AI Coding Field Notes

Field notes on AI coding agents — what they cost, where they break, what shipped. Every published figure (prices, percentages, multiples, token counts and durations) as JSON and CSV, each row carrying the sentence it came from.

View the Project on GitHub xyzs996/ai-coding-field-notes

Boosting AI Bot Conversion: A Deep Dive into Funnel Data

Boosting AI Bot Conversion: A Deep Dive into Funnel Data

Written with AI assistance. Figures without a traceable source were cut before publishing.

One reported case moved entry-group conversion from 9.1% to 55.1% by rebuilding an automated onboarding flow around what the funnel data actually showed, rather than around what the team assumed users were doing. The interesting part is not the size of the jump. It is which stage turned out to be responsible.

Understanding the Funnel: A Data-Driven Approach

Analyzing user behavior at each stage revealed the critical drop-off points, and the number that reframed the problem was this one: 83.1% of unmanaged groups contributed only 2.3% of new group entries. Unregistered groups, in other words, were not where the traffic was being lost — which is exactly where a team working on instinct would have spent the next month. After sharing AI installation steps on Douyin, the same operation attracted users who were interested in AI but had no programming background, and that audience confirmed the real bottleneck: installation. The 2.3% is the weakest number in the set, and I’d treat it as directional rather than precise.

Pricing was refined against that same feedback. Setting the course price at 9.9 yuan tested one specific proposition, that people would pay for a simpler installation path, and produced 2 sales within 48 hours, for total revenue of 19.8 yuan. Two sales is not a business. It is a signal, bought cheaply, and I think that is the correct use of a price test. At 9.9 yuan the experiment cost less than the analytics work needed to read its own result.

The broader lesson for independent developers is that the fastest path to AI adoption usually begins with a minimal closed loop rather than a platform. A zero-code novice built a company business system covering scheduling, CRM, and AI sales analytics using Codex and Claude, at a core cost of ¥560 in subscription fees, serving a team of over 40 people. That is the whole argument in one line: prototype a minimal viable workflow, iterate it against real business data, then scale. The ¥560 figure is the part I’d want independently confirmed, because Codex and Claude are billed per seat and 40 people is not 1 seat.

Enterprise AI adoption rarely starts with a fully deployed Agent. It starts when employees change how they work on low-risk, high-frequency tasks, and the usage records those tasks leave behind are what eventually expose which patterns are worth formalizing into an organizational capability rather than leaving them as one person’s private trick. Adoption is not a deployment problem. It is a working-habits problem, and the sequencing follows from that.

For independent developers, the lesson is narrower: solve a pain point someone can name. In one case, a designer built customized website demos and high-quality images to approach small restaurant owners, converting information asymmetry into revenue. The critical factor was that the solution looked non-obvious to that audience. Obvious solutions do not sell.

AI-powered recruitment automation makes the same point from a different angle. Breaking job requirements into observable, measurable dimensions is what lets AI improve both hiring efficiency and candidate retention, because it replaces a vague judgement with a checkable one. Demand was validated the cheap way: sharing a recruitment automation tutorial on WeChat produced dozens of requests for more information.

Optimizing the Funnel: Key Insights

A gradual approach to AI implementation matters more than aiming straight at an “AI-native” product, and the most expensive counterexample is well documented. Klarna replaced 700 human customer service representatives with AI assistants, then had to rehire after customer satisfaction fell. Klarna is the reference case here because the reversal is documented, not because 700 is a big number.

De-AI-fication is what a mature deployment looks like from outside: the automation is there, and the customer cannot tell where it starts. Matching AI capability to what users actually need is the whole of product-market fit here.

Implementation should move through four stages. Employees first execute the process manually, which is what establishes the standard; AI then handles the repetitive parts under human oversight, more complex tasks follow once that oversight stops catching anything worth catching, and only at the fourth stage does AI sit inside core business operations. Each stage exists to make the next one cheaper to reverse.

For independent developers, the first question is whether your service is discoverable by AI in automatic mode at all. Only after that is confirmed does optimizing conversion in development mode make sense. Visibility first, then the high-value conversion paths, because spending in the other order is how budgets disappear.

Moving from a traditional product to an AI-native one forces a rethink of product boundaries and uncertainty handling. Unlike a traditional product, an AI product must account for unstable outputs and messy user intent, which means the design has to stay reliable while the responses vary. That constraint is doing more work than most roadmaps admit.

The WeChat Small Business Agent is in internal testing, with traffic distribution shifting from ranking to AI readability, and the practical consequence for anyone selling a service through that channel is that the AI selects and executes one service directly while the user never sees a list to browse at all. Independent developers should validate discoverability in automatic mode before optimizing selection certainty on high-value paths. A ranked list you cannot appear in is not a ranking problem.

AI product managers, meanwhile, need a phased learning path across 3 steps: product knowledge, then industry understanding, then hands-on AI project implementation. Skipping step 2 is the common failure, and it shows up later as an AI roadmap nobody in the industry can price.

Technical Considerations: Handling Interfaces and Uncertainties

Separating platform interface returns from business facts keeps a technical uncertainty from being read as a business rejection. The system in this case repeatedly mistook order inquiries for qualification gates, when orders were only triggers. That single confusion produced rejections that no business rule had ever asked for.

Klarna’s experience says the same thing at company scale: implementation should progress through stages of human empowerment, from manual process to automated workflow, so that the solution stays aligned with what the business actually needs. The framework’s core claim is that employees must be allowed to adapt and refine a process before automation is layered on top of it.

Deconstructing the process into manageable components first is the most effective strategy available to an independent developer. Executing tasks manually before automating them is what tells you what a successful outcome even looks like, and a clear definition of “good” is the thing that keeps an AI system from confidently optimizing the wrong target. Assign multiple responsibilities to a single agent and you get confusion, higher token consumption, and rework. Role-specific agents outperform generalized ones because each one has less to be wrong about.

Finally, implementation succeeds or fails on translating technical output into business insight. That takes understanding both what the model can do and what the business specifically needs. Klarna’s case shows AI works best supporting existing processes rather than replacing them outright.

Replicability and Limitations

These strategies raised conversion in 1 operation, with 1 audience, at a single 9.9 yuan price point. Nothing here establishes that the same sequence transfers.

What does look transferable is the order of operations:

Define the business requirement and its constraints first, then put AI into low-risk, high-frequency scenarios where a wrong answer costs little and the usage record is still real. That staging is what lets you validate a capability before anything important depends on it.

The reported enterprise case reached 55.1% from 9.1% through exactly that loop: analyze the behaviour, change one stage, measure again. No single clever intervention appears anywhere in the account, which is the most credible thing about it.

Originally published on Medium.

Also readable on Telegraph.

Read next

All 26 write-ups

The 8 figures in this piece — each with the sentence it came from — are in the figures table, alongside 262 more, as JSON and CSV.

Topics: AI Implementation

Part of ai-coding-field-notes — field notes on AI coding agents. Found something wrong, or shipped something similar? Open an issue — corrections are the point.