Generate Content

1. The Problem

Assessment tells you what is missing. But knowing the gap exists does not close it.

You run an assessment against IPA gateway criteria, Green Book requirements, or your organisation's own framework. The platform finds 14 RED and AMBER findings. Each one is specific: "No sensitivity analysis found." "Escalation procedures not defined." "Benefits not quantified."

Now what?

Today, you write the missing content from scratch. You open a blank document, look up the expected format, dig through your business case for the relevant figures, and spend two or three hours producing a sensitivity analysis table that should have been in the Financial Case from the start. Multiply that across 14 findings and you are looking at a week of remediation work before the next gateway.

Generate Content produces an 80% complete first draft in minutes. It assembles context from the criterion definition, the evidence requirements, and the data already in your uploaded documents. You get a structured draft with real figures pulled from your business case, citations back to source pages, and clear markers where you need to add information the AI could not find.

You review it, fill in the gaps, and download as Word. The platform already knows what good looks like for each criterion. Generate Content uses that knowledge to write the first draft so you can focus on the 20% that requires human judgement.

2. How It Works

Five steps from assessment finding to downloadable first draft.

1. Assessment identifies a gap. A RED or AMBER finding flags a specific deficiency — for example, "No sensitivity analysis has been identified in the uploaded documents." The finding includes the criterion reference, evidence requirements, and a recommended action.
2. You click "Generate Draft Content." The button appears on qualifying findings — those where the platform knows the expected structure and can produce useful output. You select your output format (structured table, narrative section, or full appendix) and confirm which uploaded documents to draw from.
3. The AI assembles context. The system pulls together the criterion definition, the evidence requirements, the assessment finding, relevant chunks from your uploaded documents, and the framework-specific template for that content type. This is targeted retrieval — not your entire document set, but the specific sections that contain relevant data.
4. A first draft appears. Generated content streams into a preview panel within 5–15 seconds. Data values cite their source document and page number. Fields where data was not found are marked [ACTION REQUIRED] with an explanation of what you need to provide. Each section carries a confidence rating.
5. You review, edit, and download. Check the citations against your documents. Fill in the placeholders. Verify any yellow-highlighted values where the AI flagged uncertainty. Download as Word, incorporate into your documents, and re-upload for the next assessment run.

3. What Can Be Generated

Not every gap can be addressed with generated content. The system knows the difference.

The key distinction: can the AI produce the structure and populate it with data from your documents, or does it need information that only you possess? Each criterion in the assessment framework has a suitability rating that determines whether the Generate Content button appears and what kind of output you get.

Suitability Tiers

4. Example: Sensitivity Analysis

A Green Book business case is assessed. The sensitivity analysis is missing. Here is what Generate Content produces.

The AI assembles context from the Financial Case (capital costs, operating costs, discount rate), Economic Case (benefit estimates, BCR), and HMT Green Book sensitivity analysis template. The generated output:

5. Example: Escalation Procedure

A governance framework gap triggers a generated escalation procedure grounded in the project's terms of reference.

The AI draws on the Terms of Reference (governance roles), the Risk Management Strategy (tolerance references), and the project value (£340M) to produce a structured procedure:

6. Example: Benefits Quantification

Four of six benefits are not monetised. Generate Content produces a structured table mixing extracted data with clear placeholders.

Notice the pattern: cells with extracted data carry teal citations. Cells requiring your input carry red ACTION REQUIRED markers. The methodology column recommends the appropriate HMT valuation approach even where the value itself must come from you.

7. Understanding Confidence Scores

Every piece of generated content tells you how much it trusts its own output. Here is how to read the signals.

Section-Level Confidence

Each generated section carries an overall confidence rating based on how much source data was available and how reliably it could be extracted.

Field-Level Highlights

Even within a High-confidence section, individual values can carry uncertainty. These are highlighted in yellow in the generated output. A yellow highlight means the AI found something relevant but is not certain it extracted the right figure.

Common causes of yellow highlights:

• Value appears in a table that was OCR-processed from a scanned PDF — possible digit errors
• Multiple conflicting values found across documents (e.g., different cost estimates in different versions)
• Value was inferred from narrative text rather than extracted from a structured table
• Unit or currency not explicitly stated in the source — the AI assumed £ sterling

8. Provenance and Auditability

Every generation is fully traceable. If content is questioned during an audit, the full chain is reconstructable.

Programme documentation goes through gateway reviews, NAO scrutiny, and public accountability processes. AI-generated content in this context requires a complete audit trail. Programme Insights records every step of the generation process.

What Gets Recorded

Why This Matters

If a reviewer or auditor asks "where did this sensitivity analysis come from?", the answer is fully documented:

• The finding that triggered it (criterion G3-4.7, rated RED, from assessment run #47)
• The source data used (Financial Case pp. 23, 24, 31; Economic Appraisal p. 18)
• The AI model and exact prompt that produced it
• The raw AI output and the human-edited final version
• Who generated it, when, and what edits were made

This is not an afterthought. The provenance system is designed for environments where generated content may be scrutinised by the IPA, the NAO, or a parliamentary committee.

9. The Feedback Loop

Generated content improves over time because the system learns from your edits.

When you edit a generated draft and save it, the platform captures the difference between the AI's version and yours. This is not abstract machine learning — it is a concrete record of what you changed and why:

• Structural changes — did you reorder sections, add rows to a table, or remove a column? This tells the system the template needs adjusting for your context.
• Value corrections — did you change a figure the AI extracted? This flags a potential extraction error for that document format.
• Tone and style — did you rewrite narrative sections in a different register? This teaches the system your organisation's preferred language.
• Placeholder resolution — what did you fill in for each ACTION REQUIRED marker? Over time, this builds a picture of what data your organisation typically provides.

What This Means in Practice

If your organisation consistently changes "5 working days" to "3 working days" for Level 1 escalation response times, future generations for your projects will default to 3 working days. If you always add a "Dependencies" column to RACI matrices, future RACI generations will include it.

The learning is scoped to your organisation. Your edits improve your future generations. They do not affect other users.

10. Getting Started

From assessment findings to first draft in seven steps.

1. Run an assessment on your project documents. Upload your business case, governance documents, and supporting material. Select the appropriate module (IPA Gateway, Green Book, NEC, or your custom framework) and run the assessment.
2. Review the RED and AMBER findings. Each finding tells you exactly what is missing and what the framework expects. Focus on the criteria rated RED first — these are the gaps that will fail a gateway review.
3. Look for the "Generate Draft Content" button. It appears on qualifying findings where the platform can produce useful output. If the button is not there, the gap requires data that only you can provide — no amount of AI will help.
4. Click it and choose your output format. Select structured table, narrative section, or full appendix depending on what the criterion requires. Confirm which uploaded documents to draw from — the platform pre-selects the most relevant ones.
5. Review the draft. Check the teal citations against your source documents. Fill in every [ACTION REQUIRED] placeholder. Verify any yellow-highlighted values where the AI flagged uncertainty.
6. Download as Word and incorporate into your project documents. Edit as needed — this is a first draft, not a final submission. Your domain expertise turns the 80% draft into a 100% deliverable.
7. Re-upload the updated document. The platform re-assesses automatically. If the generated content closes the gap, the finding moves from RED or AMBER to GREEN. You have a closed loop: assess, generate, fix, verify.