Stop wasting money on AI: three costly mistakes and how to fix them

Executive overview

Dylan Davis spent $300–$500 a month on AI before identifying the three root causes of waste: under-using subscriptions, buying AI products without proper evaluation criteria, and vibe coding without preparation. The core insight is that discipline beats spend — structured habits, binary evaluation, and minimal-complexity implementation consistently outperform throwing money at more tools or bigger models. Switching from over-engineered code to a single well-prompted model took one client's accuracy from 76% to 98% at one-fifth the cost. The same principle of trusting AI and staying simple applies equally whether you are a casual ChatGPT user or a developer building production applications.

Using AI as a consumer

• Spending $20–$200/month on subscriptions while only using a fraction of features is the AI equivalent of buying an iPhone and only using the flashlight.
• Invest 20 minutes a week learning new prompting techniques or newly released features for the tools you already pay for.
• Maintain a short list of three to five high-value use cases; apply every new technique directly to that list rather than exploring abstractly.
• Many leading models offer generous free tiers — Gemini 2.5 Pro (via AI Studio) and Grok cover most occasional use cases at no cost.
• Use Grok for real-time X/Twitter data and topics that stricter models decline; use Gemini for debugging, large file processing, and generating detailed documents.
• Mix free and paid access deliberately: pay only where frequency and value justify it.

Evaluating AI products before you buy

• When a vendor promises to automate everything, establish clear evaluation criteria before signing up — not after.
• Define a specific goal for the purchase, then build binary (pass/fail) evaluation criteria tied directly to that goal.
• Binary evals are actionable; spectrum-based scores (1–10 or A–F) invite subjective debate and obscure real performance.
• Example binary evals for an AI sales rep: does it book a follow-up meeting? Does it surface a genuine pain point? Does the call end in under 60 seconds?
• Track 10–15 binary evals consistently across the vendor's trial period (typically 30–90 days) and require steady improvement.

Implementing AI with lower risk

• Start with internal automations before touching customer-facing products; internal failures are recoverable, product failures damage brand and revenue.
• Use a discovery process: identify the 12–18 month priority goal, map the blockers, rank them, then automate the highest-ranked blocker first.
• Always begin implementation with the largest, most capable model available to confirm the task is automatable at all.
• Once viability is proven, move to a smaller, cheaper model and chunk the single prompt into sub-tasks if needed — smaller models running in series are faster, cheaper, and often more reliable.

Trusting AI instead of over-engineering

• Over-engineering is expensive: 25,000–30,000 lines of custom parsers, regexes, and post-processing scripts achieved only 70–76% accuracy on PDF extraction.
• Replacing that code with a single strong prompt to one prominent model achieved 98% accuracy, was three times faster, and five times cheaper.
• Remove code first; only add complexity after proving AI cannot handle the task alone.
• Revisit and simplify implementations regularly — model improvements mean yesterday's over-engineered workaround may be unnecessary today.

Preparing before vibe coding

• Spend 60–70% of vibe-coding time on preparation, not on writing code — sharpen the axe before swinging it.
• Preparation means producing three documents before any code is generated: a specification (what to build), a blueprint (how to build it), and a to-do roadmap (the sequence the AI will follow).
• Research current API documentation and add it to the spec or blueprint; AI models have training cut-offs and will otherwise reference outdated docs and produce wrong code.
• Without this preparation, the AI loops on errors, wastes tokens, and inflates costs.

Best practices during vibe coding

• Start a fresh conversation for every new feature, test, or sub-task; context windows degrade AI quality noticeably past ~80% capacity.
• Lead each fresh conversation with the spec, blueprint, and to-do list to maintain coherent direction without relying on accumulated context.
• Avoid stacking new starter kits, scaffolding tools, multiple parallel agents, and task-manager frameworks — each layer of complexity makes it harder for the AI to act effectively.
• Use the terminal tools the AI was already trained on; surgical guidance plus familiar tooling beats elaborate orchestration every time.
• Simplicity is a force multiplier: fewer abstractions mean faster iteration, lower token usage, and easier debugging.