study was analyzed by PPC Land on June 15, 2026 and posted as an arXiv preprint, which means it has not been peer-reviewed. Research without that financial interest points to more caution. A November 2025 study from Carnegie Mellon and Stanford found agents finished tasks much faster but with real quality gaps, and reported that AI automation can slow human work by 17.7% once you count the time spent checking the output. Gartner has predicted that more than 40% of agentic AI projects will be canceled by the end of 2027 over cost and unclear value. The speed gains show up across both vendor and independent work; the claims about quality and cost are where the independent research is far less sure.
How this fits B2B marketing’s AI moment
Vendor efficiency numbers rarely stay in research papers. They move into pitch decks, board memos, and budget cases. The same year has seen AI move deeper into how B2B brands get found and sell, from AI engines becoming a top B2B discovery route, to regulators stepping into AI search with the UK order letting publishers opt out, to AI platforms selling ads inside the chat window. In that rush, a clean claim that agents cut cost by 94% is exactly the kind of figure that shapes spending, and exactly the kind that deserves a closer look at where it came from.
Talking Shift: before you trust a number, ask who wrote it
This study says AI agents cut work time by 87% and cost by 94%. Those numbers will show up in sales decks and budget meetings fast. So here is a simple habit before you let any vendor stat into a decision. Ask three things. Who paid for the study? What did they compare it against? And did they measure real quality, or just guess at it? This one fails all three. Perplexity supplied the data, only compared its own two products, and judged quality by whether people complained, not by whether the work was correct. Start Some Shift’s take: when the company that profits from the answer also runs the study, treat it like advertising. The speed gains may be real, so test them on your own work before you put them in a budget.
What B2B marketers should do about vendor AI claims
Use this story to set a standard for any AI number that reaches your team:
- Ask who funded and authored any AI study before you quote it in a plan or a pitch.
- Check what the tool was actually compared against, and be wary when a company only measures itself.
- Separate speed claims from quality claims, since the research agrees on speed and splits on quality.
- Run a small, timed test of any agent on your own real tasks before you budget for it.
- Count the time your team spends checking AI output, because that cost is often left out of vendor math.
- Keep a human review step on anything an agent produces for clients, and write that step into your process.
What to watch next for AI agent research
Watch whether independent, peer-reviewed studies confirm or trim these efficiency claims. Watch whether more vendors publish their own favorable research as a sales tool. And watch whether buyers start asking, by default, who paid for the numbers in front of them.
