Convince Me AI Can't Do It
Assume AI can solve the problem, then locate the precise constraint when it cannot
- Difficulty
- Easy
- Time to result
- ~weeks to results
- Steps
- 6
- Confidence
- 94%
Convince Me AI Can't Do It reverses the usual burden of proof in AI projects. Instead of asking whether AI might help with an established workflow, the team begins with the underlying problem and assumes that a capable AI system can address it. The team then builds the strongest practical attempt and studies the precise point of failure. A weak result might come from missing company context, inaccessible business data, inadequate tools, poor orchestration, or an actual model limitation. Each suspected cause is tested rather than asserted. The output is either a working AI-enabled process or a clear account of the remaining constraint. This reframing exposes hidden assumptions, encourages first-principles design, and helps teams master the modular capabilities they can combine into previously impossible products or campaigns.
Origin
Holland adopted this challenge with his product teams after a blog-post creator produced poor content despite having a seemingly good prompt. The diagnosis revealed missing context and led to investment in a broader context layer. Extracted from Marketing Against the Grain.
Core principles
- 01Start from the underlying problem rather than the current workflow
- 02Treat poor AI output as evidence to investigate
- 03Separate missing context from weak prompts and true capability limits
- 04Understanding each constraint creates reusable building blocks
How to run it
- 1
Restate the problem
Describe the desired outcome independently of the current workflow, staffing model, or software implementation.
Pro tip Ask what the customer actually needs rather than which existing task should be automated.
Watch out Starting from the incumbent process may preserve constraints that no longer apply.
- 2
Adopt the capability assumption
Challenge the team to demonstrate why AI cannot achieve the outcome. This shifts discussion from general skepticism to testable limitations.
Pro tip Frame the challenge as an investigation, not as pressure to exaggerate AI's reliability.
Watch out An assumption used for discovery is not proof that the system is production-ready.
- 3
Build the strongest attempt
Use the best suitable model, prompt, tools, data, and workflow available to produce a representative result.
Pro tip Evaluate a complete task using realistic inputs rather than a toy demonstration.
Watch out Testing an outdated or free model may reveal the limits of that model rather than the limits of AI.
- 4
Locate the failure
Specify what is inadequate in the output and where the process broke down. Avoid broad conclusions such as saying the AI simply is not good enough.
Pro tip Compare the output against explicit acceptance criteria.
Watch out A good-looking output can still fail on factual accuracy, brand fit, or operational reliability.
- 5
Test each suspected constraint
Add missing company context, connect relevant data, improve orchestration, or try a more capable tool. Re-run the task after each material change.
Pro tip Change one major variable at a time so the source of improvement remains visible.
Watch out Do not expose sensitive business data to an unapproved system while testing context.
- 6
Decide from evidence
Adopt the workflow if it meets the required standard; otherwise record the specific unresolved limitation and the conditions under which it should be retested.
Pro tip Preserve successful components as reusable building blocks for other problems.
Watch out Do not confuse partial capability with dependable end-to-end automation.
In the wild
Holland's team had invested heavily in prompting, yet its blog-post creator still produced poor content. By asking where AI was actually failing, the team recognized that the prompt was not the central problem: the system lacked the company's relevant context. That finding redirected the product effort toward memories and a broader context layer.
→ The team identified a structural input deficiency instead of continuing to polish the prompt.
A marketing team asks AI to create a launch brief and initially receives generic recommendations. Rather than abandoning the approach, it tests the missing inputs separately: customer research, positioning documents, previous campaign results, and channel constraints. The team then evaluates the enriched output against its normal briefing rubric.
→ The team learns whether the remaining gap is solvable through context or requires human strategic judgment.
Common mistakes
Automating the inherited workflow
Reproducing every step of an old process can obscure a simpler route to the desired outcome. Begin with the problem and its acceptance criteria.
Calling a context failure a model failure
Even a strong model produces generic work when it lacks company, customer, or brand information. Test missing inputs before rejecting the capability.
Forcing a predetermined conclusion
The challenge should uncover evidence, not compel the team to claim success. Genuine reliability, safety, or capability limits must remain visible.
Is it for you?
Best for
It is best for product and marketing teams evaluating whether AI can automate or transform an existing task.
Not ideal for
It is not ideal for high-risk decisions where an unverified AI attempt could directly affect safety, compliance, or customers.
From the transcript
“I have now started to ask my teams convince me that AI can't do that”
“if you just started with the premise of this isn't good enough, where is AI falling down?”
“But the reality is it didn't have all the context.”
From the episode
This is How the Top 1% Marketers are Using AI