AI Data Analysis & Digital Workflows

Author

Resource Hub

Date

08 Sept 2025

Between August 2025 and today, I was subcontracted to help the Local Government White Paper team ingest and synthesise 270 public submissions (received by and shortly after 31 July 2025) into a single evidence base and a 100-page “Integrated Analysis of Public Submissions” for specialist review.

The deliverables include a normalised database, quantified reports across 11 thematic areas and 5 research areas, and cross-country insights to inform options for the revised White Paper due by March 2026.

The punchline: by combining AI data analysis (for fast, consistent text classification) with digital workflows (for structured review and audit trails), we cut manual collation time dramatically, improved traceability, and gave the Reference Groups a consolidated, queryable view of the public’s proposals and solutions—right when the policy window is open.

Related keywords we’ll cover

AI in government; policy text mining; GovTech; evidence-based policymaking; participatory governance; POPIA compliance; PAIA requests; coalition governance; municipal finance; audit outcomes; digital workflow automation; data cleaning; prompt engineering; reproducible policy analysis; intergovernmental relations; service delivery dashboards.

Key Takeaways
A simple schema + regex + AI beats manual collation for speed, consistency and traceability.
AI data analysis accelerates classification; human review secures legitimacy and accuracy.
Digital workflows create an auditable path from submission to policy option.
Quantification (counts and rankings) reveals consensus and contention faster.
Compliance-by-design (PAIA/POPIA gates) avoids late surprises.
Aligning with GovTech/OECD guidance helps benchmark quality and risk posture.

AI Data Analysis & Digital Workflows for SA White Paper on Local Government

Project context and goals

Mandate. Support report writers and local government experts to turn raw submissions from civilians, political parties, government entities, NPOs, and the private sector into a coherent dataset and decision-ready analysis. Outputs: (1) a normalised database, (2) 11 thematic and 5 research-area reports reflecting public perspectives, (3) quantified counts of problems/proposals/solutions, and (4) a 100-page integrated synthesis for specialist review and policy optioning.

White Paper for Local Government timetable. The revised White Paper aims to set principles, propose policy shifts with evidence, and guide the transition roadmap—finalisation by March 2026.

Problem we solved. Disparate PDFs and emails, uneven formats, and overlapping issues made manual synthesis slow and opaque. Our approach created a single source of truth and reproducible analysis steps, letting subject-matter teams focus on first- and second-order policy choices rather than data wrangling.

Method stack (AI + human-in-the-loop)

Core tools. Google Drive apps (Sheets, Docs) for the shared data model and traceable edits; ChatGPT-5 and Gemini 2.5 for assisted classification and summarisation; light prompt engineering for consistency; and versioned templates for the 11 thematic reports and the integrated synthesis.

Why this stack. It’s accessible to government teams, keeps the audit trail in familiar systems, and aligns with GovTech guidance to strengthen core platforms and transparent workflows rather than bespoke, brittle tools. See the World Bank’s GovTech Maturity Index for how platforms and shared services underpin capability gains (global GTMI average improved from 0.519 (2020) to 0.552 (2022)).

Guardrails. We embedded privacy and transparency requirements into the workflow—open-by-default where lawful (PAIA), protective where required (POPIA). The Information Regulator’s guidance and PAIA resources informed the templates for access and redaction.

Data architecture: from messy inputs to a tidy policy dataset

Normalisation. Every submission was split into atomic “claims,” each mapped to: Thematic Area, Problem, Proposal, Solution, and evidence snippet. This mirrors how the integrated report organises insights and cross-cutting shifts (e.g., professionalisation, IGR, digital modernisation).

Schema highlights.

source_id (submission), claim_id (unique key), theme (1–11), type (problem/proposal/solution), keywords (tag set), quant_count (if quantifiable), quote (short extract).
Status flags: needs_review, escalate_reference_group, confidential_redaction.

Benefits. This structure supports AI data analysis (reliable tagging), digital workflows (assign, review, approve), and reproducible exports (by Theme/Research Area).

Smart tagging with Sheets + regex (repeatable and transparent)

We used plain-English keyword libraries per theme (expandable as reviewers discovered synonyms) and regex helpers to reduce false positives. Three examples:

=IF(B2<>"","(?i)(" & TRIM(B2) & ")", "")

=TEXTJOIN(
  ", ", TRUE,
  FILTER(Themes!B$2:B,
    REGEXMATCH('Main Database'!A3, "(?i)"&Themes!C$2:C)
  )
)

=IF(Themes!L2="","",SPLIT(REGEXREPLACE(Themes!L2,",\s*",","),","))

Counting and ranking the most frequent tagged items across all breakdown sheets:

=ARRAYFORMULA(
  QUERY(
    TOCOL(
      TRIM(
        SPLIT(
          REGEXREPLACE(TEXTJOIN(",",TRUE,'Break Down'!B2:AM266),",\s*",","),
          ","
        )
      ),
    1),
    "select Col1, count(Col1) 
     where Col1 <> '' 
     group by Col1 
     order by count(Col1) desc 
     label count(Col1) ''",
    0
  )
)

This gave reviewers transparent, reproducible counts for “most cited problems” and “top proposed solutions” per theme.

AI-assisted classification (with a human brake)

Why AI here. Submissions varied widely in style and length. AI helped chunk, label, and summarise at scale, while reviewers validated edge cases and refined the taxonomy—classic human-in-the-loop.

Prompts and patterns. We templated prompts to (1) extract claims, (2) assign Theme + type (problem/proposal/solution), (3) list citations/paragraph refs, and (4) propose LSI tags (e.g., “coalition rules,” “Blue/Green Drop,” “participatory budgeting”).

Risk management. We adhered to OECD guidance that AI in the public sector should be used with clarity on benefits, risks, and accountability—not as a black box.

Quantification: turning submissions into counts and signals

For each theme, we quantified the frequency of problem statements and mapped proposed remedies. The integrated synthesis presents first-order (structural) and second-order (framework) shifts—e.g., functional assignment, coalition stability, professionalisation, and digital modernisation—to help policymakers see system levers rather than isolated fixes.

Example signals produced for reviewers

Governance & leadership: dominance of calls to professionalise leadership, depoliticise administration, and strengthen oversight (AGSA consequences, MPACs, lifestyle audits).
Administration & capacity: merit-based hiring, protected tenure, minimum competencies, and digitised workflows across HR/finance/assets/records.
Citizen relations: proactive transparency (PAIA), ward-level forums, participatory budgeting, and local ombuds.

These signals align with recent AGSA audit insights showing persistent weaknesses and regressions—underscoring why consequence management and capability building must advance together.

Reporting: 11 thematic areas & 5 research areas (how we packaged it)

We generated 11 thematic reports (Governance; Administration; Citizen relations; Structure; IGR; Finance; Infrastructure; LED; Spatial planning; Environment/Climate/Disaster; Traditional & Khoi-San) nested within Research Areas A–E, with a consistent four-part structure per theme: Executive Summary, Main Findings, Interpretation, Implications for White Paper drafting.

The 100-page integrated document then surfaces cross-cutting shifts (e.g., rules-based IGR, finance follows function, digital modernisation, citizen power & transparency) and immediate decision prompts for the White Paper team.

Comparative research: what other countries are doing (and why it helps)

To avoid “reinventing the wheel,” we triangulated proposals against international practice and GovTech guidance:

Platform-first governance. Countries moving up the GTMI scale invest in shared platforms (identity, payments, case management) and workflow digitisation that reduce compliance duplication—exactly the kind of gains local government needs for HR/SCM/records.
Responsible AI adoption. OECD’s 2024 work stresses transparent use cases, risk registers, and algorithmic accountability instruments—features we mirrored in our AI review sheets and decision logs.

These references framed the “what good looks like” checklist for our local context.

Digital workflows that kept everyone aligned

We built end-to-end digital workflows around four boards: Ingest → Tag → Review → Publish.

Key controls

Assignment & SLAs in Sheets/Docs with owner columns and due dates.
Redaction gates guided by PAIA (open-by-default where lawful) and POPIA (protect personal info).
Versioned exports: Theme PDFs and the integrated synthesis are generated from the same dataset, so a fix in one place updates everywhere.

Why this matters. Digital workflows reduce manual hand-offs, create an audit trail, and let leadership see progress at a glance—exactly the kind of consequence-friendly culture push AGSA keeps calling for.

What changed for reviewers (before vs after)

Before: Dozens of PDFs; inconsistent counts; late-stage reconciliation; limited traceability.
After: One database; consistent tags; filterable by Theme/Research Area; quantified; citations to the original text; exportable to report templates.

Policy value. This allowed Reference Groups to focus on first-order questions (e.g., differentiated municipal model, IGR compacts) and second-order frameworks (e.g., professionalisation rules, digital standards) surfaced across themes.

Lessons learned (practical notes for SA policy teams)

Start with a shared schema (what is a “claim”?) before turning on any AI.
Find-and-replace your way to consistency (regex + validation lists beat free-text every time).
Quantify early. Publish interim counts—even if imperfect—to expose taxonomy gaps.
Make privacy the default (label fields; build redaction steps into the flow).
Tie reporting to decisions. Frame outputs as decision prompts (now/next/later), not just summaries.

How to replicate this in your organisation (mini-playbook)

Set up (week 1–2)

Build a claims table with the fields listed above.
Draft Theme keyword lists and teach the team how to add synonyms.
Stand up the four-board workflow (Ingest → Tag → Review → Publish).

Process (rolling)

Batch submissions; run regex-aided tagging; have AI propose tags; reviewers confirm.
Use TEXTJOIN/FILTER/QUERY patterns (above) to auto-generate top terms and counts by theme.
Export to standardised H2/H3 report templates with Executive Summary → Findings → Interpretation → Implications.

Governance

Appoint an evidence custodian (sign-off authority).
Publish open data where lawful (and redacted where required) to boost public trust and enable external review.

Conclusion

The White Paper team asked for a coherent, quantified picture of what South Africans said—and a way to act on it. By combining AI data analysis with digital workflows, we built a transparent pipeline from raw text to decision prompts that reflect where the strongest public signals are. The 100-page integrated analysis is now a living evidence base: it updates as reviewers refine tags or add late submissions, and it exports cleanly into theme reports and policy option packs. That lets Reference Groups concentrate on weighing first-order structure (e.g., differentiated municipal models, rules-based IGR) and second-order frameworks (professionalisation, performance and data systems) against trade-offs and feasibility.

Most importantly, this approach fits our context: it’s affordable, uses familiar tools, respects POPIA/PAIA, and is grounded in public-sector AI guidance. As the process moves toward March 2026, the combination of clean data, clear workflows, and credible sources is what will keep the drafting focused on outcomes that communities can feel.

AI-Readiness & Workflow Audit
Automate repetitive work. Save hours weekly. Modernise ops.
For teams swamped by admin/reporting/content tasks
Get: 3–5 high-impact automations + live setup of your first workflow
Outcome: More capacity today, scalable foundation for tomorrow
Book this Audit

Frequently Asked Questions

1) How did you ensure POPIA/PAIA compliance while using AI?

Data minimisation, redaction gates before publishing, and a register of lawful bases for processing. We followed the Information Regulator’s PAIA guide and POPIA guidance notes.

2) What counts did you publish?

Per theme: most-cited problems; top proposals/solutions; and cross-theme roll-ups—grounded in the 265-submission dataset documented in the integrated analysis.

3) How do AI outputs get verified?

Every auto-tag is reviewed by a theme lead; disagreements are logged; prompts evolve. This mirrors OECD advice on accountable AI in the public sector.

4) Which global frameworks guided the digital workflow design?

World Bank GovTech materials (shared platforms, service digitisation) and OECD public-sector AI guidance.

Author

Romanos Boraine

Resource Hub

Case Studies

Date

08 Sept 2025

Between August 2025 and today, I was subcontracted to help the Local Government White Paper team ingest and synthesise 270 public submissions (received by and shortly after 31 July 2025) into a single evidence base and a 100-page “Integrated Analysis of Public Submissions” for specialist review.

The deliverables include a normalised database, quantified reports across 11 thematic areas and 5 research areas, and cross-country insights to inform options for the revised White Paper due by March 2026.

The punchline: by combining AI data analysis (for fast, consistent text classification) with digital workflows (for structured review and audit trails), we cut manual collation time dramatically, improved traceability, and gave the Reference Groups a consolidated, queryable view of the public’s proposals and solutions—right when the policy window is open.

Related keywords we’ll cover

AI in government; policy text mining; GovTech; evidence-based policymaking; participatory governance; POPIA compliance; PAIA requests; coalition governance; municipal finance; audit outcomes; digital workflow automation; data cleaning; prompt engineering; reproducible policy analysis; intergovernmental relations; service delivery dashboards.

Key Takeaways
A simple schema + regex + AI beats manual collation for speed, consistency and traceability.
AI data analysis accelerates classification; human review secures legitimacy and accuracy.
Digital workflows create an auditable path from submission to policy option.
Quantification (counts and rankings) reveals consensus and contention faster.
Compliance-by-design (PAIA/POPIA gates) avoids late surprises.
Aligning with GovTech/OECD guidance helps benchmark quality and risk posture.

Project context and goals

Mandate. Support report writers and local government experts to turn raw submissions from civilians, political parties, government entities, NPOs, and the private sector into a coherent dataset and decision-ready analysis. Outputs: (1) a normalised database, (2) 11 thematic and 5 research-area reports reflecting public perspectives, (3) quantified counts of problems/proposals/solutions, and (4) a 100-page integrated synthesis for specialist review and policy optioning.

White Paper for Local Government timetable. The revised White Paper aims to set principles, propose policy shifts with evidence, and guide the transition roadmap—finalisation by March 2026.

Problem we solved. Disparate PDFs and emails, uneven formats, and overlapping issues made manual synthesis slow and opaque. Our approach created a single source of truth and reproducible analysis steps, letting subject-matter teams focus on first- and second-order policy choices rather than data wrangling.

Method stack (AI + human-in-the-loop)

Core tools. Google Drive apps (Sheets, Docs) for the shared data model and traceable edits; ChatGPT-5 and Gemini 2.5 for assisted classification and summarisation; light prompt engineering for consistency; and versioned templates for the 11 thematic reports and the integrated synthesis.

Why this stack. It’s accessible to government teams, keeps the audit trail in familiar systems, and aligns with GovTech guidance to strengthen core platforms and transparent workflows rather than bespoke, brittle tools. See the World Bank’s GovTech Maturity Index for how platforms and shared services underpin capability gains (global GTMI average improved from 0.519 (2020) to 0.552 (2022)).

Guardrails. We embedded privacy and transparency requirements into the workflow—open-by-default where lawful (PAIA), protective where required (POPIA). The Information Regulator’s guidance and PAIA resources informed the templates for access and redaction.

Data architecture: from messy inputs to a tidy policy dataset

Normalisation. Every submission was split into atomic “claims,” each mapped to: Thematic Area, Problem, Proposal, Solution, and evidence snippet. This mirrors how the integrated report organises insights and cross-cutting shifts (e.g., professionalisation, IGR, digital modernisation).

Schema highlights.

source_id (submission), claim_id (unique key), theme (1–11), type (problem/proposal/solution), keywords (tag set), quant_count (if quantifiable), quote (short extract).
Status flags: needs_review, escalate_reference_group, confidential_redaction.

Benefits. This structure supports AI data analysis (reliable tagging), digital workflows (assign, review, approve), and reproducible exports (by Theme/Research Area).

Smart tagging with Sheets + regex (repeatable and transparent)

We used plain-English keyword libraries per theme (expandable as reviewers discovered synonyms) and regex helpers to reduce false positives. Three examples:

=IF(B2<>"","(?i)(" & TRIM(B2) & ")", "")

=TEXTJOIN(
  ", ", TRUE,
  FILTER(Themes!B$2:B,
    REGEXMATCH('Main Database'!A3, "(?i)"&Themes!C$2:C)
  )
)

=IF(Themes!L2="","",SPLIT(REGEXREPLACE(Themes!L2,",\s*",","),","))

Counting and ranking the most frequent tagged items across all breakdown sheets:

=ARRAYFORMULA(
  QUERY(
    TOCOL(
      TRIM(
        SPLIT(
          REGEXREPLACE(TEXTJOIN(",",TRUE,'Break Down'!B2:AM266),",\s*",","),
          ","
        )
      ),
    1),
    "select Col1, count(Col1) 
     where Col1 <> '' 
     group by Col1 
     order by count(Col1) desc 
     label count(Col1) ''",
    0
  )
)

This gave reviewers transparent, reproducible counts for “most cited problems” and “top proposed solutions” per theme.

AI-assisted classification (with a human brake)

Why AI here. Submissions varied widely in style and length. AI helped chunk, label, and summarise at scale, while reviewers validated edge cases and refined the taxonomy—classic human-in-the-loop.

Prompts and patterns. We templated prompts to (1) extract claims, (2) assign Theme + type (problem/proposal/solution), (3) list citations/paragraph refs, and (4) propose LSI tags (e.g., “coalition rules,” “Blue/Green Drop,” “participatory budgeting”).

Risk management. We adhered to OECD guidance that AI in the public sector should be used with clarity on benefits, risks, and accountability—not as a black box.

Quantification: turning submissions into counts and signals

For each theme, we quantified the frequency of problem statements and mapped proposed remedies. The integrated synthesis presents first-order (structural) and second-order (framework) shifts—e.g., functional assignment, coalition stability, professionalisation, and digital modernisation—to help policymakers see system levers rather than isolated fixes.

Example signals produced for reviewers

Governance & leadership: dominance of calls to professionalise leadership, depoliticise administration, and strengthen oversight (AGSA consequences, MPACs, lifestyle audits).
Administration & capacity: merit-based hiring, protected tenure, minimum competencies, and digitised workflows across HR/finance/assets/records.
Citizen relations: proactive transparency (PAIA), ward-level forums, participatory budgeting, and local ombuds.

These signals align with recent AGSA audit insights showing persistent weaknesses and regressions—underscoring why consequence management and capability building must advance together.

Reporting: 11 thematic areas & 5 research areas (how we packaged it)

We generated 11 thematic reports (Governance; Administration; Citizen relations; Structure; IGR; Finance; Infrastructure; LED; Spatial planning; Environment/Climate/Disaster; Traditional & Khoi-San) nested within Research Areas A–E, with a consistent four-part structure per theme: Executive Summary, Main Findings, Interpretation, Implications for White Paper drafting.

The 100-page integrated document then surfaces cross-cutting shifts (e.g., rules-based IGR, finance follows function, digital modernisation, citizen power & transparency) and immediate decision prompts for the White Paper team.

Comparative research: what other countries are doing (and why it helps)

To avoid “reinventing the wheel,” we triangulated proposals against international practice and GovTech guidance:

Platform-first governance. Countries moving up the GTMI scale invest in shared platforms (identity, payments, case management) and workflow digitisation that reduce compliance duplication—exactly the kind of gains local government needs for HR/SCM/records.
Responsible AI adoption. OECD’s 2024 work stresses transparent use cases, risk registers, and algorithmic accountability instruments—features we mirrored in our AI review sheets and decision logs.

These references framed the “what good looks like” checklist for our local context.

Digital workflows that kept everyone aligned

We built end-to-end digital workflows around four boards: Ingest → Tag → Review → Publish.

Key controls

Assignment & SLAs in Sheets/Docs with owner columns and due dates.
Redaction gates guided by PAIA (open-by-default where lawful) and POPIA (protect personal info).
Versioned exports: Theme PDFs and the integrated synthesis are generated from the same dataset, so a fix in one place updates everywhere.

Why this matters. Digital workflows reduce manual hand-offs, create an audit trail, and let leadership see progress at a glance—exactly the kind of consequence-friendly culture push AGSA keeps calling for.

What changed for reviewers (before vs after)

Before: Dozens of PDFs; inconsistent counts; late-stage reconciliation; limited traceability.
After: One database; consistent tags; filterable by Theme/Research Area; quantified; citations to the original text; exportable to report templates.

Policy value. This allowed Reference Groups to focus on first-order questions (e.g., differentiated municipal model, IGR compacts) and second-order frameworks (e.g., professionalisation rules, digital standards) surfaced across themes.

Lessons learned (practical notes for SA policy teams)

Start with a shared schema (what is a “claim”?) before turning on any AI.
Find-and-replace your way to consistency (regex + validation lists beat free-text every time).
Quantify early. Publish interim counts—even if imperfect—to expose taxonomy gaps.
Make privacy the default (label fields; build redaction steps into the flow).
Tie reporting to decisions. Frame outputs as decision prompts (now/next/later), not just summaries.

How to replicate this in your organisation (mini-playbook)

Set up (week 1–2)

Build a claims table with the fields listed above.
Draft Theme keyword lists and teach the team how to add synonyms.
Stand up the four-board workflow (Ingest → Tag → Review → Publish).

Process (rolling)

Batch submissions; run regex-aided tagging; have AI propose tags; reviewers confirm.
Use TEXTJOIN/FILTER/QUERY patterns (above) to auto-generate top terms and counts by theme.
Export to standardised H2/H3 report templates with Executive Summary → Findings → Interpretation → Implications.

Governance

Appoint an evidence custodian (sign-off authority).
Publish open data where lawful (and redacted where required) to boost public trust and enable external review.

Conclusion

The White Paper team asked for a coherent, quantified picture of what South Africans said—and a way to act on it. By combining AI data analysis with digital workflows, we built a transparent pipeline from raw text to decision prompts that reflect where the strongest public signals are. The 100-page integrated analysis is now a living evidence base: it updates as reviewers refine tags or add late submissions, and it exports cleanly into theme reports and policy option packs. That lets Reference Groups concentrate on weighing first-order structure (e.g., differentiated municipal models, rules-based IGR) and second-order frameworks (professionalisation, performance and data systems) against trade-offs and feasibility.

Most importantly, this approach fits our context: it’s affordable, uses familiar tools, respects POPIA/PAIA, and is grounded in public-sector AI guidance. As the process moves toward March 2026, the combination of clean data, clear workflows, and credible sources is what will keep the drafting focused on outcomes that communities can feel.

AI-Readiness & Workflow Audit
Automate repetitive work. Save hours weekly. Modernise ops.
For teams swamped by admin/reporting/content tasks
Get: 3–5 high-impact automations + live setup of your first workflow
Outcome: More capacity today, scalable foundation for tomorrow
Book this Audit

Frequently Asked Questions

1) How did you ensure POPIA/PAIA compliance while using AI?

Data minimisation, redaction gates before publishing, and a register of lawful bases for processing. We followed the Information Regulator’s PAIA guide and POPIA guidance notes.

2) What counts did you publish?

Per theme: most-cited problems; top proposals/solutions; and cross-theme roll-ups—grounded in the 265-submission dataset documented in the integrated analysis.

3) How do AI outputs get verified?

Every auto-tag is reviewed by a theme lead; disagreements are logged; prompts evolve. This mirrors OECD advice on accountable AI in the public sector.

4) Which global frameworks guided the digital workflow design?

World Bank GovTech materials (shared platforms, service digitisation) and OECD public-sector AI guidance.

Author

Romanos Boraine

Resource Hub

Case Studies

Date

08 Sept 2025

Between August 2025 and today, I was subcontracted to help the Local Government White Paper team ingest and synthesise 270 public submissions (received by and shortly after 31 July 2025) into a single evidence base and a 100-page “Integrated Analysis of Public Submissions” for specialist review.

The deliverables include a normalised database, quantified reports across 11 thematic areas and 5 research areas, and cross-country insights to inform options for the revised White Paper due by March 2026.

The punchline: by combining AI data analysis (for fast, consistent text classification) with digital workflows (for structured review and audit trails), we cut manual collation time dramatically, improved traceability, and gave the Reference Groups a consolidated, queryable view of the public’s proposals and solutions—right when the policy window is open.

Related keywords we’ll cover

AI in government; policy text mining; GovTech; evidence-based policymaking; participatory governance; POPIA compliance; PAIA requests; coalition governance; municipal finance; audit outcomes; digital workflow automation; data cleaning; prompt engineering; reproducible policy analysis; intergovernmental relations; service delivery dashboards.

Key Takeaways
A simple schema + regex + AI beats manual collation for speed, consistency and traceability.
AI data analysis accelerates classification; human review secures legitimacy and accuracy.
Digital workflows create an auditable path from submission to policy option.
Quantification (counts and rankings) reveals consensus and contention faster.
Compliance-by-design (PAIA/POPIA gates) avoids late surprises.
Aligning with GovTech/OECD guidance helps benchmark quality and risk posture.

Project context and goals

Mandate. Support report writers and local government experts to turn raw submissions from civilians, political parties, government entities, NPOs, and the private sector into a coherent dataset and decision-ready analysis. Outputs: (1) a normalised database, (2) 11 thematic and 5 research-area reports reflecting public perspectives, (3) quantified counts of problems/proposals/solutions, and (4) a 100-page integrated synthesis for specialist review and policy optioning.

White Paper for Local Government timetable. The revised White Paper aims to set principles, propose policy shifts with evidence, and guide the transition roadmap—finalisation by March 2026.

Problem we solved. Disparate PDFs and emails, uneven formats, and overlapping issues made manual synthesis slow and opaque. Our approach created a single source of truth and reproducible analysis steps, letting subject-matter teams focus on first- and second-order policy choices rather than data wrangling.

Method stack (AI + human-in-the-loop)

Core tools. Google Drive apps (Sheets, Docs) for the shared data model and traceable edits; ChatGPT-5 and Gemini 2.5 for assisted classification and summarisation; light prompt engineering for consistency; and versioned templates for the 11 thematic reports and the integrated synthesis.

Why this stack. It’s accessible to government teams, keeps the audit trail in familiar systems, and aligns with GovTech guidance to strengthen core platforms and transparent workflows rather than bespoke, brittle tools. See the World Bank’s GovTech Maturity Index for how platforms and shared services underpin capability gains (global GTMI average improved from 0.519 (2020) to 0.552 (2022)).

Guardrails. We embedded privacy and transparency requirements into the workflow—open-by-default where lawful (PAIA), protective where required (POPIA). The Information Regulator’s guidance and PAIA resources informed the templates for access and redaction.

Data architecture: from messy inputs to a tidy policy dataset

Normalisation. Every submission was split into atomic “claims,” each mapped to: Thematic Area, Problem, Proposal, Solution, and evidence snippet. This mirrors how the integrated report organises insights and cross-cutting shifts (e.g., professionalisation, IGR, digital modernisation).

Schema highlights.

source_id (submission), claim_id (unique key), theme (1–11), type (problem/proposal/solution), keywords (tag set), quant_count (if quantifiable), quote (short extract).
Status flags: needs_review, escalate_reference_group, confidential_redaction.

Benefits. This structure supports AI data analysis (reliable tagging), digital workflows (assign, review, approve), and reproducible exports (by Theme/Research Area).

Smart tagging with Sheets + regex (repeatable and transparent)

We used plain-English keyword libraries per theme (expandable as reviewers discovered synonyms) and regex helpers to reduce false positives. Three examples:

=IF(B2<>"","(?i)(" & TRIM(B2) & ")", "")

=TEXTJOIN(
  ", ", TRUE,
  FILTER(Themes!B$2:B,
    REGEXMATCH('Main Database'!A3, "(?i)"&Themes!C$2:C)
  )
)

=IF(Themes!L2="","",SPLIT(REGEXREPLACE(Themes!L2,",\s*",","),","))

Counting and ranking the most frequent tagged items across all breakdown sheets:

=ARRAYFORMULA(
  QUERY(
    TOCOL(
      TRIM(
        SPLIT(
          REGEXREPLACE(TEXTJOIN(",",TRUE,'Break Down'!B2:AM266),",\s*",","),
          ","
        )
      ),
    1),
    "select Col1, count(Col1) 
     where Col1 <> '' 
     group by Col1 
     order by count(Col1) desc 
     label count(Col1) ''",
    0
  )
)

This gave reviewers transparent, reproducible counts for “most cited problems” and “top proposed solutions” per theme.

AI-assisted classification (with a human brake)

Why AI here. Submissions varied widely in style and length. AI helped chunk, label, and summarise at scale, while reviewers validated edge cases and refined the taxonomy—classic human-in-the-loop.

Prompts and patterns. We templated prompts to (1) extract claims, (2) assign Theme + type (problem/proposal/solution), (3) list citations/paragraph refs, and (4) propose LSI tags (e.g., “coalition rules,” “Blue/Green Drop,” “participatory budgeting”).

Risk management. We adhered to OECD guidance that AI in the public sector should be used with clarity on benefits, risks, and accountability—not as a black box.

Quantification: turning submissions into counts and signals

For each theme, we quantified the frequency of problem statements and mapped proposed remedies. The integrated synthesis presents first-order (structural) and second-order (framework) shifts—e.g., functional assignment, coalition stability, professionalisation, and digital modernisation—to help policymakers see system levers rather than isolated fixes.

Example signals produced for reviewers

Governance & leadership: dominance of calls to professionalise leadership, depoliticise administration, and strengthen oversight (AGSA consequences, MPACs, lifestyle audits).
Administration & capacity: merit-based hiring, protected tenure, minimum competencies, and digitised workflows across HR/finance/assets/records.
Citizen relations: proactive transparency (PAIA), ward-level forums, participatory budgeting, and local ombuds.

These signals align with recent AGSA audit insights showing persistent weaknesses and regressions—underscoring why consequence management and capability building must advance together.

Reporting: 11 thematic areas & 5 research areas (how we packaged it)

We generated 11 thematic reports (Governance; Administration; Citizen relations; Structure; IGR; Finance; Infrastructure; LED; Spatial planning; Environment/Climate/Disaster; Traditional & Khoi-San) nested within Research Areas A–E, with a consistent four-part structure per theme: Executive Summary, Main Findings, Interpretation, Implications for White Paper drafting.

The 100-page integrated document then surfaces cross-cutting shifts (e.g., rules-based IGR, finance follows function, digital modernisation, citizen power & transparency) and immediate decision prompts for the White Paper team.

Comparative research: what other countries are doing (and why it helps)

To avoid “reinventing the wheel,” we triangulated proposals against international practice and GovTech guidance:

Platform-first governance. Countries moving up the GTMI scale invest in shared platforms (identity, payments, case management) and workflow digitisation that reduce compliance duplication—exactly the kind of gains local government needs for HR/SCM/records.
Responsible AI adoption. OECD’s 2024 work stresses transparent use cases, risk registers, and algorithmic accountability instruments—features we mirrored in our AI review sheets and decision logs.

These references framed the “what good looks like” checklist for our local context.

Digital workflows that kept everyone aligned

We built end-to-end digital workflows around four boards: Ingest → Tag → Review → Publish.

Key controls

Assignment & SLAs in Sheets/Docs with owner columns and due dates.
Redaction gates guided by PAIA (open-by-default where lawful) and POPIA (protect personal info).
Versioned exports: Theme PDFs and the integrated synthesis are generated from the same dataset, so a fix in one place updates everywhere.

Why this matters. Digital workflows reduce manual hand-offs, create an audit trail, and let leadership see progress at a glance—exactly the kind of consequence-friendly culture push AGSA keeps calling for.

What changed for reviewers (before vs after)

Before: Dozens of PDFs; inconsistent counts; late-stage reconciliation; limited traceability.
After: One database; consistent tags; filterable by Theme/Research Area; quantified; citations to the original text; exportable to report templates.

Policy value. This allowed Reference Groups to focus on first-order questions (e.g., differentiated municipal model, IGR compacts) and second-order frameworks (e.g., professionalisation rules, digital standards) surfaced across themes.

Lessons learned (practical notes for SA policy teams)

Start with a shared schema (what is a “claim”?) before turning on any AI.
Find-and-replace your way to consistency (regex + validation lists beat free-text every time).
Quantify early. Publish interim counts—even if imperfect—to expose taxonomy gaps.
Make privacy the default (label fields; build redaction steps into the flow).
Tie reporting to decisions. Frame outputs as decision prompts (now/next/later), not just summaries.

How to replicate this in your organisation (mini-playbook)

Set up (week 1–2)

Build a claims table with the fields listed above.
Draft Theme keyword lists and teach the team how to add synonyms.
Stand up the four-board workflow (Ingest → Tag → Review → Publish).

Process (rolling)

Batch submissions; run regex-aided tagging; have AI propose tags; reviewers confirm.
Use TEXTJOIN/FILTER/QUERY patterns (above) to auto-generate top terms and counts by theme.
Export to standardised H2/H3 report templates with Executive Summary → Findings → Interpretation → Implications.

Governance

Appoint an evidence custodian (sign-off authority).
Publish open data where lawful (and redacted where required) to boost public trust and enable external review.

Conclusion

The White Paper team asked for a coherent, quantified picture of what South Africans said—and a way to act on it. By combining AI data analysis with digital workflows, we built a transparent pipeline from raw text to decision prompts that reflect where the strongest public signals are. The 100-page integrated analysis is now a living evidence base: it updates as reviewers refine tags or add late submissions, and it exports cleanly into theme reports and policy option packs. That lets Reference Groups concentrate on weighing first-order structure (e.g., differentiated municipal models, rules-based IGR) and second-order frameworks (professionalisation, performance and data systems) against trade-offs and feasibility.

Most importantly, this approach fits our context: it’s affordable, uses familiar tools, respects POPIA/PAIA, and is grounded in public-sector AI guidance. As the process moves toward March 2026, the combination of clean data, clear workflows, and credible sources is what will keep the drafting focused on outcomes that communities can feel.

AI-Readiness & Workflow Audit
Automate repetitive work. Save hours weekly. Modernise ops.
For teams swamped by admin/reporting/content tasks
Get: 3–5 high-impact automations + live setup of your first workflow
Outcome: More capacity today, scalable foundation for tomorrow
Book this Audit

Frequently Asked Questions

1) How did you ensure POPIA/PAIA compliance while using AI?

Data minimisation, redaction gates before publishing, and a register of lawful bases for processing. We followed the Information Regulator’s PAIA guide and POPIA guidance notes.

2) What counts did you publish?

Per theme: most-cited problems; top proposals/solutions; and cross-theme roll-ups—grounded in the 265-submission dataset documented in the integrated analysis.

3) How do AI outputs get verified?

Every auto-tag is reviewed by a theme lead; disagreements are logged; prompts evolve. This mirrors OECD advice on accountable AI in the public sector.

4) Which global frameworks guided the digital workflow design?

World Bank GovTech materials (shared platforms, service digitisation) and OECD public-sector AI guidance.

Similar Resources

AI Data Analysis & Digital Workflows for SA White Paper on Local Government

Between August 2025 and today, I was subcontracted to help the Local Government White Paper team ingest and synthesise 270 public submissions.

Date

08 Sept 2025

Topic

Case Studies

Author

Romanos Boraine

How I Sped Up a UNICEF Research Project with AI Custom GPTs for Data Analysis and Synthesis

Over a three-month engagement with a UNICEF-aligned research team, I used two custom-built GPTs and a spreadsheet-first architecture to dramatically increase the speed and rigour of a child poverty study in Zambia.

Date

08 Sept 2025

Topic

Case Studies

Author

Romanos Boraine

Crafting a Digital Brand: A Case Study on the Launch of Craft Cask

This case study documents the journey of building Craft Cask, a digital-first media brand focusing on whisky, from the ground up.

Date

20 Jun 2025

Topic

Author

Romanos Boraine

7 Ways NPOs Waste Money on Inefficient Digital Systems

South African non-profit organisations (NPOs) are the lifeblood of social progress, driven by passionate, purpose-driven teams committed to making a significant impact with limited resources.

Date

17 Jun 2025

Topic

Author

Romanos Boraine

Similar Resources

AI Data Analysis & Digital Workflows for SA White Paper on Local Government

Between August 2025 and today, I was subcontracted to help the Local Government White Paper team ingest and synthesise 270 public submissions.

Date

08 Sept 2025

Category

Case Studies

Author

Romanos Boraine

How I Sped Up a UNICEF Research Project with AI Custom GPTs for Data Analysis and Synthesis

Over a three-month engagement with a UNICEF-aligned research team, I used two custom-built GPTs and a spreadsheet-first architecture to dramatically increase the speed and rigour of a child poverty study in Zambia.

Date

08 Sept 2025

Category

Case Studies

Author

Romanos Boraine

Crafting a Digital Brand: A Case Study on the Launch of Craft Cask

This case study documents the journey of building Craft Cask, a digital-first media brand focusing on whisky, from the ground up.

Date

20 Jun 2025

Category

Author

Romanos Boraine

7 Ways NPOs Waste Money on Inefficient Digital Systems

South African non-profit organisations (NPOs) are the lifeblood of social progress, driven by passionate, purpose-driven teams committed to making a significant impact with limited resources.

Date

17 Jun 2025

Category

Author

Romanos Boraine

Similar Resources

AI Data Analysis & Digital Workflows for SA White Paper on Local Government

Between August 2025 and today, I was subcontracted to help the Local Government White Paper team ingest and synthesise 270 public submissions.

Date

08 Sept 2025

Category

Case Studies

Author

Romanos Boraine

How I Sped Up a UNICEF Research Project with AI Custom GPTs for Data Analysis and Synthesis

Over a three-month engagement with a UNICEF-aligned research team, I used two custom-built GPTs and a spreadsheet-first architecture to dramatically increase the speed and rigour of a child poverty study in Zambia.

Date

08 Sept 2025

Category

Case Studies

Author

Romanos Boraine

Crafting a Digital Brand: A Case Study on the Launch of Craft Cask

This case study documents the journey of building Craft Cask, a digital-first media brand focusing on whisky, from the ground up.

Date

20 Jun 2025

Category

Author

Romanos Boraine

7 Ways NPOs Waste Money on Inefficient Digital Systems

South African non-profit organisations (NPOs) are the lifeblood of social progress, driven by passionate, purpose-driven teams committed to making a significant impact with limited resources.

Date

17 Jun 2025

Category

Author

Romanos Boraine