Guide · Cost

Cut AI and automation costs before your 2027 budget

By Mayank Raj · Kingdom of Kumar · Updated 4 November 2026

To cut LLM and automation spend, work in this order: measure cost per task, send easy work to a small model and escalate only when needed, cache repeated prompt prefixes, move work nobody is waiting on to batch APIs, trim context and output, and only then look at replacing per-task automation pricing. We give no savings percentages for your stack because we cannot know them. The vendor discounts below are quoted from their own docs.

1. Measure cost per task first

A monthly invoice tells you what you spent, not what for. Before you change a prompt, log one row per task: workflow name, model, input tokens, output tokens, retries, and whether the result was accepted. Divide by completed tasks, not by calls. A workflow that retries three times costs three times what its invoice line suggests.

Two or three task types will usually carry most of the bill. Fix those and leave the rest. If you do not have this table, build it before anything else on this page, because every later step needs it to prove it worked.

2. Route by difficulty: small model first, escalate

Classification, extraction, tagging and routing rarely need your largest model. Run the cheapest model that passes your own checks, and escalate to a larger one only when a validator fails or the model reports low confidence. Keep the escalation rate in your cost-per-task table. If most tasks escalate, the routing is not helping and you have only added a hop.

Test with your own examples. A published benchmark does not tell you how a model handles your invoices or your support tickets.

3. Cache repeated prompt prefixes

If every call starts with the same long instructions, schema or reference document, cache it. Per Anthropic’s prompt caching docs, cache reads cost 10% of the base input price (with some per-model exceptions listed there), and writing a 5-minute cache entry costs 1.25x base input. So caching pays when the same prefix is reused within the cache window, and costs slightly more if it is not. OpenAI’s prompt caching guide describes a discount of up to 95% on cached input, depending on the model. Put stable content first and changing content last, or the prefix will not match.

4. Batch anything nobody is waiting for

Nightly enrichment, backfills, bulk classification and report drafts do not need an instant reply. Anthropic’s Message Batches API is priced at 50% of standard API prices, and most batches finish in under an hour. OpenAI’s Batch API gives a 50% discount versus synchronous APIs with separate, higher rate limits, and completes within 24 hours. Keep interactive work, such as a chat a person is typing into, on the normal API.

5. Trim context and output

• Send the fields the task needs, not the whole record or the whole thread.
• Summarise long history once and pass the summary on later calls.
• Ask for structured output with short fields instead of prose you then parse.
• Set a sensible maximum output length for each task type.
• Remove few-shot examples that no longer change the result, checked against your own test set.

6. Per-task automation pricing: Zapier, Make, n8n, or code

Two different billing units matter here. Zapier bills tasks: only successful actions count and trigger steps do not. From 1 December 2026, Next Gen Zaps and the Zapier SDK also use tasks, and each successful Next Gen run uses at least one task. Higher tiers lower the cost per task, and overage switches to pay-as-you-go unless you turn it off (see Zapier pricing). Check whether your plan will change on that date.

n8n Cloud bills per workflow execution, which is one full run however many steps it has. Its pricing page lists Starter at 20 EUR a month billed annually for 2.5K executions and Business at 667 EUR a month billed annually for 40K executions, with a 50% Startup discount on Business for companies under 20 employees. The Community Edition can be self-hosted from GitHub.

Count runs and steps per workflow for a month, then price the same workload under each unit. That tells you more than any headline price. We do not quote a Make figure here because we have not verified one.

When not to migrate

• The bill is small. A migration has a build cost and a maintenance cost, and a small saving may not cover either.
• The workflow is brittle or poorly documented and nobody wants to touch it. Document it first.
• Nobody on your team will own the replacement. Self-hosted tools need updates and monitoring.
• The workflow relies on a connector that only the current tool provides well.
• Volume is falling or the process is about to change. Wait until it settles.

Often the right move is partial: keep the long tail of small zaps where they are, and move the two or three busiest workflows.

A one-week plan

1. Day 1: export last month’s LLM and automation invoices and list every workflow behind them.
2. Day 2: add per-task logging to the top three.
3. Day 3 to 4: try routing, caching and trimming on those three, measured against your test examples.
4. Day 5: move any no-rush job to a batch API, then rewrite the 2027 line items from measured cost per task.

Questions

Related guides

• A 2027 AI plan for agencies
• AI agent for holiday customer support
• EU AI Act disclosure checklist