How to Use StackAI to Check Ad Copy Against Brand Rules

The short answer: StackAI's free plan, with one seat, two projects and 500 runs a month (pricing page, checked on 24 October 2026), is enough to build a checker that reads draft ad copy, compares it with your brand rules and approved claims, and returns a verdict with the rule it cites: put the rules in a knowledge base, build an agent grounded in it with citations on, upgrade it to a workflow so the answer comes back as fixed fields, run ad variants through a batch table, and have a person decide on every flag.

Disclosure: I may earn a commission through some links on this site. It does not change what this page says.

The checker does not write copy and does not publish anything. It reads, compares and points. That narrowness is what makes it safe to run on every variant, and it is the same job a brand agent does in review mode, which how to build a brand agent treats from the rules side. For the platform itself, see the StackAI glossary entry.

Size it before you build:

1. Runs. StackAI's pricing FAQ defines a run as one execution of a workflow, and each batch row is its own set of workflow inputs (batch interface guide). For example, 30 ad variants and 10 emails a week come to (30 + 10) x 4 = 160 runs a month, under a third of the free 500.
2. Inputs you already hold. The brand book, the list of claims legal has approved, the disclaimers each channel needs, and 20 pieces of past copy whose verdict a senior reviewer has already given.
3. The measure. The share of the checker's flags your reviewer agrees with, tested against known answers in Step 9. A checker whose flags are mostly overruled adds review time instead of saving it.

What you need first

• A StackAI account on the free plan, which includes community support on Discord.
• Permission from whoever owns AI in your company. StackAI is built for IT teams to govern AI across an organisation, and the docs assume an administrator who approves models and access.
• Your brand rules written as rules, each with an example of a breach.

Step 1: Build the knowledge base, settings first

Open Knowledge Bases and select New Knowledge Base. Before uploading anything, open More Options, Update Settings and check chunking, OCR and the embedding model: StackAI's knowledge base guide says these become read-only once the base holds files. Turn on OCR if any rule lives in a scanned PDF. Upload, then wait for each document to show Indexed, because upload completion alone does not make it searchable. Open a document's Chunks view to see the passages retrieval will use, and run a question in the Evaluate tab to check that the right rule comes back.

Step 2: Create the agent

In Projects, select New Project, then Create New Agent under Quick Start, as the first agent guide describes. Choose a model your organisation has approved. Write instructions that state the job and the limits: check the supplied copy against the brand rules and approved claims; quote the rule each problem breaks; never approve a claim that is not on the approved list; if the rules do not cover something, say so. The guide's own template line is worth keeping: do not invent facts or claim to have searched a source that is not connected.

Step 3: Ground it in the rules, with citations

In Agent Builder, select Add Knowledge Sources, Add Knowledge Base and choose yours, then enable Use citations, per the grounding guide. Citations matter more here than anywhere: a flag without the rule behind it is an opinion, and the reviewer needs to open the source to judge it.

Step 4: Test it in Draft Preview

Paste three pieces of copy: one clean, one with a claim that is not on the approved list, one with a tone problem. Then ask something the rules do not cover. The grounding guide expects the agent to say what is missing rather than guess. If it guesses, tighten the instructions before going further.

Step 5: Upgrade to a workflow

A workflow lets you fix the shape of the answer and run many inputs at once. Export a copy of the project first: the upgrade guide warns that conversion cannot be undone and removes the Agent Builder view. Then select Upgrade to Workflow, review the input, knowledge and output nodes on the canvas, and run one input.

Step 6: Return fixed fields, not prose

In the model node's settings, set Answer Format to JSON Schema, per the structured outputs guide. Three fields are enough: a verdict limited to "pass" or "revise", a list of breaches, and the rule cited for each. The guide makes a point to keep in the instructions: a required field does not justify inventing a value, so allow an empty list when nothing is wrong.

Step 7: Run the variants as a batch

Choose the Batch interface and map the copy column to the input and the three fields to output columns. The batch guide says to start with one representative row, check its mapped values, then import the rest from your spreadsheet, keeping the source file so failed rows can be retried. Select Run Batch and read the status of each row: rows can fail on their own while others complete. For a single draft, a Form interface does the same job one submission at a time.

Step 8: Keep the decision with a person

The checker proposes; the reviewer decides. Every row marked "revise" goes to one named person, who reads the cited rule, accepts or overrules the flag, and notes overrules in the sheet. If you later add a step that sends or publishes, StackAI's human input step can pause the run for a response, and its own documentation warns that showing a question is not the same as receiving approval, and that an unattended schedule cannot supply a human answer.

Step 9: Measure against known answers

Open Evals, then Datasets, and build a dataset from the 20 past pieces with their reviewed verdicts, per the evaluation dataset guide. Run it after every change to the rules or instructions. The guide's warning applies to the people too: an incorrect reference answer makes a correct workflow fail.

About the free plan

The free plan is sized to prove the idea to one person, not to run it for a team: one seat and two projects. The next step on the pricing page is a custom Enterprise quote, so the first number you see for team use comes from StackAI's sales team. Arrive with your monthly run count and seat list, and the conversation starts from your usage rather than theirs.

What StackAI's docs warn about

• Settings you cannot change later. Chunking, OCR and embedding settings lock once a knowledge base holds files.
• A one-way upgrade. Converting an agent to a workflow cannot be undone; export first.
• Failed rows hidden in a good batch. Each row needs its own check before you export results.
• Spend that passes the limit. Credit policies gate new runs, and a run's final cost can still take usage past the remaining allowance, per the usage and limits guide.

FAQ

Can it check a PDF or an image, not just text?

The form interface accepts a Files input, and the agent then needs a prompt reference and file processing set up, per the form guide.

Does the Asana acquisition change anything?

StackAI's announcement says no: it continues as its own product and brand, workflows stay as they are, and nothing has to connect to Asana.

Can the whole marketing team use the free plan?

No. It includes one seat. Seats, projects and runs beyond that are set in a custom Enterprise plan (pricing page).