Retool’s AI bet: automate the workflow, keep engineers in control

Customers before Demo Day, traction before hiring

In March 2017, Retool was supposed to present at YC Demo Day. It had no customers and no traction. David Hsu recalls choosing to wait six months to build a product and find customers the team could be proud to present. Watching peers raise money was painful, but customer value mattered more than completing the startup ritual.

Retool gives developers reusable building blocks for assembling internal applications—the interfaces and back ends businesses need to run their operations. Building that business initially meant roughly a year to a year and a half of writing code and talking to customers. Hsu remembers raising about $1 million–$1.5 million coming out of YC, with two founders paying themselves around $30,000–$40,000 annually. In his recollection, cash could last years; motivation was the scarcer resource.

That explains his preference for a small team before product-market fit: finding a product customers want. Employees need direction even while founders are discovering what to build. Repeatedly supplying conviction without clear answers can exhaust founders and leave employees in an unstable position. Hsu favors hiring once customers and product direction provide a firmer foundation. When Retool expanded, he estimates that former founders made up about half the team at the 20-, 30-, and 40-person stages; he credits them with establishing a self-starting culture focused on customer outcomes.

Retool’s fundraising approach extended that preference for substance: raise less money at a lower valuation and accept leaving money on the table. The argument distinguishes two costs: - Dilution reduces existing ownership. A financing that dilutes existing holders by 10% proportionally reduces employees’, founders’, and previous investors’ stakes. - A valuation driven by enthusiasm creates a future hurdle. Investors can pay more without the business changing. If that enthusiasm disappears, subsequent fundraising and employee morale can suffer. The retrospective assessment is that Retool benefited from avoiding an exceptionally high valuation. It had not predicted the market downturn.

Faster code generation leaves a verification problem

Internal applications involve repetitive but necessary work: state management, data validation, error messages, and interfaces over business data. Retool packages that work for developers. It deliberately avoids the “low-code” label because the term can suggest developers are not the intended audience; approximately 95% of customers are described as developers.

AI pair programming appeared on Retool’s roadmap as a joke around 2019. By the early-2024 conversation, the team was building it. The strategy separated two opportunities: - Use AI to build applications faster: generate SQL queries and user interfaces to reduce implementation work. - Help developers build AI-enabled applications: make AI easier to incorporate into everyday business software. Retool was investing in both, with most investment going to the second.

The proposed bottleneck is that generating JavaScript does not remove the need to understand it. An engineer must still read the result, verify behavior, and diagnose subtle errors. The respondent estimates that assistants such as Copilot make coding perhaps 10–20% faster—an informal assessment from conversations with engineers, not a benchmark. The hypothesis is that existing languages and frameworks leave too much work at a low level of abstraction.

The analogy is automated punch-card production: producing cards faster helps, but a bug still sends the programmer back to reading them. Retool’s opportunity is to reduce the work of assembling an application through reusable building blocks, beyond accelerating its initial code generation.

Business context supplies the AI use case

Retool’s AI application strategy rested on an early-2024 prediction: practically every internal application would incorporate AI within roughly three years. A reported Fortune 500 sales-forecasting application illustrates the idea. It pulls deals from Salesforce and uses previous deal dynamics to estimate likely closing periods; it could then prewrite the salesperson’s forecast report. The useful unit is a business task connected to its existing data.

[Retool AI](https://retool.com/products/ai) packages AI actions, while [Retool Vectors](https://docs.retool.com/ai/concepts/vectors) brings vector storage into the application environment. The aim is to reduce integration work, including model selection and testing. Retool chose pgvector partly because open source lets engineers inspect implementation details, fix bugs, and contribute changes upstream. The respondent expects open source to win in tooling and supporting infrastructure, while explicitly leaving the model question open.

A more surprising example comes from an unnamed multinational clothing manufacturer. Its seasonal design process involved studying sales data, identifying popular patterns by region, and designing the next collection. Its reported Retool application lets analysts examine those regional trends and feed the findings into DALL-E and other image-generation APIs to produce patterns the company prints.

Finding an AI application can require learning how a particular business works. In January 2024, the respondent saw many AI founders still searching for use cases, while established enterprise customers already had processes worth improving. The clothing example makes that opportunity concrete, though it provides no measured savings or design-quality results.

Developers control the sequence; AI works inside the steps

Customers were already building AI workflows in [Retool Workflows](https://retool.com/products/workflows). The attraction starts with a limitation of chat: someone must bring it the task. Chat works well for a one-off document draft, but a person still has to supply the data and transfer the answer back into the working system. That reactive interface covers only part of a role such as a bank relationship manager.

Retool’s proposed next step is developer-controlled orchestration: developers assemble and edit the sequence of operations, while individual steps can use AI. An AI-generated draft may provide a starting point, but the developer must be able to modify it. The rationale is that model errors can compound across a multistep agent; the conversation’s reliability example illustrates that concern without establishing a measured end-to-end error rate.

The clothing manufacturer’s workflow makes the division concrete: 1. Ingest daily sales data into a database. 2. Analyze it with SQL to identify relevant patterns. 3. Generate design ideas with AI. 4. Send designs to analysts for approval or rejection. The reported volume is 10,000 generated designs per day. That is generation volume, not accepted designs or a measured productivity gain. Human review remains part of the process.

Workflows with human review are the respondent’s nearer-term bet for business value: automate the repeated transfer of data and prompts while preserving review. Retool’s broader product philosophy follows the same allocation of control. Developers want reusable tables, state management, and API access so they can build the final application themselves. For a sales-operations tool, the stated competitive choice is often React or Retool.

The build-versus-adopt rule is build where Retool can offer a distinctive customer benefit. Workflows is treated as a core competency, positioned between Zapier and Airflow. Vector databases already offered a thriving ecosystem, so adopting pgvector made more sense than creating another implementation.

Stack Overflow: reported substitution, not a traffic measurement

The discussion of Retool’s 2023 State of AI report opens with roughly 1,600 respondents surveyed the previous August and an interviewer’s citation that 52% considered AI overrated. By early 2024, the respondent saw less hype and slightly more useful applications, but continued uncertainty about returns on substantial investment. They predicted greater skepticism if returns failed to appear within roughly a year.

The consequential Stack Overflow finding is a conditional one: 94% of respondents who used Stack Overflow less attributed that reduction to Copilot and ChatGPT. The interviewer is reporting why a subgroup changed its behavior. This does not establish an overall percentage decline in usage or measure site traffic.

The companion links to a Stack Overflow traffic discussion as a dispute reference, without specifying the scope of the disagreement. The distinction matters: developers changing where they seek help and the magnitude or cause of a website’s traffic decline are different claims. The reported survey result supports the first.

An interviewer then compares expected job impact: operations respondents rated it at 8 out of 10, while designers—the lowest group cited—rated it at 6.8. The clothing example makes the contrast interesting without explaining why respondents differed. The response shifts to implementation: engineers need to experiment with prompts, model limitations, and supporting data infrastructure to make AI useful for a business task.

Business priorities, engineering execution, human review

The proposed responsibilities are distinct: - The business chooses the bottleneck: identify a function that is difficult to scale and worth improving. - Engineering builds the intervention: experiment with models, prompting, and integration. Retool’s support queue provides the example. The product’s flexibility generates many customer questions, and the company was exploring AI-drafted replies. Doubling support-agent productivity was a hoped-for outcome, not an achieved result.

Hiring raises a related distinction. An interviewer cites a survey finding that 45% of companies made engineering interviews harder to compensate for AI assistance, then questions why an unchanged job should require a harder interview. The respondent values competent AI-tool use while still testing critical thinking and computer science fundamentals.

Retool’s described interviews use screen sharing or a shared coding pad without Copilot to observe candidates’ reasoning. Internally, engineers are encouraged to use Copilot. The distinction is assistance versus dependence: candidates should still be able to reason through basic code, such as a for loop, without relying exclusively on generation.

Deployment brings the discussion back to internal software. A core Retool belief is that most software serves businesses’ own operations: a clothing manufacturer employs engineers to help sell clothing rather than software. Internal AI applications can tolerate more unreliability when a person checks the output. The respondent is less certain about AI embedded in consumer banking or support without explicit consumer awareness, where fabricated answers could be costly.

Model choice turns on quality, cost, and data control

An interviewer reports that 80% of survey respondents used OpenAI and asks about alternatives. The respondent sees growing enterprise acceptance of hosted models, including access through Azure for employee experimentation. Some data remains harder to send outside the company; unreleased earnings results are the example. Open-source models could help address those privacy constraints.

In the respondent’s early-2024 assessment, GPT-4’s quality advantage over Llama 2 helps explain customer preference, while OpenAI’s aggressive pricing discourages switching. They see that combination as especially difficult for model startups to challenge, though well-funded competitors such as Google and Facebook may be harder to displace. These are strategic judgments, not task-specific benchmark results.

The proposed switching condition is explicit: if open models reach comparable quality, speed, and cost, data control may favor on-premises hosting. That comparison is hypothetical; it does not claim parity had arrived.

For the next survey, the respondent wants similar questions to track changes in model and tooling satisfaction. An interviewer proposes separating demand for image generation, visual comprehension, audio, text-to-speech, and video. Multimodality bundles distinct product needs: interest in AI spanning several media types does not tell a builder which capability users need first.

Sales proves demand; developer adoption expands reach

Retool initially relied heavily on sales because close customer work exposed needs and guided product development. Brief usage after a launch was a weaker signal: people trying an AI project and then leaving did not clearly establish product-market fit. The sales-first approach generated revenue and traction while allowing slower hiring.

Its longer-term ambition required broader adoption: roughly 10% of the world’s code written in Retool, or 10% of software running on it. These were aspirations, not market-share figures. The respondent says Retool changed its primary success metric in the preceding year from revenue or new annual recurring revenue to the number of developers building on the platform. Reaching tens of millions of developers would require self-service adoption beyond what a sales force could support.

Amazon illustrates the gap between purchasing and awareness. The respondent estimates that 11 separate Amazon business units used Retool, while suggesting that the average engineer there had probably never heard of it. Wider use would depend on positive product experiences spreading through coworker recommendations. Enterprise sales and bottom-up adoption address different parts of distribution.

Must intelligence look human?

The closing exchange returns to Hsu’s early interests. His Oxford degree combined philosophy and computer science, including questions about computers, intelligence, and logic. The discussion of artificial general intelligence, or AGI, asks what would make a participant recognize a system as generally intelligent.

One proposal connects human intentionality—here, the pursuit of goals—to evolutionary pressures toward survival and reproduction, with happiness tentatively added as another aim. The participant wonders whether analogous goals in a simulated world of AIs might produce something recognizable as AGI. It is a philosophical thought experiment.

Other participants explore alternative frames: - Selection through datasets: one voice suggests that knowledge transferred between models might constitute an evolutionary process. They speculate that a sentient model could seek to spread its outputs into future training data to influence later models. - Intelligence at a different scale: another uses Hofstadter’s ant-colony analogy to ask whether connected models, rather than individual weights, might be the useful units for thinking about emergent intelligence. - Machines need not reproduce biology: a participant challenges strict reliance on human evolution and development as templates for AI. These remain competing analogies, without an experiment resolving them.

The final analogy is an airplane: it flies without being a bird. One speaker argues for judging AI by whether it achieves desired outcomes, without requiring humanlike internal mechanisms. Successful behavior need not imply a humanlike implementation. The exchange leaves thought and consciousness unresolved.