Skip to main content

Alpha This is a new service. Help us improve it and give your feedback by email.

Step 4: Review Extracted Fields - Planning AI Walkthrough

Review the AI-extracted planning application fields and verify accuracy

Walkthrough progress

Step 4 of 4 • 3 minutes

Step 4 3 minutes

Review Extracted Fields

Examine the AI-extracted fields and see how they match the original planning application document.

Find your Planning AI application URL in CloudFormation Outputs
The application landing page with document upload area and sample document option

Expected outcome

  • AI-generated committee summary is displayed prominently
  • Application completeness assessment shows status
  • Classification identifies application type with confidence level
  • Extracted fields match original document content

What you should see

After processing completes, you'll see a results section with extracted fields. The exact fields depend on which sample document you uploaded:

AI analysis results

The most prominent section shows three AI-generated outputs from Amazon Bedrock:

Committee Summary
A 2-3 sentence summary of the application suitable for planning officer reports and committee papers. This is the hero output — replacing 20-30 minutes of manual writing per application.
Completeness Check
Shows whether the application appears complete or incomplete. If incomplete, lists the missing required fields — saving back-and-forth with applicants.
Application Classification
Identifies the application type (Householder, Full, Outline, etc.) with a confidence level. Enables intelligent routing to the right planning team.

Understanding the results

Each extracted field includes:

Field Name
The type of information extracted (e.g., "Applicant Name", "Site Address"). These map to standard planning system fields.
Extracted Value
The actual text extracted from the PDF. This is what would be automatically populated in your planning database.
Confidence Score
AI's confidence level (0-100%). Scores above 95% are highly reliable. Lower scores may need human review.

Verify accuracy

Compare the extracted fields against the original PDF document:

  1. Open the sample PDF you downloaded

    Find it in your downloads folder and open it in a PDF reader

  2. Locate the applicant name in the PDF

    Usually on page 1, in a section like "Applicant Details" or "Personal Information"

  3. Compare with extracted "Applicant Name" field

    Does it match exactly? Notice the AI found this even if it wasn't labeled clearly.

  4. Check the address extraction

    Look for "Site Address" or "Application Site" in the PDF. The AI should have captured the full address including postcode.

  5. Review the development description

    Find the "Proposed Development" or "Description of Works" section. The AI extracts the key summary, not just copying entire paragraphs.

What's impressive: The AI parsed a 17-page document containing legitimate planning content and deliberately irrelevant material (Victoria Sponge recipes, holiday packing lists, 5G conspiracy theories, and a seven-year-old's crayon drawing) — and produced a coherent committee-ready summary focusing only on the material planning considerations. It knew the difference between an applicant name and an agent name, understood UK address formats, and correctly classified the application type despite pages of noise.

Key observations

Notice these important capabilities:

Format independence

AI extracted fields regardless of where they appeared on the page. No template matching required - it works with any planning application format.

Accuracy on typed text

97-99% confidence scores on typed PDF documents. Better than human transcription accuracy, with no typos or misread characters.

UK-specific recognition

Correctly identified postcodes, understood "Land adjacent to..." addresses, and recognized planning terminology like "Householder" and "Change of Use".

Speed vs accuracy tradeoff

8-20 seconds to extract 5-6 fields vs 45 minutes manual entry. That's 135-270× faster with higher accuracy.

Real-world integration

In a production system, these extracted fields would:

  • Auto-generate committee report drafts: AI summaries flow directly into committee papers, saving 20-30 minutes per application
  • Flag incomplete applications: Completeness checks identify missing fields before officer review, reducing back-and-forth with applicants
  • Auto-populate planning database: Extracted values flow directly into your planning system's application record
  • Intelligent routing: Application classification automatically assigns cases to the right planning team (Householder, Full, Listed Building, etc.)
  • Link to case management: Automatically create case files, assign to planning officers, generate acknowledgment letters
  • Audit trail: Original PDF stored with extracted metadata and AI analysis for compliance and appeals

Calculate your time savings

Based on what you just saw, estimate your council's savings:

Troubleshooting

Extracted fields don't match document

If extracted values are incorrect or incomplete:

  • Check the source document quality - blurry scans reduce accuracy
  • Ensure text is machine-readable, not handwritten
  • Verify document uses standard planning application format
  • Look at confidence scores - <95% indicates uncertainty
  • Try re-uploading with a higher quality scan
  • Check if PDF contains actual text or is just images

Our sample documents should extract with 97-99% accuracy. If not, there may be a service configuration issue.

Some fields missing entirely

If expected fields aren't extracted at all:

  • The field may not exist in this document type (e.g., outline applications lack detailed descriptions)
  • Check if field is labeled clearly in the PDF - ambiguous labels confuse AI
  • Some fields may need custom extraction rules beyond Textract's default capability
  • Verify the extraction Lambda function includes logic for all required fields
  • Check CloudWatch logs to see what Textract returned
Confidence scores very low (<90%)

If confidence scores are unexpectedly low:

  • Low-quality scans or poor image resolution
  • Unusual fonts or formatting in the PDF
  • Handwritten text (Textract OCR works best on typed text)
  • Document skewed or rotated (Textract auto-corrects but this reduces confidence)
  • Field values in unexpected document locations

In production, you'd set a confidence threshold (e.g., 95%) and flag lower scores for human review.

Need help?

If you're stuck, check the troubleshooting section or contact us.