CASE STUDIES WE STUDY

Public AI deployments
SMBs can learn from.

We study these in detail — not because they’re ours, but because the patterns underneath them translate cleanly to 30–300 person businesses.

None of these are our client engagements. Every organisation, outcome and link below is publicly verifiable.

Khan AcademyAmerican Red Cross + DataKindSecond Harvest Food Bank + DataKindDataKind Edvise (UCF & Google.org)Save the Children + DataKind + MicrosoftUK boutique law firm (composite, anonymised)US specialty clinic group (composite, anonymised)

Khan Academy

Mountain View, CA · 2023–2025

150M+

Learners Reached

Khanmigo: A Free AI Tutor for Every Student on Earth

Challenge

Khan Academy’s mission is to provide free, world-class education globally — but scaling personalized 1-on-1 tutoring for 150M+ learners without hiring thousands of teachers seemed impossible.

Approach

Launched Khanmigo in 2023 — a GPT-4-powered AI tutor and teacher assistant that guides students through problems using the Socratic method, never just giving answers. Available free to every learner on the platform.

Outcomes

  • Deployed to 150M+ registered learners worldwide
  • Teachers save hours weekly on lesson planning via AI assistance
  • Writing Coach launched with AP & SAT prompt library
  • Khan Academy questions now integrated directly into ChatGPT
Read the full case studyCopilot Studio agent patternsMicrosoft Copilot consultingSource

American Red Cross + DataKind

Washington, D.C. · 2015–2020

2.5M+

Homes Protected

Predicting Home Fire Risk to Save Lives at Scale

Challenge

The American Red Cross responds to a home fire every 24 minutes in the US — but had no way to proactively identify and reach the highest-risk households before tragedy struck.

Approach

DataKind DC volunteers built a national fire risk prediction model using open data (census, fire incident records, housing age, poverty indices) to score every US census tract. The Red Cross used the model to direct their Home Fire Campaign canvassers to highest-risk neighbourhoods.

Outcomes

  • Model identifies highest-risk communities across all 50 states
  • Volunteers installed smoke alarms in 2.5M+ high-risk homes
  • Over 1,000 lives saved through targeted prevention campaigns
  • Model iterated and improved multiple times with open-source code
Read the full case studyAzure AI consulting5 questions that disqualify a use caseSource

Second Harvest Food Bank + DataKind

Orlando, FL · 2025

Open

Access Platform

“Ladder” — A Community Data Platform to Fight Food Insecurity

Challenge

Food insecurity and poor health are closely linked to economic instability in Central Florida, but social sector organizations were siloed, duplicating effort and missing the people who fell through the gaps.

Approach

DataKind co-developed “Ladder” with Second Harvest Food Bank of Central Florida — a freely available platform that makes fragmented community data accessible, actionable, and equitable. Organizations can identify unserved populations and coordinate services using shared data layers.

Outcomes

  • Launched September 2025 as an open-access platform
  • Integrates food insecurity, health, and economic stability data
  • Enables cross-sector coordination to eliminate service duplication
  • Adopted by multiple Central Florida social sector partners
Read the full case studyAI governance for SMBsAI strategy & roadmapSource

DataKind Edvise (UCF & Google.org)

Orlando, FL · 2025

68K+

Students Supported

AI Advising Platform Scaling Student Success at 68,000-Student University

Challenge

The University of Central Florida — with 68,000 students — struggled to proactively identify students at risk of dropping out. Academic advisors lacked the tools to act on signals before it was too late.

Approach

DataKind built “Edvise” — a freely available AI advising support platform (backed by Google.org) — that surfaces predictive risk signals to advisors, enabling proactive, data-driven interventions. Launched nationally across partner institutions in 2025.

Outcomes

  • Deployed at UCF (68K students) and Lee College among others
  • Advisors act earlier on at-risk students using AI-generated signals
  • Partnered with National Student Clearinghouse for data integration
  • Expanding to Historically Black Colleges via Ed Advancement partnership
Read the full case studyMicrosoft Copilot ROI for SMBsAI ROI calculatorSource

Save the Children + DataKind + Microsoft

Washington, D.C. / Global · 2023–2024

10x

Faster Insights

Generative AI Transforming Humanitarian Data Access for Crisis Response

Challenge

During humanitarian crises, data exists in abundance — but it is fragmented, inaccessible, and often too slow to act on. Save the Children’s responders struggled to translate raw data into timely, actionable insights in the field.

Approach

The Humanitarian Data Insights Project (HDIP) — a partnership between Save the Children, DataKind, and Microsoft — used generative AI to build a platform that synthesises crisis data streams and surfaces plain-language insights for humanitarian responders, dramatically cutting the time from data to decision.

Outcomes

  • Generative AI reduces time-to-insight from days to hours
  • Natural language querying of complex humanitarian datasets
  • Validated and extended in a new phase launched in 2023
  • Model for responsible AI use across the humanitarian sector
Read the full case studyAzure AI consultingMicrosoft AI stack SMBs already ownSource

UK boutique law firm (composite, anonymised)

London, UK · 2025–2026

−70%

First-pass review hours

Cutting first-pass discovery review 70% inside the firm tenant

Challenge

A 40-fee-earner UK litigation boutique was losing margin on fixed-fee disclosure exercises. First-pass review of discovery production routinely consumed 400–600 associate hours per mid-size matter — work the client would not pay for at full rate, but the firm had to do for privilege screening and relevance tagging before substantive review could begin.

Approach

Deployed an Azure OpenAI + AI Search classifier inside the firm’s own Microsoft 365 tenant. Pipeline ingests the disclosure production, tags each document by custodian, date range, doc type and topic, and flags potential privilege candidates for human review. Microsoft Purview sensitivity labels per matter prevent cross-matter leakage. The deployment memo — covering data residency, retention, audit and the duty-of-competence narrative — was signed off by the firm’s GC and shared with their PI insurer before go-live.

Outcomes

  • First-pass review hours cut ~70% on the three matters measured
  • Privilege-flagged subset reviewed by a partner inside one day (vs one week)
  • Zero cross-matter data leakage incidents in 6 months of operation
  • Deployment memo accepted by the firm’s PI carrier without premium uplift
Read the full case studyAI consulting for law firmsAI governance pack for SMBsAzure AI consultingSource

US specialty clinic group (composite, anonymised)

Texas, USA · 2025–2026

−65%

Front-desk intake time

Front-desk intake reduced 65% inside the BAA boundary

Challenge

A US specialty clinic group with eight locations and ≈60 clinicians was drowning in pre-visit paperwork. Front-desk staff manually keyed scanned intake forms, insurance cards and referring-provider documents into the EHR — ≈12 minutes per new patient, with a measured intake error rate above 8% feeding downstream prior-auth denials. Anything they deployed had to stay inside the Microsoft HIPAA BAA and could not act as a clinical-decision-support system that would attract FDA SaMD scrutiny.

Approach

AI Builder forms-processing for intake packets and insurance cards. Azure AI Document Intelligence for referring-provider documents and operative reports. Power Automate writes structured output to the EHR via FHIR APIs (athenahealth). All inside the BAA boundary; front-desk staff confirm the extracted fields rather than type them. Deployment is explicitly administrative — no clinical-decision-support, no autonomous patient communication — with the clinician as the decision-maker on every chart.

Outcomes

  • Per-patient intake time dropped from ≈12 to ≈4 minutes
  • Intake error rate fell from 8% to below 2% (measured on a 90-day sample)
  • Prior-auth denials traceable to intake errors down ~55% in the same window
  • BAA boundary preserved; no PHI touched a service outside the agreement
Read the full case studyAI consulting for healthcare SMBsAI governance pack for SMBs5 questions that disqualify a use caseSource

Want to see your own pattern here?

We help SMBs apply the same patterns at the right scale.

None of the deployments above happened by accident. Each one started with a clear problem, defensible data access, and a named owner. That’s where we start every engagement.

Take the free assessmentTalk to us about a patternRead the field notes

Hi, I'm Nova. Chat, speak, or show me — I'll point you at the right tool.