Decision intelligence in action: Augmenting human judgment with predictive data in high-stakes operational environments.

AI to Support Public Decisions

How Faculty builds AI into decisions in counter-terrorism, health care and teaching

0%of Daesh propaganda detected in tests of the Home Office and Faculty tool, with 0.005% false positives

£0.0M+estimated daily savings to the NHS and government from Cera's AI-supported home care (Faculty analysis, 2025)

0+lesson plans Inspired Education teachers shared for Faculty's AI lesson planner

Executive Summary

The Center for Human Progress wrote this page from Faculty's published case studies and other public sources.

Marc Warner co-founded Faculty, an applied AI company in London, in 2014. Faculty builds AI into the processes organizations use to make decisions. Its projects include the National Health Service's COVID-19 Early Warning System, a tool developed with the Home Office to detect Daesh propaganda, a hospital-risk model for the home care company Cera and a lesson planner for the Inspired Education school group.

Most results below are Faculty's or its clients' own reports. A peer-reviewed paper covers the accuracy of the Cera model.

Who's behind this

Key terms

Two problems Faculty's case studies describe

Dashboards that describe without acting
In Faculty's view, AI falls short of its potential when "the output is a series of charts and dashboards that describe the world without acting upon it." In the Home Office project, the analysts who understood the propaganda worked in government, and the people who could remove it worked at the tech companies.
Admin that crowds out teaching
"A mountain of paperwork doesn't draw anyone into teaching - but it certainly drives them out of it." Inspired Education's chief information officer, Torben Lundberg, says teachers are wary of tools that add "more tracking" without helping them teach.

Three lessons from Faculty's applied AI work

Faculty's book Ten Lessons From Ten Years of Applied AI draws a lesson from each of ten client projects. Three come from projects on this page.

Building AI into operations (Home Office)
"AI is an operational discipline, not an analytical one." On this view, AI belongs inside the tools people use to run a process, linked to the actions they take. The Home Office tool was designed to run inside a platform's upload process.
Automating some tasks and supporting others (Inspired Education)
"Augment human tasks that require judgement. Automate those that don't." At Inspired, one AI tool drafts lesson materials and another drafts tests, and teachers review and edit both before use.
Testing hypotheses before modeling (Cera)
"It's data SCIENCE, not DATA science." Before building a model, the team tested hypotheses about what causes and predicts hospital admissions. Answers about bladder function, which people were uncomfortable giving, showed little link with later admissions. Drowsiness, together with other signs, could signal a likely admission several days ahead.

Case studies

Detecting Daesh propaganda with the Home Office

In 2017, propaganda from Daesh (ISIS) was spreading on the largest social media platforms, and people inspired by the group carried out a series of attacks in the UK and elsewhere. In Faculty's account, the platforms could not remove it fast enough and said accurate automated detection was not yet possible. The Home Office's Office for Security and Counter-Terrorism turned to Faculty, then called ASI Data Science, and the two developed a tool together.

The tool had to detect propaganda within two hours of release, the window in which most links to Daesh videos are shared. YouTube engineers told the Home Office they would consider it feasible only if its false-positive rate fell to 0.005%, or five in every 100,000 videos. It also had to tell propaganda apart from prayer videos, news reports and other legitimate content. It used only the videos themselves, not data about users.

In tests, the tool "was able to detect 94% of the propaganda, with only a 0.005% false-positive rate." At YouTube's volume at the time of five million uploads a day, that would mean 250 incorrectly flagged videos a day, which Faculty says one moderator could check. The Home Office announced the tool in February 2018 and said it would share the method with smaller platforms.

Read Faculty's case study.

The NHS Early Warning System and Cera's risk model

During the COVID-19 pandemic, Faculty worked with the NHS to build the Early Warning System (EWS). It forecast COVID-19 hospital demand for every NHS hospital trust in England three weeks ahead. Its models combined positive case numbers, calls to the NHS 111 helpline and mobility data from Apple and Google. The forecasts warned hospitals of likely spikes so they could move staff, beds and equipment. Faculty says the EWS had more than 1,000 users.

Read Faculty's case study.

Faculty also worked with Cera, a home care company whose carers record observations at each visit in Cera's app. Faculty's analysts and Cera's data scientists, clinicians and managers built a model that predicts which clients are likely to be admitted to hospital within a week. Each day a registered nurse reviews a risk-ranked list and arranges low-cost help, such as a call with the client's family doctor (GP) or a pharmacist's medication review.

Nathan Windle and colleagues at Cera and Faculty tested the model in a peer-reviewed study. The model and 10 clinicians assessed the same 150 client records, and the model was better at telling which clients would be admitted. It took under a minute, and the clinicians typically took over 40 minutes. Cera's own data scientists later adapted the model to predict falls. In September 2025, Cera announced Faculty's estimate that Cera's AI-supported care model was saving the NHS and government more than £1.5 million a day and had saved over £1 billion in total.

Read Faculty's case study.

Lesson planning at Inspired Education

Inspired Education runs more than 125 private schools in 30 countries. Faculty helped it survey more than 700 school leaders and teachers, who reported spending on average over six hours a week on lesson planning and over four hours creating and marking weekly tests.

Inspired's teachers shared more than 50,000 of their best lesson plans. Using these, Faculty built a lesson planner on GPT-4, OpenAI's large language model. A teacher writes a lesson plan and objectives, with help from the AI. The planner then finds the most relevant lessons in the bank and drafts slides, class activities and assignments from them in Inspired's template, a method called retrieval-augmented generation. Drafts are ready to review in under a minute.

Faculty says the tool lets teachers focus their professional judgment on making learning interactive and directly engaging with students.

Read Faculty's case study.

Who did what in three of the projects

Row	What the software did	What people did
Home Office tool	Could flag likely Daesh propaganda at upload, before publication	Would review the small share of videos needing it, which the Home Office put at 50 in a million random videos
NHS Early Warning System	Forecast COVID-19 hospital demand for each NHS trust three weeks ahead	Used the forecasts, with other information, to plan where staff, beds and equipment went
Cera's risk model	Ranked home care clients each day by their risk of a hospital admission within a week	A nurse reviewed the list and arranged help, such as a GP call

Human-led AI

Faculty says Warner founded the company "to help organisations make better decisions using human-led AI." Its Inspired case study adds that most high-value tasks "require professional judgement," and that "It is usually unwise to try and replace this human judgement."

Questions for public bodies considering AI

These are the Center's questions, drawn from Faculty's case studies.

1. Which decision would the model inform?
Faculty says it works "decision-by-decision." On this page the decisions were which videos to block, where to send staff and equipment, and which client a nurse should contact first.
2. Which tasks need professional judgment?
Faculty's Inspired lesson separates tasks that need judgment, which AI should support, from routine tasks, which it can automate.
3. How far should staff rely on the model?
The EWS showed users which inputs drove each forecast. For the Home Office tool, YouTube engineers named the false-positive rate at which they would consider it feasible.

Published by Center for Human Progress (US) Inc., a 501(c)(3) organization