Running out of AI tokens quickly? Here's why and how to save on token usage.
- The AISI Team
- Apr 16
- 4 min read
In 2026, every business is an AI business. But while most leaders focus on the capabilities of Generative AI, very few are looking at the meter.
At AISI, we’ve audited enterprise AI workflows where over 30% of the monthly spend was attributed to "token waste"—redundant, poorly formatted data that adds zero value to the output but drives the bill straight to the ceiling.
If you aren't managing your "Tokenomics," you aren't just using AI; you’re subsidising the compute costs of Big Tech. Here are 7 expert tricks to slash your token usage without sacrificing a shred of quality.
1. Beware of the "Chat Debt" Cumulative Cost
This is the single biggest invisible cost. AI models don't actually "remember" your past messages; they re-read the entire conversation history every time you send a new reply.
The Waste: By the 20th message in a thread, you are paying to send those first 19 messages back to the server over and over again.
Our Tip: Use the "New Chat" Rule. Once a specific task is done, hit "New Chat." If you need to carry context over, ask the AI to "Summarise our key decisions in 3 bullets," then paste that summary into a fresh session.
2. Convert PDFs to Markdown or Plain Text
PDFs are the "heaviest" way to feed data to an AI. They contain complex layout information (fonts, margins, positions) that the AI must navigate.
The Waste: PDF "bloat" can triple the token count of a document compared to its text equivalent.
Our Tip: Copy the text into a .txt file or, better yet, a Markdown (.md) file. Markdown uses simple symbols (like # for headings) that help the AI understand structure with almost zero token overhead.
3. Optimize Your Visuals (The Image Resize)
Modern models like GPT-4o and Claude 3.5/4 "see" images by breaking them into tiles. A high-resolution 4K image is chopped into dozens of tiles, each costing hundreds of tokens.
The Waste: Uploading a high-res screenshot of a dashboard can cost more than a 10-page essay.
Our Tip: If you’re asking for a critique or a data summary, clip only the part you require (rather than the whole screenshot) and resize the image to roughly 512x512 or 1024x1024 before uploading. The AI is smart enough to "see" the content without needing every pixel.
4. Stop Uploading Raw Excel Files (The CSV Rule)
An Excel file (.xlsx) is essentially a giant container of XML code, formatting instructions, and metadata. When you upload it, the AI has to "parse" all that invisible structure.
The Waste: A 50-row spreadsheet in Excel might cost 5,000 tokens just to "read" the formatting.
Our Tip: Save your file as a CSV (Comma Separated Values). It strips away the styling and leaves only the raw data. It’s faster, leaner, and often 80% cheaper to process.
5. Force a "No Preamble" Policy
AI models are naturally verbose. They love to start with, "As an AI language model, I would be happy to assist you with..."
The Waste: Over thousands of employees, these polite "fluff" sentences add up to millions of wasted output tokens.
Our Tip: Standardise your prompts with a "No fluff" clause. Explicitly state: "Provide the answer directly. No preamble, no pleasantries, no post-script."
6. Use "System Instructions" for Recurring Style
If you find yourself telling the AI "Write in a professional tone" in every single prompt, you are wasting tokens.
The Waste: Repetitive instructions in every prompt accumulate massive waste across a department.
Our Tip: Use the "System Prompt" or "Custom Instructions" feature. This defines the AI's behaviour once at the start of the session, rather than you having to pay for those instructions in every single message you send.
7. Master Context Caching (The Enterprise Secret)
Leading providers (Google Gemini and Anthropic Claude) now allow for Context Caching. If you frequently ask questions about the same 200-page company manual, you can "cache" that document on the server.
The Waste: Re-uploading your knowledge base for every new employee query is a financial disaster.
Our Tip: By using caching, the AI "freezes" that data. Subsequent queries cost up to 90% less because the AI doesn't have to "re-read" the base document.
The Bottom Line
At AISI, we specialize in AI Consulting and AI Solutions. We don't just build agents; we build efficient agents. In the race to automate, the companies that win are those that understand how to scale their intelligence without exploding their infrastructure costs.
Ready to stop the leak? We’ve created a "Token-Efficiency Cheat Sheet" for your team to keep their AI usage lean and high-impact.
Download our 1-Page Token Usage Efficiency Cheat Sheet here: -
Want a professional audit of your AI workflows?
AISI is a Singapore-based AI solutions provider specialising in AI Agents, AI Video Analytics and Enterprise AI implementation. Reach us here or email us at specialist@aisi-asia.com
Sources: * OpenAI (2026) "Understanding Vision Token Usage."
Anthropic (2026) "Prompt Engineering for Cost-Efficiency."
Google DeepMind (2026) "Context Caching: Technical Implementation and Savings."
Comments