In March, something changed: our users stopped searching.
That is not a red flag. For us, it's music to our ears. It means our users are adapting with us.
Figure 1: Over the last 23 months, our platform has shifted from finding information for users to understanding, analyzing, and doing the work for them.
Over the last 23 months, we’ve witnessed this shift building. As AI has continued to advance and AlphaSense’s capabilities have significantly expanded with it, our users moved with us too. They went from asking AlphaSense to find things to asking AlphaSense to do things.
Years ago, the most common searches on the platform were fragments like: “revenue guidance,” “iPhone sales,” and “Samsung market share.” Search phrases averaged six words and returned a list of the most relevant documents. Today, the average Generative Search query is a 100-word prompt and looks more like a note for your analyst: “Track the capital investments breakdown for NVDA, AMD, CoreWeave, and other hyperscalers and neoclouds; when there are impactful changes, update my coverage deck.”
The change is clear: AlphaSense is truly becoming an extension of the team. To understand what that meant in practice, we dug deeper.
The Shift, Dissected
Figure 2: Five query types, showing how user needs shifted over time. Navigation collapsed when Generative Search took over, while open-synthesis and agentic-action have increased steadily as model and harness quality have improved.
To dig deeper, we used anonymized classification metadata to understand usage signals. Each query was sorted into two buckets, finding (e.g., "get me this document") and delegating (e.g., "do this thing for me”). Within those buckets are five query types (navigational, pinpoint, multi-hop, open-synthesis, and agentic-action), and each one reflects more user trust than the last.
The first shift happened between October 2024 and February 2025. Navigational queries dropped from 42% to 18%. That volume moved to pinpoint queries like, “What was Boeing's margin in Q3.” So what happened? AlphaSense's Generative Search was finding the right documents reliably, so users skipped ahead and asked for the answer like they would ask a human analyst. Total "finding" queries barely changed, but underneath was the beginning of an even bigger behavior shift.
Next, in mid-2025, the shift changed character. Pinpoint queries began to recede, and people started handing over the work itself with delegating queries. Two kinds of handover appeared. The first, open synthesis, is work the user waits for. They type a real task like: "explain the competitive landscape,” or "summarize what changed in guidance and why," and then read the output, clicking into source documents for the deeper context on the narrative. Through 2025, open synthesis accounted for the majority of delegating query growth.
The second, agentic action, is work the user delegates to the system. In March 2026, this type of query started to grow, and by September it was 19%. These queries describe the job once and go, and nobody is watching when it runs. Our CFO, Samantha Greenberg, described the shape of it on stage at Citi’s Global TMT Conference last month: a semiconductor banker goes to sleep in Silicon Valley, ASML makes an acquisition overnight in Europe. By morning, there is a deal memo with broker research, comparables updated for the pro-forma multiples, before a single query was issued. This is the 19%: scheduled agents that no one was awake for, which barely existed at the start of the window (under 1% of queries). These runs read several times as much source material as a watched one and take far longer.
With agentic workflows rising, new risks are introduced, because every mistake can compound in a bigger way than ever. Our CEO, Jack Kokko, put it plainly: “Allow a step in an agentic process to have 95% accuracy and a 10-step process becomes a coin toss.” This is why AlphaSense is so strict about sourcing.
Figure 3: Paired dot plot of 14 capabilities, first month versus last. Fact lookup is the single capability whose share falls, from 91% to 43%. All others rise.
There’s another way to study this trend. If you sort the same queries by what skill the answer needs, the same movement appears from a second direction. Fact lookup is the only one of 14 skills whose share fell between October 2024 and September 2026, from 91% to 43%. A search box used to mean one job: find the thing; now it means infinite.
Deliverables, Not Documents
Deliverable growth is significant. Two years ago, roughly one query in a thousand asked us to build something. It is now one in seven. Now our users have generated over a million slides and 100,000 slide decks. As of September, delegation represents 56% of queries and deliverables represent about 14%. The gap is not missing data, because most delegated work asks for an answer, not an object. "Explain what changed in the guidance and why" is an example of a job that a user could give AlphaSense and has named nothing to be built.
Figure 4: Requests for any kind of artifact have risen to one in seven queries over the last 22 months. Requests for a table or spreadsheet rise to 9.6%, for a deck to 4.5%, and for a chart to 1.8%.
One in seven still understates it, because it only counts the people who asked for it in their question. By September, 2.4% of queries asked for slides in text. A further 2.1% of users asked an ordinary research question and then pressed the button that turns the answer into a deck. The two groups barely overlap; only 0.07% did both. The button did not replace the asking. It found a second population who wanted slides and had never thought to ask a search box for them. Counting both, one in 22 conversations now ends in slides. Two years ago it was one in 7,000.
Figure 5: Net growth in requests for deliverables by type over the last 23 months.
The handful of users who asked in 2024 weren't asking us to build anything. They were building the slide for themselves and asking us for parts, for example: “Help me generate a table I can put in a slide,” or “I am making a slide on this industry; what are the key insights to include?” The deliverable was the analyst’s job. Now the system builds the whole slide deck end to end; and AlphaSense’s PowerPoint plug-in lets the users iterate on slides in natural language, a massive acceleration in a key workflow for our user base, who are constantly presenting their research-backed ideas to colleagues and leadership.
The World Going Forward: Delegating and Augmenting
The trend from October 2024 to September 2026 maps to our product releases, beginning with Generative Search, Deep Research in June 2025, and Scheduled Agents in February 2026. Then came the crossover in March 2026, when standing agentic work went from 7% to 19% over the following two quarters.
Figure 6: Change in model capabilities required to answer user questions in SuperAnalyst over Generative Search. As capabilities increase, reasoning is center stage as pure artifact generation and fact lookup are solved.
What comes next? Anyone reading a chart being asked to think about the future might expect that the delegation trend keeps growing until all work is delegated. We do not think that is what happens, and our new product is the reason why.
SuperAnalyst is an agent-first surface with August being its alpha month. If you tag its current usage with the same metadata classifiers, delegation goes from 56% to 58%. This isn’t a surprise but a design. With SuperAnalyst, the handover questions are largely solved, and what comes next moves downstream. The requests it sees do not delegate more, but they expect substantially more work per request. The median prompt is 50% longer, and 39% of prompts require three or more distinct capabilities, compared to 29% for Generative Search. Users are asking harder questions than ever before, and they are building alongside their SuperAnalyst.
Figure 7: Words per query at three quantiles, log scale. The typical query grew about fourfold, from 5 to 18 words. The longest tenth grew sixteen-fold, from 14 words to 229. The mean, 6 to 110, sits near the top because the distribution is heavily right-skewed.
From October 2024 to September 2026, the number of queries a person types did not grow exponentially. However, each of those queries now carries seventeen times the words and pulls in thirteen times the documents. The platform does far more work on far fewer keystrokes, and all of the growth in the raw request count comes from scheduled runs that nobody types at all.
On stage at Citi, when discussing what share of searches on the platform are now issued by a machine rather than a person, our CEO put it at close to 99%. On a run-rate basis, agents are contributing 2 billion actions compared to 20 million by human users. People are still asking a lot, and queries are growing rapidly, but SuperAnalyst (along with its sub-agents) runs 100 queries in the time a user would complete one. Research now goes much broader and deeper, and still gets completed a lot faster. This has opened up areas of research that simply weren’t possible before.
We think the first two years of this AI boom have been about moving some work from the person to the machine. The next two are about how people choose to augment and delegate work — and the creativity they deploy when they can now ask the machine to do much more complex work.
