Don't Spend a Model on What a Rule Can Do: A Local-First Manual for Everyday Automation

Your receipt does not need a conversation.

Your thermostat does not need an agent. Your photo folder does not need a data center. A date in a PDF needs Find. A total needs a calculator. Printed text needs OCR. A fixed warning light needs a rule.

Sending all of that to a general-purpose generative model is not sophistication. It is failing to choose a tool.

When routine computing becomes a metered cloud conversation, people pay in token credits, network dependency, disclosure risk, and electricity they cannot inspect. The problem is not that generation exists. The problem is making it the tollbooth for work that ordinary computers already know how to do.

This is not a model ranking, a buying guide, or an affiliate pitch. It takes the user's side:

Use the least powerful tool that can finish the task. Climb only when the lower rung fails.

Local-first is a routing rule, not a purity test. The goal is not zero AI. It is justified AI.

The LOCAL ladder

Run this test before opening a chatbot.

The 30-second version:

1. Is the answer already present?
2. Can an exact rule produce it?
3. Can a narrow detector recognize it locally?
4. What does generation add that extraction cannot?
5. Does anything really need permission to act?

Use probabilistic generation for ambiguity. Use deterministic software for facts and rules.

Most “AI tasks” are smaller than they sound

The useful question is not “Which model should I use?” It is “What operation is this?”

These tools are not “less intelligent.” They are narrower—and narrow is often better. Their inputs are bounded, their output can be tested, and their failure does not arrive disguised as fluent prose.

Classical machine learning still matters. Logistic regression, Naive Bayes, decision trees, support-vector machines, random forests, and isolation forests can run on ordinary CPUs without generating a single token. OpenCV provides local thresholding, template matching, and background subtraction. Tesseract turns local images into searchable text. None requires a remote conversation to complete the core job [3, 4, 5].

Energy: measure the task, not the slogan

First, keep the units straight:

- Watts (W) measure power at a moment. - Watt-hours (Wh) measure energy used over time. - Tokens are billing or workload units, not units of electricity.

There is no honest universal energy price for “one AI prompt.” Model size, task, output length, hardware, batching, location, and accounting boundary all matter. Reasoning, agentic workflows, and video can require hundreds or thousands of times more electricity per query than simple text generation [1].

The system-level trend is clear. The IEA reports that global data-center electricity demand rose 17% in 2025, while electricity use by AI-focused data centers rose 50%. It projects total data-center use to grow from 485 TWh in 2025 to about 950 TWh in 2030 [1]. That is not a per-prompt figure, and not all data-center demand is AI.

So make the narrow claim you can defend: skipping an unnecessary model call avoids that model call. It saves the associated token charge and remote inference. It does not make the local device impact-free.

Do not buy an oversized GPU to perform a job that a dormant CPU, spreadsheet, or ten-line rule can handle. Local-first includes the hardware already on the desk.

Five everyday jobs, routed correctly

1. Find a deadline in a PDF

1. Get the original document and confirm its issuer and date.
2. Search for deadline, before, within, expires, and the relevant month.
3. If the page is an image, run OCR locally, then search again.
4. Record the exact sentence and page number. Inspect the page yourself.

Only escalate when conditions are scattered across sections or genuinely ambiguous. If generation is needed, send the relevant pages—not the complete personal file—and demand an answer that separates exact wording, interpretation, and uncertainty.

The final record should be four fields: date, action, exact source line, page. That is more useful than a polished paragraph.

2. Check a bill or receipt

Keep the raw document on the device.

1. OCR the merchant, date, line items, tax, credits, and printed total.
2. Validate fixed fields with rules: date format, currency, known account suffix, and required invoice pattern.
3. Put the figures in a spreadsheet.
4. Calculate sum(line items) + tax − credits and compare it with the printed total.
5. Mark unreadable digits as UNKNOWN; never fill them with a plausible guess.

Use a model only if you need a plain-language explanation of a verified fee. A language model may organize figures, but the formula remains the authority.

3. Sort files, photos, and messages

Start with metadata, not content:

- filename and extension - sender or domain - creation date and location metadata - exact hashes for duplicate files - perceptual hashes for near-duplicate images - a known folder or label vocabulary

If rules leave a meaningful remainder, train a small local classifier on examples you have reviewed. Keep an UNKNOWN bucket. Measure false positives and false negatives before moving files automatically.

For mail, scored rules plus a small Bayesian classifier can be effective and inspectable. Quarantine uncertain messages; do not auto-delete them. Apache SpamAssassin's own documentation warns about false positives and supports per-user Bayesian learning [6].

A remote model should not read an entire mailbox merely to learn that receipts from one domain belong in one folder.

4. Detect a household or workshop condition

Use a rule before a predictor.

IF temperature stays above 30 °C for 10 minutes:
  alert once
RESET only after temperature falls below 28 °C

That two-degree gap is hysteresis. It prevents repeated alerts near the boundary. Add a moving average if the sensor is noisy. If a fixed threshold cannot capture normal variation, test a local anomaly detector on recorded data and publish its false-alarm rate.

Detection is not diagnosis. A detector may say “different from baseline”; it cannot safely infer a gas leak, electrical fault, fire, or medical event without purpose-built equipment and accountable human judgment.

5. Research without enclosing the commons

Begin with public, inspectable sources: official documents, standards, public datasets, libraries, archives, and maintained community documentation.

Create a source ledger before drafting:

| Author | Title | Date | Link | Claim supported | |---|---|---|---|---|

Search, filter, bookmark, and extract locally. Escalate to generation only when synthesis across sources adds real value. Then cite the originals, distinguish source support from inference, and publish in a portable format with a correction route.

A public link is more democratic than an uncitable answer trapped in a private chat.

Public research should not become a private tollbooth for ordinary computing. Preserve the trail that lets the next person verify, reuse, and improve the work.

Prove that the “efficient” workflow is efficient

Measure a fixed batch—say 100 receipts, 100 images, or one week of sensor readings—rather than showing one favorable demo.

For a rough incremental measurement:

task energy (Wh) ≈ (average task W − idle W) × task hours

State what the measurement excludes. A serious accounting may also need device manufacture, networking, training, water, batteries, and e-waste; current ITU guidance therefore treats AI environmental assessment as a life-cycle problem [7].

Local is not automatically greener. Cloud is not automatically more efficient. The defensible comparison is the smallest sufficient workflow, measured end to end.

The public-interest test

A tool serves the person using it when the answer to most of these questions is yes:

- Can the basic job run without an account or permanent connection? - Can I inspect the rule, threshold, source, or decision log? - Can I export my inputs and outputs in a documented format? - Can I replace one component without rebuilding the whole workflow? - Can I contest an error and correct the record? - Can the organization operate it on-premises when privacy or continuity requires that? - Does an external action require a preview, explicit approval, and recovery path?

A free tier is not a commons. An API coupon is not ownership. “Weights available” is not automatically open source.

The Open Source AI Definition requires the freedoms to use, study, modify, and share a system, plus access to its preferred form for modification: sufficient data information, code, and parameters [8]. UNESCO's AI ethics recommendation adds proportionality, human oversight, sustainability, and a prior question public authorities should ask: is AI appropriate here, and if so, which method? [2].

Open code alone does not guarantee safety or democratic governance. But democratic technology has concrete properties: people can inspect it, contest it, leave it, and keep operating. Open standards and portable data make that exit real [9].

When a generative model earns its cost

Generation is useful when the remaining work has no fixed answer:

- reconcile messy or conflicting evidence - explain a difficult idea for a particular reader - translate tone and context, not just words - synthesize several verified sources - draft something genuinely new

Even then:

1. Extract and validate facts locally first.
2. Send the smallest necessary excerpt.
3. Choose the smallest capable model; use an approved local or on-premises model when practical.
4. Cap the output, avoid retry loops, and reuse verified results.
5. Require source links for factual claims.
6. Keep sending, buying, deleting, publishing, and safety decisions behind human approval.

The cloud is an escalation path, not a reflex.

Keep this rule

Locate the source. Use ordinary software. Compute close to the user. Ask a generative model only for the ambiguity that remains. Link it to other systems last.

Efficiency is not asking a data center to be slightly greener while routing every task through it. Efficiency is refusing to call it when a search box, formula, checksum, decoder, threshold, or narrow local classifier will do.

Save the model for the part that needs a model. Keep the sources, files, decisions, and right to switch in human hands.

Cite this guide

A Nguyễn (Nguyễn Hưng Tuấn Anh / Kody Anh Nguyễn). (2026). Don't spend a model on what a rule can do: A local-first manual for everyday automation. EFFOMA Practice Guide. https://effoma.com/blog/dont-spend-a-model-local-first-ai-manual

References

1. International Energy Agency. Key Questions on Energy and AI: Executive Summary. 2026 Apr 16. Accessed 2026 Aug 3.
2. UNESCO. Recommendation on the Ethics of Artificial Intelligence. 2021. Accessed 2026 Aug 3.
3. Tesseract OCR. Documentation. Accessed 2026 Aug 3.
4. OpenCV. “Image Processing Tutorials.” Accessed 2026 Aug 3.
5. scikit-learn. “Novelty and Outlier Detection.” Accessed 2026 Aug 3.
6. Apache SpamAssassin. “Mail::SpamAssassin::Conf.” Accessed 2026 Aug 3.
7. International Telecommunication Union. ITU-T L.1801: Framework for Environmental Impact Assessment of Artificial Intelligence Systems. Approved 2026 Feb 6. Accessed 2026 Aug 3.
8. Open Source Initiative. The Open Source AI Definition 1.0. Accessed 2026 Aug 3.
9. Foundation for Public Code. Standard for Public Code. Accessed 2026 Aug 3.