New on the blog: why AI belongs in the feed →

AI Application Review in Gravity Flow: Score with AI Feed, Decide with Multi Path Step

Grant committees, scholarship panels and vendor programs all have the same bottleneck: someone has to read every application before anyone can decide. This setup gives the first read to AI Feed and keeps the decision with a person. Claude scores each application against your rubric and lists what’s wrong with it. A reviewer then picks one of five paths in Gravity Flow: Fund, Revise, Decline, Defer or Escalate.

What happens to one application

  1. The applicant submits the form. Organization, project, amount, budget line items, what the grant funds, how they’ll know it worked, and an email.
  2. AI review (workflow step 1). The step runs an AI Feed text feed. Claude scores the application from 1 to 10, writes a two-sentence summary, lists each concern with the field it came from, and suggests a path. All four answers land in admin-only fields on the entry.
  3. Committee decision (workflow step 2). A Multi Path Step assigned to your reviewers. They see the application and the AI’s notes, then choose Fund, Revise, Decline, Defer or Escalate. Each path sets its own workflow status and can send the entry to its own next step.

What it caught in our test

We sent an invented application through the demo on this site: a youth rowing club asking for $18,500, with budget lines that add up to $17,700. The AI review scored it 5, flagged the $800 gap and named the two fields involved, pointed out that the plan had no schedule or coach details, and suggested Revise. The entry then waited at Committee decision for a person to choose. The AI never moved it forward on its own.

Set it up

You need Gravity Forms, Gravity Flow, GravityOps AI Feed with your own Anthropic API key, and GravityOps Multi Path Step.

  1. Import the form. Download ai-application-review-form.json and import it under Forms → Import/Export. It brings eleven fields (seven for the applicant, plus AI score, AI summary, AI concerns and AI suggested path with Visibility set to Administrative, so applicants never see them) and both Gravity Flow steps. The AI Feed feed doesn’t come with the import: its settings are below, and all three feeds are in ai-application-review-feeds.json for reference.
  2. Add the AI Feed text feed. Provider Anthropic, a Claude model of your choice (we used claude-sonnet-5-5 at low effort), and JSON output mapped to fields: score → AI score, summary → AI summary, concerns → AI concerns, suggested_path → AI suggested path. Paste the two prompts below.
  3. Connect the workflow. Open the imported AI review step and select your new text feed. Then assign the Committee decision step to your reviewers. Its five paths (Fund, Revise, Decline, Defer, Escalate) come with the import.
  4. Send each path somewhere useful (optional). For example, point Revise at a user input step assigned to the applicant’s email so they can fix their budget, and Escalate at an approval step for the committee chair.

System prompt

You prepare grant applications for a human review committee. You score; you never decide.
Score 1-10 against four things, equally weighted: a clear need, a budget that adds up (the line items must sum to the amount requested), a believable plan, and a way to know it worked.
Use only what the application says. Do not reward polish or length. Name the field each concern comes from.
Suggest one path: fund, revise, decline, defer or escalate. Suggest escalate when the amount is over 25000 or anything looks like a conflict of interest.

User prompt

Organization: {Organization:1}
Project: {Project title:2}
Amount requested: {Amount requested:3}
Budget line items:
{Budget line items:4}
What the grant will fund:
{What will the grant fund?:5}
How they will know it worked:
{How will you know it worked?:6}

Before you put it on a client’s site

  • The score is a first read, not a verdict. Rewrite the rubric in the system prompt to match the client’s criteria, and test it on past applications before it sees real ones.
  • Every application is one API call on the client’s own key. AI Feed sends it straight from the site to Anthropic and records the model and tokens of each run in its usage Ledger.
  • Only send what the rubric needs. The prompt above leaves out the contact email on purpose.