SaaS

Adding AI to SaaS: Start With the User’s Unfinished Work

Rob Walling’s eBay example shows why generating text can leave the hard work untouched. His 6 patterns help locate a more useful AI feature.

Listen to the audio analysis

Create your own

Written for builders and creators.

Most SaaS companies are currently wasting engineering cycles on "Lazy AI.

Editor's note

Why this matters now

Rob Walling's criticism of an AI feature at eBay is specific: generating a listing description leaves the seller entering the product details. He would rather see a photo help identify information such as the manufacturer, year and UPC, where available.

For a SaaS founder choosing what to add next, that is a useful distinction. Which part of the user's task does the feature remove, and which part does it leave behind? Walling offers 6 patterns for finding a more useful intervention. His suggested target of saving 15 minutes is a product heuristic, not evidence that every shorter task is unworthy of automation.

The source

What it says

Distilled from the original. The notes above and below are the editor's own.

Start With the Work Around the Text Box

In the example discussed, Walling estimates that a generated eBay description might save about 30 seconds. The seller still has to supply the details that make the listing accurate and useful.

The point is about where assistance lands in a workflow. Better prose may help, but it does not necessarily address the part the user finds slow or difficult. His alternative, extracting item details from a photo, is a proposed improvement. The episode does not establish how accurately that proposed feature would identify every field.

Walling calls superficial additions “lazy AI.” His time-saving threshold gives product teams a way to challenge a feature whose appeal comes mainly from adding a familiar AI button. It works best as a question about the user's outcome, rather than a rigid stopwatch rule.

6 Patterns for Locating the Opportunity

Walling organizes AI features by the kind of work they perform. The categories are useful because they expose options beyond generating another paragraph.

Conversation can simplify a complicated search. His example replaces a screen full of filters with a request such as finding US agencies with fewer than 10 employees. The useful result is the correctly filtered set of records. A pleasant chat interface alone does not complete that task.

Generation helps create an output. Text, images and code fit here. Walling points to SessionLab's guidance and suggestions as a way to show users what they can ask for, rather than leaving them with an empty prompt box and no idea where to begin.

Categorization handles repeated sorting. Tagging emails or digital assets can save someone applying the same labels across a large collection. The task is narrower than an open-ended assistant, which can also make its value easier to explain.

Ingestion turns an existing input into a usable record. A receipt, whiteboard photo or voice note contains information someone would otherwise retype. This pattern removes a step between having the information and using it in the application.

Analysis evaluates information against a purpose. Walling's examples include reviewing a sales call for whether a representative addressed the customer's goals. A useful response needs to explain the assessment, so the person using it can judge the reasoning.

Task execution lets an agent act through software. Scheduling a meeting is one example. Here the product needs an interface through which another system can request and complete an action. In the episode, Walling discusses both MCP and command-line interfaces, while describing adoption and interface choices as unsettled.

These categories can overlap. A voice-note workflow might ingest the note, categorize the request and initiate an action. Naming the steps helps a team see where a proposed feature ends and what still needs to happen.

Ingestion Shows Why the Last Step Matters

The construction example discussed in the episode begins with a foreman sending a WhatsApp voice note about materials. The intended result is a structured material request matched to inventory.

That requires more than a transcript. The system has to identify the item, map it to the application's records and check that the resulting fields are usable. A correctly typed field can still contain the wrong item, so structural validation and correct interpretation are different requirements.

This makes ingestion a useful place to look when a product asks users to enter information they already have elsewhere. The feature earns its place by producing a record the next part of the workflow can use. Simply moving an unstructured note into another screen leaves the user to finish the translation.

A Useful Feature Still Needs Viable Economics

Walling describes 3 ways to charge for AI use: absorb the cost, include a usage allowance in a subscription tier, or sell additional usage separately.

The tradeoff follows the task. An occasional import may be easier to include than a feature a customer can run repeatedly all day. A separate allowance can help the business control exposure, but the buyer needs to understand what the allowance buys. An unexplained token balance can obscure the cost of completing a job.

This is also where an apparently helpful feature can fail commercially. A team needs to know both what work it removes and what each completed task costs to deliver. The pricing discussion on changing rates versus changing buyers addresses the broader offer behind that calculation.

The Workflow Can Matter More Than the Model

Walling warns bootstrapped companies about competing directly with foundation-model providers. He describes a TinySeed-backed company whose model advantage was overtaken by larger providers. That is one investment experience, not proof that every specialized model business will fail.

His product argument is to build around the customer's workflow: its inputs, records, integrations and required result. Those are also relevant to the defensibility questions raised during a SaaS sale. Adding AI capability and creating a business that is difficult to replace are related but separate jobs.

Editor's note

What to do with this

The eBay example is useful because it follows the task past the generated output. The seller still needs a complete listing. In the construction example, the foreman needs the right material request to reach the next person.

That is the standard a roadmap discussion can apply: explain what the user starts with, what the feature produces and what they still have to do. Time saved helps quantify the improvement. The more basic test is whether the feature completes a meaningful part of the job.

The original

Stop Adding Lazy AI to Your SaaS

Rob Walling · 6 July 2026

  • SaaS

    Why SaaS Growth Can Stall Around $1M ARR

    Julian Marzouki and Rob Walling explain how founder approvals, competing priorities and gaps in sales data can hold back an established SaaS.

    4 min read7 min listen

All posts