Small language models are AI models with relatively few parameters, typically from under one billion to around ten billion, built to run fast and cheaply, sometimes on a laptop or phone. In marketing operations they save the most time on narrow, repeated jobs: tagging content, extracting fields, routing enquiries and cleaning data, where a large frontier model would be slower and costlier than needed.
What counts as a small language model?
There is no official cut-off, but the term usually covers models such as Microsoft’s Phi family, Google’s Gemma, Meta’s smaller Llama variants, Mistral’s compact models, Qwen’s smaller sizes and Apple’s on-device models. Many are open-weight, so you can download and run them yourself under their licence terms.
Their strengths are speed, low cost and privacy. Running a small model on your own server or machine means data never leaves your control. Their weakness is breadth: they know less, reason less well over long chains, and make more mistakes on open-ended creative tasks than the largest models.
The trick is to give them jobs that suit them. A workflow with clear inputs, clear outputs and checkable results plays to their strengths.
Which marketing jobs fit small models?
Good candidates share three traits: high volume, simple decisions and outputs you can verify automatically. Examples:
- Classifying blog posts or products into a fixed list of categories.
- Pulling company name, job title and intent from inbound form messages.
- Sentiment tagging of reviews or social comments.
- Detecting language and routing tickets to the right team.
- Normalising messy UTM parameters or campaign names.
- Writing short alt text from existing product attributes.
Tasks that do not fit well include long strategy documents, subtle brand copy and anything needing up-to-date world knowledge.
Workflow 1: tagging a content library
Large sites often have thousands of posts with missing or inconsistent tags. A small model can propose tags from a fixed list.
- Export titles and the first 300 words of each post to a CSV.
- Write a prompt that lists the allowed tags and asks for up to three, returned as JSON.
- Run a sample of 100 posts and have a person check them.
- Tighten the prompt where errors repeat, for instance by adding a one-line definition per tag.
- Run the full set, then import tags as suggestions for an editor to approve in bulk.
Because the output is constrained to a known list, invalid answers are easy to catch with a simple script.
Workflow 2: lead form triage
Free-text fields in contact forms hide useful signals. A small model can read each message and return structured fields: request type, budget mentioned (yes or no), urgency and language.
Connect the form to an automation tool such as Make, n8n or Zapier, send the message to the model, and write results into the CRM as custom properties. Sales then filters by request type rather than reading every message. Keep the original message attached, and review a weekly sample to check that the model has not drifted.
Choosing where to run the model
- Local: tools like Ollama or LM Studio run open models on a workstation; good for testing and private data.
- Self-hosted server: vLLM or similar serving software on a cloud GPU; good for steady volume.
- Hosted API: several providers serve small models cheaply by token; easiest to start.
Workflow 3: data cleaning before reporting
Campaign names in ad accounts are often inconsistent: “Summer_Sale_TR”, “summer sale tr v2”, “SS-TR”. Before building a dashboard, a small model can map each raw name to a standard naming scheme you define, returning the platform, market, campaign theme and funnel stage.
Store the mapping table so the same raw name always maps the same way, and only send new names to the model. That keeps cost near zero and results consistent.
How do you keep quality acceptable?
Small models make more errors, so every workflow needs a check. Useful controls:
- Constrain output to JSON with fixed fields and allowed values.
- Validate every response with a script and retry or flag failures.
- Keep a labelled test set of 50 to 200 examples and re-run it whenever you change model or prompt.
- Send low-confidence or invalid cases to a larger model or a person.
That last point describes a cascade: the small model handles the easy majority, and only hard cases go to a bigger, more expensive model. It is a common pattern for keeping costs under control.
How Alien Road applies small language models
In a typical engagement, discovery looks for the repetitive, rule-like jobs that eat staff hours: tagging, form triage, naming clean-up and similar. We estimate monthly volume and how errors would be caught, because a task without a check is not ready for automation.
Setup usually starts with a hosted small model for speed, moving to a local or self-hosted one if data sensitivity or volume justifies it. Prompts return structured output, and a validation script sits between the model and any system it writes to.
Measurement uses a fixed test set scored before launch and re-scored monthly, plus a log of how many items fall back to a person or a larger model. Reporting to the client shows hours saved on the task, error rates from spot checks and running cost, so the case for keeping or dropping each workflow is clear.
How much do small models cost compared with large ones?
Prices change too often to quote reliably, but the pattern is stable: per token, small hosted models typically cost a fraction of the flagship models from the same providers. Self-hosting shifts the cost from tokens to hardware, which pays off when volume is steady and high.
For a classification job running on tens of thousands of short items a month, the difference between a small and a large model can decide whether the workflow is worth automating at all. Run the numbers with your real volumes and current price pages before choosing.
Speed is the other saving. Small models answer quickly, which matters in workflows like live form triage where a person or system waits for the result.
Where to start this week
Choose one repetitive classification task in your marketing operations. Collect 100 real examples and label them by hand; that becomes your test set.
Install a local runner such as Ollama, try two small open models on the examples with the same prompt, and count correct answers. If one reaches a level you would accept from a junior assistant, connect it to a small pilot with human review, and track how often reviewers change its answers.
Keep reading: AI Optimization · AI Optimization