Original · Unique Research · 2026-07-07
Editor's note: The first-person report and its judgments belong to the original Chinese author. This English rendition retains the Mizzen AI / Sun Keqiang interview narrative, the themed sections, and the complete 10-question Q&A. All named people, companies, and figures are preserved. Interviewee statements are source attributions, not independently verified findings.
AI Industry Observation
From 30 interviews a month to 300 in one night: AI starts "doing real work" in the enterprise.
From "tool" to "strategist," AI is rewriting the underlying logic of enterprise decision-making and execution.
From 30 interviews a month to 300 in one night.
This is the real data from a leading Chinese consumer-electronics brand last week. They used Mizzen AI (觅深科技)'s platform to complete in-depth interviews with over 200 respondents in a single night, for a final total above 300. What used to take a whole month, a team, and flights across the country, now gets done in one night.
When Sun Keqiang (孙克强) told me this case, he made it sound routine, but I was still struck; it felt like AI was redefining "efficiency" as something on another dimension — speed itself.
From 30 to 300: The Time Dimension Is Compressed
Let's first talk about what Mizzen AI actually does. Mizzen AI's positioning is clear: an enterprise-grade AI user-insight platform. In plain terms, it uses AI to collect, analyze, and consolidate the real voice of users, helping enterprises make more scientific decisions.
Why does this matter? Because when an enterprise makes any key decision — whether to greenlight a new product, whether users will pay for a prototype, which direction a brand campaign should take, whether communication feedback has stepped on a landmine, whether brand health is rising or falling, what brand personality actually looks like in users' minds — essentially all of it needs one thing: hearing enough real user voices.
But "hearing" is famously slow and expensive in the traditional research industry. You have to find the target demographic, design a questionnaire, recruit interviewees, assign researchers to analyze, and finally write a report. A project starts at two or three weeks and a budget of several hundred thousand. Sample size still can't go up; 100 in-depth interviews in a month? Basically impossible.
Mizzen AI's logic is: replace traditional human researchers with AI researchers, replace scheduled interviews with asynchronous voice/text interaction, and replace manual coding with intelligent analysis. From 30 people to 300 people. The time dimension compresses from "month" to "night."
But Sun Keqiang says this isn't even the point he values most.
"Speed is of course good, but what's truly important behind speed? It's that you build a super-efficient feedback system inside the enterprise."
He drew an analogy: many enterprises make decisions like driving blindfolded on a highway, only occasionally checking the rearview mirror — that mirror being the annual report or quarterly user research. What Mizzen wants to do is install a real-time radar, sensing road conditions every second.
When real user voice can enter the decision process quickly, cheaply, and at scale, the enterprise shifts from the long loop of "gut call → wait for launch → watch market feedback → maybe crash" to the agile mode of "there's a problem → ask users tonight → have a direction tomorrow."
The data assets this feedback system consolidates may, over the long run, be more valuable than any single research result. You ask users today about a new feature; three months later you look back at how the same people's attitudes shifted — this kind of continuous insight is almost impossible under the traditional model.
A Tool? No — a "Strategist"
In AI-app discussions, the word people love is "tool" — AI is an efficiency tool, a productivity tool. Sun Keqiang isn't buying it.
"A tool provides value passively. You give me a hammer; the hammer doesn't judge whether this wall should be smashed, at what angle, or what to do next."
He made a key distinction: an executor actively judges. An excellent executor, given a task, thinks for itself: what resources do I still need? What information is missing? What's the next step? Should I adjust direction? A tool doesn't; a tool waits to be used.
"So I'd rather see AI as the enterprise's decision supporter, even a think tank and strategist."
This view deserves expansion. What is a strategist? Not the person charging for you, but the person beside you as you make big judgments, helping sort information, run through possibilities, and point out blind spots.
Sun Keqiang gave a counterexample — Meta's metaverse. From around 2017, Facebook poured investment into VR/AR; Zuckerberg pushed all-in on the metaverse, even renaming the company Meta. It sank about $10 billion a year into Reality Labs, sustained for years. The result? Oculus Quest sales never met expectations; Horizon Worlds' user retention was dismal; Meta's stock briefly crashed in 2022. Tens of billions burned bought, in effect, one strategic misjudgment.
If Meta had had a "strategist" like Mizzen that could, before investing $10 billion, first ask 300 target users "do you really want a virtual office?", the ending might have been completely different.
"You ask if Meta lacked talent? Money? Technology? It lacked none. What it lacked was someone standing up at the very first strategic judgment and telling them: do users actually want this thing?"
This is what's truly expensive in an enterprise. The right or wrong of strategic judgment itself is far more expensive than the budget of any single execution action. AI writing copy, doing data analysis, generating PPT — these are all single-point efficiency gains, good but not enough. The big problems enterprises really face are: "should we do this direction at all?" "will users pay for this feature?" "should we change this brand positioning?"
These questions were answered in the past by executives' gut instinct, consultancies' PPTs, and small-sample qualitative research. Now, AI may become a new decision-support method — using large-scale real user voice to give strategic judgment its backing.
"In plain terms, a CEO saying 'I think users will like it' in the meeting room and saying 'we talked to 300 target users last night, 73% showed clear interest in this concept' are two completely different decision qualities." From "I think" to "73% say so" — that's the strategist's value.
One Pre-sales Supporting Six Sales: AI-Reworked Workflows Are Already Here
Here, Sun Keqiang shared two internal changes that impressed him. The first is pre-sales support.
"Before, one pre-sales engineer could only support two salespeople. Now they support six." How? The pre-sales role's traditional workflow is: the salesperson books a client, throws the client's needs to pre-sales, and pre-sales starts working — digging client background, understanding industry pain points, rummaging through internal materials to match product capabilities, pulling past success cases, then assembling these into a customized proposal draft. This whole routine takes a day or two. So the number of salespeople one pre-sales can support is limited; essentially they're tied up by foundational, repetitive work.
Now AI has eaten the middle chunk of the process. Client background? AI automatically crawls public information and generates a summary. Company capability matching? AI reads product manuals and the knowledge base, recommending the most relevant module in seconds. Past cases? AI pulls the three most similar from the project library and rewrites them for the new client. Proposal draft? AI generates a 70-plus version in minutes, and pre-sales only does final review and tuning.
"The pre-sales people weren't replaced, but their work changed. Before, 70% of their time was moving bricks and 30% thinking. Now it's reversed."
The second example surprised me even more: 360-degree reviews. Doing a 360 review once used to take HR enormous effort — sending questionnaires, chasing progress, collecting data, analyzing, writing feedback. Between leaders and team members you also had to arrange mutual reviews, coordinate time, organize meetings, guide discussion; the whole process was long and costly in management attention.
"Now AI runs the whole process. From sending questionnaires to collecting data, from generating analysis reports to giving development suggestions, AI completes it independently. HR only does a final pass to check for anomalies." What's reduced isn't just workload, but the internal drain on management energy.
These two cases point to the same trend: AI has upgraded from "helping you write faster" to "running the whole process for you." Sun Keqiang emphasized a concept — "clearly-bounded scenarios." "In scenarios where boundaries are clear, rules are explicit, and inputs/outputs are quantifiable, AI can today independently and completely do all the work of a certain function. Pre-sales drafts are one, reviews are one, and user research is one."
But he turned: "Once the boundary blurs, needs a lot of subjective judgment, or involves multi-stakeholder gaming, AI can't handle it. At least not yet."
What Enterprises Pay For Isn't Just Smart, It's Peace of Mind
Many founders in enterprise services might pause — haven't we been making AI smarter? Wrong. What enterprise clients pay for isn't smart, but peace of mind.
Think about it: having AI organize interview notes, generate weekly reports, sync system data — at worst it wastes half an hour if it messes up. But having AI auto-approve a $500,000 payment, sign a contract, give an employee a "failing" performance review — once these go wrong, who's accountable?
So Sun Keqiang believes three hard metrics judge whether a task can be executed directly by AI: reversibility, risk level, and degree of uncertainty. Organizing information, generating weekly reports, creating tasks, sending reminders — these low-risk, reversible, clear-rule things, AI should just do them. But large payments, contract signing, personnel decisions, external commitments — these high-risk, irreversible things AI itself can't be sure of must keep a human-confirmation gate.
Sounds simple, right? But most enterprise AI systems do either full automation (explode when something happens) or full human approval (AI reduced to a form-filling tool). Sun Keqiang thinks a truly usable system should be dynamic authorization. What does that mean? When AI starts a new task, permissions are locked down — every step needs human confirmation. But after it does 100 in a row without error, the system automatically loosens permissions. The more reliable it gets, the larger execution space it earns. Conversely, once it errs, permissions retract immediately. This isn't a technical problem; it's the design of a trust mechanism.
At this point you may ask: what do enterprise clients actually want? I've summed up four things, none dispensable. Clear permission boundaries — what AI can and can't touch, at a glance. Key actions confirmable — spending, signing, committing must pop up for a click. Full-chain auditable — what AI did, why, on what basis, can be checked afterward. Errors rollback-able — if it messes up, it can be undone, recovered, and attributed.
Mizzen has thought this through. Their approach is three layers of protection: controllable, traceable, isolatable. Researchers can watch in real time what interview AI is doing and step in to adjust direction; before key insights emerge, human review is kept — that's controllable. Every insight can be clicked back to see which user, which original interview, supports it — that's traceable. The whole system runs in an independent environment, single-tenant deployed, even on the client's own cloud — that's isolatable.
"Enterprises aren't unwilling to use AI, but they can't hand a key process to an uncontrollable, untraceable, non-isolatable black box."
In plain terms, enterprises buying AI work the same logic as hiring people. Would you stamp "full authority" on a new hire's forehead? No. You observe, assess, set boundaries, and gradually delegate as they prove themselves. AI is the same. No matter how smart, without controllability it's a ticking time bomb in the enterprise.
What Enterprises Want Isn't Chat, but Execution
After "controllability," a more essential question surfaces: most "enterprise AI" we see now is essentially still a chatbot. You ask, it answers, round after round, at most helping write a document or query data. But what the enterprise really needs is a system that understands business state, makes judgments based on evidence, calls tools to take action, and stays controllable throughout. That's completely different from a chatbot. It's an execution system.
Sun Keqiang breaks it into four core abilities, one by one. First is continuously understanding context. It must truly understand a client's history, a project's progress, a task's current state, the cause-and-effect of a series of decisions. Work in an enterprise is continuous action embedded in a complex business network, never isolated conversation. If AI needs you to re-explain the background every time, it's no different from a new colleague, retrained every time.
Second is reliable judgment. When AI advises the enterprise, it must be clear about what the basis is, where the evidence is, what's fact, what's inference, and what needs human confirmation. What enterprises fear most isn't AI saying "I don't know," but AI confidently stating a wrong conclusion without letting you see where it's wrong.
Third is actually executing actions. Generate tasks, update systems, trigger flows, notify relevant people, keep watching the results. It must actually do it, not generate text telling people "you should go do this." The execution chain in an enterprise is: see a problem → decide what to do → do it → confirm it's done → see the effect → adjust. AI must run this whole chain, not stop at "we suggest you..."
Fourth is controllable. This's already been covered — permissions, audit, human intervention, exception handling, none can be missing. Stripped down, these four abilities string together: know the situation → make a sound judgment → actually do the work → while doing it, someone's watching.
Where has Mizzen actually run through this? In the user-research scenario, they've already achieved understanding context from complex unstructured information and forming traceable conclusions. User research itself is complex enough — research design, recruiting respondents, dozens to hundreds of interviews, information organization, analysis and distillation, output reports. AI participates in every link of this chain, and every final insight can be clicked back to see which user's original words support it.
This process already solves the four hardest problems of enterprise AI: long-context understanding, multi-turn task execution, human-AI collaboration, and evidence chains with reviewable results. But Sun Keqiang also frankly told me two things are still being chewed on. One is longer-term, more dynamic memory and state management. Today's AI can remember the context within a project, but if a client has collaborated for three years, gone through five research rounds, with four product iterations in between, can AI automatically correlate this cross-timeline information? This needs far more than "long context" — dynamic state memory. The second is truly moving from "insight" to "closed-loop action." Now Mizzen's AI can help enterprises find problems and form judgments, but the next actions — changing products, adjusting strategy, driving implementation — still need people. How to make AI truly move from "finding problems" to "solving problems"? That's the next mountain. But the direction is clear.
A true enterprise-grade AI execution system is nothing like a chatbot. It's a system that continuously understands state, makes evidence-based judgments, calls tools to execute, and stays controllable.
Humans Define Goals, AI Bears Execution
After the technical architecture, let's talk about something more essential — in the future enterprise, how exactly do people and AI divide the work? Sun Keqiang drew a timeline. The first stage is Copilot. AI sits beside you, helping you write, look up, analyze. Humans are the main driver, AI the co-pilot. It writes what you tell it, answers what you ask. This is the form we're most familiar with now. The second stage is Agent. AI can independently complete concrete tasks. You say "organize the key insights from all last week's client interviews," and it pulls recordings, transcribes, analyzes, outputs conclusions itself, without you teaching step by step. But essentially it's still executing a clearly defined task.
The third stage, I call it Execution System — execution infrastructure. AI goes from a tool "you go use" to a layer of infrastructure embedded in enterprise operations. It continuously understands business state, collaborates across systems, autonomously executes actions, and tracks results in a closed loop. What's the biggest change at this stage? To introduce AI as infrastructure, the enterprise must redesign "which work is done by humans and which by AI."
This sentence carries weight. It means AI won't just exist as a plugin to existing processes; it will reconstruct the processes themselves. An example. A product manager's current workflow might be: do user research → organize requirements → write PRD → review → follow up development → acceptance → launch observation. In the future this process may split in two: AI handles execution of research (recruiting, interviewing, organizing, initial analysis), continuously tracks data, auto-generates weekly reports, syncs project state, and flags risks. The product manager? Only three things: define the goal (what problem are we actually solving), make key judgments (based on AI's analysis, decide whether and how to do it), and bear final responsibility (after launch, whether it's good — you're the one accountable).
In the past many people's value lay in remembering processes, chasing progress, moving data, and doing repetitive execution. Honestly, this work is hard and patient. But in the future, these are exactly what AI is best at. Where does human value migrate? I'd say eight words: define the goal, bear responsibility.
AI can help you get things done, even do them well. But it doesn't know whether the thing "is worth doing" — that's goal definition. It also can't be responsible for the result — when something goes wrong, the boss takes the blame, not the AI.
In Sun Keqiang's view, the scarcest ability in future enterprises is no longer "getting things done," but knowing what's worth doing and why, and being willing to take responsibility for results. This shift will be painful. Many people have proven themselves at the execution level — I work more overtime, I'm more efficient, my processes go smoother. But the stronger AI gets, the less these are worth. The more people can't only be execution nodes in a process, the more they must become goal definers, boundary setters, and final responsible parties.
Ultimately, what truly changes enterprises isn't a tool upgrade. Some Monday morning three years from now, a product manager walks into the office and AI has already organized all last week's user feedback, flagged three key questions needing his own judgment. He sits down not answering emails or chasing progress, but thinking: what's the most important thing this month? Turning people driving the process into AI driving execution and humans defining goals and bearing responsibility. This shift won't happen overnight. But once it starts, there's no going back.
More Conversation Details
Q1: In plain "human" language, what exactly does Mizzen AI do?
Sun Keqiang: In one sentence, Mizzen AI wants to be an enterprise-grade AI user-insight platform, using real user voice to support scientific enterprise decisions. Enterprises doing product innovation, prototype validation, brand marketing, brand-health tracking, even brand-personality profiling, essentially all need to hear the voice of many real users. But in the past this was very slow and heavy. Often an enterprise needs a month just to interview 30 people. Now with AI, this can happen overnight. Speed matters, but behind speed what matters more is that we want to build a super-efficient feedback system, consolidate real user voice, and turn it into decision assets the enterprise can reuse.
Q2: How do you view the judgment that "AI is no longer just a tool, but is starting to become an executor"?
Sun Keqiang: I partly agree. The biggest difference between a tool and an executor is that a tool provides value passively, while an executor actively judges: what resources do I still need, what information do I still need, what should the next step be. But I don't fully agree with defining AI only as an "executor." In my view, AI's value in the enterprise isn't only helping you land actions; it can also help you lower decision risk and reduce the probability of error. So I'd rather see it as the enterprise's decision supporter, even a think tank and strategist in a sense. It can help you get the information strategic analysis needs, then land that information into execution recommendations, helping the enterprise make higher-quality decisions.
Q3: After AI enters the enterprise, has value upgraded from "improving efficiency" to "replacing part of process execution"?
Sun Keqiang: I think it has upgraded to "replacing part of process execution." A very concrete change is pre-sales. Before, I thought one pre-sales could only support two salespeople, but now one pre-sales can support about six. Because a lot of foundational pre-sales work — understanding client background, organizing company capabilities, pulling past cases, forming proposal drafts — can be assisted or even done directly by AI. This isn't just making people faster; in some clearly-bounded scenarios, AI can already independently and completely do part of some function's work. Of course not all processes will be replaced, but as a trend, AI's value is no longer just an efficiency tool, but is becoming a new execution unit.
Q4: Why are single-point AI capabilities — writing copy, doing analysis, generating content — no longer enough?
Sun Keqiang: Because for enterprises, the most expensive cost isn't the spend at a single point, but the cost of doing things wrong. Writing copy, doing analysis, generating content can improve single-point efficiency, but what enterprises really face isn't "can this copy be written faster," but "should this direction be done at all." A good decision can give an enterprise long growth; a wrong decision can drag it down. So what enterprises really need is higher, stronger intelligence: able to end-to-end help enterprises get information, understand information, form judgments, and land judgments further into execution.
Q5: Is Mizzen AI more like workflow automation, or an AI decision/execution system?
Sun Keqiang: Mizzen AI clearly wants to build an AI decision/execution system. Our goal isn't simply to do workflow automation, but to build a decision engine. If it's just workflow automation, it's more automating the traditional user-research chain — auto-generating interview outlines, auto-organizing interviews, auto-outputting reports. That's valuable, of course, but it's essentially still responsible for one process step. But the AI decision/execution system we want doesn't just make one step faster; it must have complete information-acquisition ability, consolidate and integrate this information on the platform, and ultimately become a platform that supports humans in various decisions and underwrites downstream business. The biggest difference between the two is who it's responsible to. Workflow is responsible for a process step; a decision engine is directly responsible to the business department, to business output, and in a sense even to strategic decisions.
Q6: For AI to truly complete an enterprise workflow, is the hardest part understanding the task or executing across systems?
Sun Keqiang: Honestly, both directions now face big challenges. First is understanding the task itself. A task in an enterprise can't be clearly defined by one instruction; it needs understanding of task background, phased expectations, quality requirements, task style, and the real state of results. Many tasks can be done by people because teams share an unspoken rapport and lots of unwritten context. Second is cross-system execution. Internal enterprise work isn't a single step; it usually involves coordination among dozens of systems, including cross-department, cross-system, cross-platform switching. If AI truly completes a workflow, it must be compatible with different platforms, permissions, data structures, and business states. So it's not an either/or. The truly hard part is that AI must simultaneously understand the enterprise's real context and execute actions in real systems.
Q7: What core abilities should a truly usable enterprise-grade AI execution system have?
Sun Keqiang: I think at least four core abilities. First, continuously understanding context. It can't only answer questions in a single conversation; it must understand clients, projects, tasks, historical decisions, and current state in the enterprise, knowing where things stand. Second, reliable judgment. In enterprise scenarios, AI can't just generate a plausible-sounding answer; it needs evidence, sources, traceability, knowing what's fact, what's inference, and what still needs human confirmation. Third, actually executing actions. It doesn't just tell you "suggest following up this client," but can further generate tasks, update systems, trigger flows, notify relevant people, and continuously watch execution results. Fourth, controllable. Including permissions, audit, human intervention, exception handling. Enterprises won't hand core processes to a fully black-box Agent, so human-AI collaboration and governance are critical.
Q8: How do you define the difference between an AI Assistant and an AI Execution System?
Sun Keqiang: The core difference is: an AI Assistant helps you do things; an AI Execution System helps you get things done. The Assistant is usually human-initiated, AI-responsive, focused on answering questions, generating content, giving advice. It solves "is this answer good" and "is this advice helpful." An Execution System faces a goal. It needs to continuously understand context, call different tools and systems, autonomously judge the next action, bring in people when necessary, and ultimately form a closed loop of results. So simply, an Assistant is responsible for answer quality; an Execution System is responsible for task results. That's the most fundamental difference.
Q9: Which tasks suit direct AI execution, and which must keep human confirmation?
Sun Keqiang: I think the key isn't a one-size-fits-all by task type, but three dimensions: reversibility, risk, and uncertainty. Low-risk, reversible, clear-rule tasks can be directly executed by AI. For example, information organization, system sync, generating weekly reports, creating tasks, sending internal reminders — even if wrong, these are easy to find and fix. But anything high-risk, irreversible, or where AI's own confidence is low should keep human confirmation. For example, large payments, contract signing, personnel decisions, external commitments, or actions that significantly affect client relationships. My principle is: AI can autonomously execute low-risk actions; the closer to irreversible decisions and major responsibility, the more Human-in-the-loop is needed. Long-term, the best system isn't fixed approval but dynamic authorization. The more reliably AI proves itself, the larger execution permission it gets.
Q10: From real cases, what change can AI actually bring in user research?
Sun Keqiang: We have a large client, a very leading Chinese consumer-electronics brand. They have a huge user base, but had never understood their users well. Before, doing one study could take a whole month and only interview 30 people. In our first POC, in five or six hours we helped the client interview 30-plus respondents. After the brand saw the results, it began to believe in the platform's AI-moderator ability. Later we helped them interview 200-plus respondents in one night, and a few days later the final interview count exceeded 300. The key here isn't the number itself, but that in-depth research at this scale was almost infeasible before. AI lets enterprises get a large amount of real user voice in a very short time, not staying at the questionnaire level, but through the AI moderator's continuous follow-up consolidating deeper information. Finally, these results help enterprises better understand user distribution, structure, and characteristics, and are widely collaborated on and circulated across internal departments to support more business decisions.