Guest: Mengqi, founder of invoko.ai, as introduced at release | Host: Qu Kai
Published: 2026-05-23 | Duration: 68:42 | Prepared: 2026-10-18
Official episode
When AI can quickly generate similar functions, is there room for software startups? Mengqi’s answer reverses: coding progress first caused anxiety, then work on specific experiences restored her confidence. Her accounts of sourcing agents, marketing tools, and desktop software are valuable for tracing how apparently coherent decisions were overturned by actual use.
This article preserves a personal account from May 2026. Criticism of peers, markets, and business models was not independently investigated. Anonymous observations do not establish wrongdoing by a company, and one pivot cannot define all enterprise entrepreneurship.
1. Choosing a category before demand can confuse shared technology with shared customers
Listen: 08:45–16:39. The team first avoided a crowded general-agent market by choosing vertical agents, then sourcing: finding creators, candidates, experts, or customers. All involve finding people and seem to share search capabilities. But that starting point answered how to avoid competitors before identifying a payer’s most important task.
After launch, two feedback streams conflicted. Founders wanted results without operating the product. Professional marketers wanted auditable processes, custom metrics, and confirmation at intermediate steps. Following professional feedback produced something resembling traditional SaaS without incumbents’ accumulated data. Professional feedback was not inherently wrong; the team had not settled the target role before trying to satisfy incompatible preferences.
Mengqi compares flour and eggs: they can make flatbread or pasta without one customer ordering both. Shared ingredients do not establish cross-selling. The team deepened creator marketing instead of covering every sourcing scenario. An editorial implication is to expand through adjacent tasks of the same user, rather than a feature tree based only on technical capabilities.
2. Completing delivery and postponing rejection of an old direction can coexist
Listen: 12:28–20:53. Mengqi estimates the original tool covered roughly 30% of customer work while omitting time-consuming communication. Search quality depended heavily on external API data. These are estimates about her business, not industry measurements. Completing the easiest AI segment may miss the largest value.
Qu asks why the team could not perfect one task and suggests expansion might postpone admitting the original direction was weak. Mengqi says it was not wholly ineffective; its value was insufficient, making communication and cooperation worth adding. Preserving the disagreement is more useful than a generic instruction to focus. Did expansion fill a delivery gap or merely enlarge the market story?
She also reports companies earning service revenue while customers do not use the demonstration product. This does not establish deception by unnamed companies. The usable question is how software, managed operations, consulting, and human fallback are priced separately. Services are not inherently bad, but teams need an honest account of what customers buy, who bears costs, and whether it persists.
Enterprise sales cycles and organization did not suit her preferences and team. That is a personal choice. High-paying heavy users and subscribers with lower consumption also create different cost structures. Her “whale” and “gym” metaphors do not establish a superior model or justify making unused subscriptions a product objective.
3. One button removing friction can matter more than a new concept
Listen: 30:21–38:31. The consumer pivot was gradual. The team first built Clico, a browser extension for locating information across tabs, citing context, and writing near the cursor. Instead of screenshots, switching to a chat tool, explaining, and copying back, a preset button shortened the path from intention to action.
Mengqi initially considered it too simple. User feedback changed her view. Value did not require an idea nobody had imagined; a task became easier. Additional functions still needed restraint so buttons and information organization did not become new burdens. A valuable feature does not automatically create a coherent experience.
The desktop expansion reportedly took about three months and forty-nine versions, with bugs remaining. This describes the team’s compatibility, stability, and edge-case work, not a universal software schedule or independently proven superiority to competitors. Opening, running, and smoothly solving a problem are different acceptance levels.
Permissions also shape experience. The team used a comic to explain data storage for users unwilling to read long agreements. Clear explanation does not prove security: permission scopes, actual storage, access, and revocation need to match. This last limitation is editorial; no implementation audit was conducted.
4. Research should change decisions; internal use has sampling limits too
Listen: 39:07–42:46; 44:59–47:48. Mengqi describes decomposing decisions into questions and seeking Reddit feedback on pain points, previous solutions, priorities, and payment. She reports forty to fifty posts per question, some without mentioning the product. This is a self-reported method, not a representative sample or a posting strategy recommended here.
The useful principle is to connect each decision to a question feedback can overturn. Self-selection, community rules, wording, and repeated posting affect results. A more defensible editorial suggestion is transparent recruitment of target users, source labeling, and comparison with actual trials, repeat use, and payments. Forums are not an automatic answer bank.
Doing growth for its consumer product led the team to rebuild creator marketing as an internal Skill, adding communication, metrics, and execution details each day. The output was a spreadsheet: collaborators, prices, value assessments, content checks, communication progress, and filming and upload dates. “Delivering a result” became concrete.
Dense internal context helps iteration without proving external sale will work equally well. Data, permissions, responsible people, and shared objectives may be precisely what outside customers lack. Contacting and negotiating also require company authorization, and generating a sheet does not mean an agreement has been fulfilled. Those two boundaries are editorial additions.
5. Technology and staffing should follow the present bottleneck
Listen: 24:44–29:55; 48:30–55:16; 59:26–01:01:11. Mengqi criticizes early decisions shaped by financing narratives, impressive teams, and technical concepts. Exposure and engagement appear to offer natural rewards for training a growth agent, yet training costs, attribution quality, and foundation-model progress may make internal training uneconomic. This is her retrospective judgment, not a rule against training or algorithm hires across all application businesses.
Agent, workflow, chat, and graphical interfaces likewise cannot be ranked by novelty alone. A stable workflow may solve most of a customer’s problem. Adding fashionable concepts can increase expression and maintenance costs. Task outcomes, total cost, and experience remain the relevant comparison.
Development automation shifts difficulty toward the ends of the process. The team experiments with multiple agents receiving user issues, adding context, prioritizing, assigning coding, and notifying users. Simple interface problems may support automatic checks; complex issues still require people or users. Mengqi has not settled the external interface and therefore avoids rushing a platform release. A working internal pipeline and a viable external product remain separate achievements.
An editorial implication is to evaluate which problem should be fixed and whether it was truly resolved as code volume rises. Sending a notification is delivery activity; user verification is another piece of evidence. Issues, permissions, changes, tests, and final readback need to connect into a complete loop.
6. Efficiency and emotional value can coexist without concealing each other
Listen: 21:18–24:34; 55:38–58:27; 01:01:37–01:07:26. Pomodoro timers, habit tools, and desktop companions on Steam showed Mengqi that productivity, games, and emotional experience can overlap. Users may want completed work and a comfortable daily experience. Mentions of ADHD concern personal experience and product scenarios, not demonstrated treatment effects.
Her final formulation has an order: solve the problem while caring about the person and their current mood. Personality without reliable results cannot fulfill an efficiency promise. A correct result does not prove the interaction avoids anxiety. An editorial approach seeks separate evidence for task outcomes and experience rather than treating time spent as happiness or productivity.
Moving from proving her ability toward a product she wants to maintain is a personal reflection. Affection can sustain work without replacing user feedback; funding and famous hires are insufficient. She compares software with a lawn needing maintenance, while acknowledging that maintenance may itself become automated. Current difficulty is a work requirement, not necessarily a permanent moat. Her optimism rests on specific unresolved problems, not a guarantee for every project.
Exercise: review one decision that seemed logically inevitable
Write the original reason, such as “all these scenarios need sourcing, so customers will cross-buy.” Examine one actual user’s recent task: who operates, who pays, which step takes the most effort, where control matters, and what is delivered?
Judge technical reuse, user reuse, and business reuse separately. Technical reuse alone does not establish one market. Specify evidence that would stop expansion and behavior demonstrating that a valuable missing step was completed. For an experience improvement, compare switching, explanation, and rework between intent and result. This editorial exercise is not a financing or business-success formula.
Sources and listening guide
Prepared from the complete transcript and official introduction and chapters, without replaying audio or testing products. Anonymous allegations are not reported as established facts. Stability, version counts, spending, and feedback retain self-report attribution. No name is invented for the unnamed desktop product; closing lyrics are excluded from interview claims.
- 08:45: Starting with a vertical category.
- 11:40: Professional and non-professional user conflict.
- 12:28: Easy segments versus costly work.
- 30:21: From extension to desktop experience.
- 39:07: Research tied to decision questions.
- 44:00: Internal demand and concrete results.
- 01:05:00: Task value and attention to people.
Related themes: user research, creative tools, AI organizations, and companionship. Compare Wang Wenfeng on AI managing AI and Albert on success probability and consumer orientation for different product choices under similar technical conditions.