A&A INSIGHTS
Does an AI research service have to answer instantly? Designing a deliverable people will wait for
Build a chat window and instant answers become the promise, which leaves out the depth you wanted to sell. From Anthropic’s Genspark case study and its guidance on trading latency for task performance: how to separate instant consultation from scheduled research, and which signals to set.
日本語で読む
THE STARTING POINT
For solo and small operators who sell research as a product. The argument: because the value of research AI is not only the speed of an instant answer, an asynchronous offer is worth testing. Anthropic’s Genspark case study quotes co-founder and CTO Kay Zhu saying that positioning the product as search meant returning to the user in one second or at most ten, and that ten seconds was too short for the system to do meaningful work. Anthropic’s agent design guide states that agentic systems often trade latency and cost for better task performance. It sets out a two-tier offer and the receipt, progress and completion signals. Source facts, A&A interpretation and examples stay distinct.
Assume instant answers and the research you wanted to sell never fits
A&A perspective
A solo operator who sells research and decides to build an AI front door usually ends up at a chat screen. The other person types a question and an answer comes back on the spot. It is easy to build, easy to demo, and everyone else does it. Then it runs, and the answers are shallow: the questions that would reward real digging are exactly the ones where the conversation trails off. A&A’s reading is that the cause is not capability but the speed the screen promised. A chat window promises an instant reply the moment it opens, and the research you wanted to sell does not fit inside that promise.
From the sources
Anthropic’s published Genspark case study states this constraint plainly. It quotes Kay Zhu, co-founder and CTO at Genspark, saying that if they positioned the product as search, it needed to return to the user in one second, or at most ten seconds, and that ten seconds was too short for the system to do meaningful work. According to the page, Genspark launched out of stealth in mid-2024 as an AI search product, then steadily expanded into parallel search and asynchronous deep research that would email users a finished analysis after running for several minutes. The page says the team kept expanding the runtime because each expansion let them tackle harder problems for users. The page is published by Anthropic on its own site; it is vendor testimony, not an independent third-party audit.
Change how the work is received and the questions you can take change
From the sources
The same page gives an example from the user’s side. One paying user spent a week using a traditional search engine to compare credit cards against a long list of personal criteria, then handed the same question to Genspark’s asynchronous agent and got the same conclusion in 15 minutes. The page also explains that the architecture underneath at the time was a directed graph of predefined workflow nodes, and that this design was hitting a ceiling. Zhu is quoted saying it was too rigid, and that it often broke on edge cases.
A&A perspective
Reading that 15 minutes as an example of "AI is fast" misses the point. In A&A’s reading, what is doing the work here is the comparison. Fifteen minutes looks valuable because the thing being compared against is not a search engine’s response time but a week of the user’s own labour. What going asynchronous bought was not speed but a change of kind: work that previously had to be done yourself became something you could hand to someone else. When receipt moves from a screen to an inbox, the thing your buyer compares you against moves from other chat windows to their own hours. For a one-person research business, that move should matter more than shaving the response time. This is A&A’s interpretation, not a statement in the source.
| What you decide | Instant consultation | Research you take time over |
|---|---|---|
| Where the answer comes from | Material and experience you already hold | Primary sources you open after the request |
| What you promise on receipt | Answer now, or say you cannot | Start date, hand-over date, and what is out of scope |
| Contact along the way | None | One check-in when an assumption breaks |
| Signal of completion | The conversation ends | The deliverable is handed over and acknowledged |
| Unit of pricing | By time or by month | Per engagement |
| Questions it suits badly | Ones that need primary sources to answer | Ones where the decision is needed today |
Making people wait is not payment; it is an exchange
From the sources
Anthropic’s agent design guide, published on December 19, 2024, states the exchange directly. It recommends finding the simplest solution possible when building with LLMs and only increasing complexity when needed, adding that this might mean not building agentic systems at all. It then states that agentic systems often trade latency and cost for better task performance, and that you should consider when this tradeoff makes sense. It also says that for many applications, optimising single LLM calls with retrieval and in-context examples is usually enough. A note at the top of the article says much of the tooling landscape it describes has changed since December 2024.
A&A perspective
The framing as an exchange carries straight into offer design. Waiting is a cost you charge to the other side, and in return you owe them a depth they would not reach themselves. So "make it asynchronous and you can charge more" does not hold. What holds is the order: you open primary sources, reconcile several of them, and split the conclusion by condition on their behalf, and you take the time that costs. Conversely, attaching a multi-day turnaround to a question you could answer from what you already know is simply a delay. A&A’s test is that if you cannot say in one line what you are exchanging the wait for, that offer does not yet have a reason to be asynchronous.
Sell instant consultation and scheduled research as two things
A&A perspective
The practical conclusion is not to cram both into one door. Put what you answer instantly and what you take time over side by side as separate offers. Split them not by difficulty but by where the answer comes from: if material and experience you already hold will answer it, it belongs on the instant side; if you need to open primary sources after receiving the request, it belongs on the scheduled side. That test can be applied the moment you read an enquiry. Splitting by difficulty instead makes the call ambiguous between hard-for-them and hard-for-you, and you end up forcing an answer on the spot. With both on the menu, the instant door turns into a place where you can say "I cannot answer that" and immediately point at the other offer.
Hypothetical example
As a hypothetical, consider a solo operator selling competitor research in one industry. Two doors. The first is a 30-minute consultation answered on the spot, where the answers come from industry material already read and from past engagements. When a question arrives that cannot be answered, they say so plainly and point to the second offer. The second is per-engagement research: after the request, they open public sources and hand over in five working days. What is handed over is the conclusion, the list of material it rests on, and a list of what this engagement could not establish. That third list lets the buyer set the scope of the next request. This is a hypothetical setup for explanation, not a real operator or a delivered result.
Decide the receipt, progress and completion signals in advance
From the sources
The same guide also describes where agents meet people. It says agents begin their work with either a command from, or interactive discussion with, the human user, and that once the task is clear, agents plan and operate independently, potentially returning to the human for further information or judgement. During execution, it says, it is crucial for agents to gain ground truth from the environment at each step, such as tool call results or code execution, to assess progress. It adds that agents can then pause for human feedback at checkpoints or when encountering blockers, and that while the task often terminates upon completion, it is also common to include stopping conditions such as a maximum number of iterations to maintain control. The article’s appendix lists the conditions under which agents add the most value: tasks that require both conversation and action, have clear success criteria, enable feedback loops, and integrate meaningful human oversight.
A&A perspective
That is machine design, but A&A reads it as transferring directly to a research service sold by a person. Three signals move across. First, receipt: when the request arrives, return the start date, the hand-over date, and what this engagement will not cover. Second, mid-course: when an assumption breaks, meaning the material you expected does not exist or the question reads two ways, send exactly one check-in. Turn that into "report once everything is finished" and an off-target deliverable arrives five days later. Third, completion: treat the engagement as complete when the buyer acknowledges receipt, not when you send the file. And set, for your own sake, the condition under which you stop if no conclusion emerges after days of digging. The absence of a stopping condition produces the same result in a person as in a machine.
What does not carry over from Genspark
From the sources
The same case study also shows how different the premises are. It states that Genspark surpassed $250 million in annual recurring revenue after pivoting to the Super Agent in early 2025, that Claude orchestrates more than 150 specialised tools inside a single agent, and that roughly 50 engineers produce all of the company’s code through AI tools. These are the company’s own account as published by Anthropic, not independently verified figures. The page also states that users invented use cases nobody at Genspark had planned for, giving the example of a New York real estate analyst who now produces investor pitch decks in hours, work that used to take weeks of multi-person research.
A&A perspective
What a solo operator can carry over is neither that scale nor that revenue. It is only the first two observations: that the response time a screen promises decides how deep a question you can take, and that changing how work is received changes what the buyer compares you against. The 150 tools and the 50 engineers are independent of both. There is also a reading to avoid: the causal claim that going asynchronous is what produced the growth. The page does not say so, and it carries no material that would let anyone check what happened between the pivot and the revenue. The statement that users invented unplanned use cases points the other way, suggesting the builders did not know the shape of the demand in advance. When you launch an asynchronous offer of your own, whether the demand exists can only be established after launch, from how people respond.
Where this shape does not fit, and the next step
A&A perspective
Three cases where it does not fit. First, when the buyer’s decision is needed today: a five-day turnaround is meaningless for information needed in tomorrow’s meeting. Second, when the subject is shallow: making a question you could answer from existing knowledge asynchronous just leaves the buyer waiting. Third, when willingness to pay has not been tested: an offer that takes time only works for questions the buyer believes would cost them a week to do themselves, and whether they believe that has to be asked, not assumed. This article did not measure demand for asynchronous offers. Of the two sources, one is a company’s own account on a case study page and the other is general design guidance; neither is evidence that your customers will wait.
A&A perspective
The next step to recommend is not rebuilding the offer but sorting recent enquiries into two piles: those you answered from material already at hand, and those you answered vaguely because answering properly would have required research. Count the second pile, and look for the same kind of question recurring in it. A recurrence is the candidate for your first asynchronous offer. No recurrence means there is nothing to split yet. If you would rather review the whole picture of the work you sell, A&A’s "AI-native GTM: a practical guide for solo founders and small teams" explains how customer acquisition connects through to retention. If you are stuck on which offer to build, or on pricing and turnaround, talk to us on an advisory basis.
The first decision in a business that sells research is not which model or which screen, but what you are asking people to wait for. Separate the consultation you answer instantly from the research you take time over, and sort by where the answer comes from. On the scheduled side, write the three signals — receipt, mid-course check, completion — and the condition under which you stop if no conclusion emerges. Making people wait is an exchange, not a value in itself, so if you cannot say in one line what the wait is exchanged for, that offer does not yet have a reason to be asynchronous.
Sources & editorial note
Primary pages read for this article. Publication dates below belong to the sources; access dates record our research.
- Genspark's Super Agent orchestrates 150+ tools with Claude
Anthropic · n.d.
Accessed 2026-09-24 - Building effective agents
Anthropic · 2024-12-19
Accessed 2026-09-24
AI-assisted editorial production
A&A uses AI for research, writing, translation and editorial checks. Source facts, our analysis and hypothetical examples are labeled separately.
Editorial check: 2026-09-24