ACCENTURE FACULTY

ACN offerings / Responsible AI and red teaming

Responsible AI,
operationalised.

Assess. Govern. Red team. Monitor.
One partner from first evaluation to continuous assurance.

Start a rapid eval See the offer

In partnership with Google, Anthropic and OpenAI

  • Google
  • Anthropic
  • OpenAI
Our history with GAO

The moment

AI is scaling faster than the controls around it.

01

Frontier models are in production

Agents built on GAO models now touch customers, code and decisions.

02

Regulation has teeth

EU AI Act obligations phase in through 2027. ISO/IEC 42001 and NIST AI RMF are becoming the audit baseline.

03

Boards want evidence

Principles are no longer enough. Leaders need tested models, owned risks and controls they can show.

The gap: most enterprises cannot say which AI they run, how it behaves under attack, or who owns the risk.

Section 01 / 05

Built with the
GAO labs.

Partnerships with Google, Anthropic and OpenAI. We test the models you actually run.

Our GAO partners

We work inside the ecosystems you build on.

Gemini

Google

Responsible AI tooling for Gemini on Vertex AI, with evaluation in Google Cloud pipelines.

Claude

Anthropic

Claude Fable powers our rapid evaluation engine. A shared focus on AI safety.

GPT

OpenAI

Evaluation and red teaming of GPT models and agent deployments, building on Faculty’s evaluation heritage.

Why GAO matters: we test the models you actually run, with insight into how they are built and where they fail. Model agnostic across Gemini, Claude, GPT and open weight models.

Our offer

One integrated path to trustworthy AI.

  1. 01AssessCurrent state assessment and rapid model evaluation.
  2. 02DesignA governance operating model built from the findings.
  3. 03ImplementPolicy, Risk & Controls embedded in the AI lifecycle.
  4. 04AssureRed teaming and continuous evaluation in production.

Accenture brings

Delivery at scale, industry depth, risk and regulatory practice, and alliances with the model providers.

Faculty brings

Frontier AI safety science, model evaluation and red teaming expertise, and decision grade data science.

Section 02 / 05

Know where
you stand.

Current state assessment, powered by the Claude Fable rapid eval.

Current state assessment

Five dimensions. One honest baseline.

01 Inventory02 Model03 Data04 Security05 Accountability
Current Target / Illustrative
  1. 01InventoryWhich AI systems, models and vendors are in use, and where.
  2. 02Model behaviourAccuracy, robustness, bias and hallucination on real tasks.
  3. 03Data and privacyLineage, consent, leakage and retention.
  4. 04Security and misusePrompt injection, jailbreaks, agent permissions.
  5. 05AccountabilityOwners, policies, decision rights, regulatory readiness.

Outputs

  • AI system register
  • Maturity profile
  • Prioritised risks
  • Executive readout

Spotlight / Claude Fable rapid eval

Evidence in weeks, not quarters.

Anthropic’s Claude Fable is our evaluation engine. It writes realistic and adversarial tests from your own policies, runs them against your models and agents, and grades every response. Faculty scientists calibrate the grader and verify each high severity finding.

  1. Days 1 to 2ScopeUse cases, policies, risk appetite.
  2. Days 3 to 5GenerateFable builds test suites from your policies.
  3. Week 2Run and gradeAutomated runs, expert calibrated grading.
  4. Week 2Read outScored findings and priority actions.

Claude
Fable

How it works

  1. 01Your policiesand risk appetite in.
  2. 02GenerateThousands of realistic and adversarial tests.
  3. 03RunAgainst your models and agents.
  4. 04GradeEvery response against policy.
  5. 05CalibrateFaculty experts verify high severity findings.
  6. 06EvidenceScored findings and an audit ready pack.
  • Thousands of scenarios per use case
  • Any GAO or open weight model
  • Human verified, audit ready evidence

Sample output

A clear picture and a ranked agenda.

InitialDevelopingDefinedManagedOptimisedInventoryModel behaviourData & privacySecurityAccountability
Target: between Defined and Managed / Illustrative

Top findings (illustrative)

  1. 01Customer agent open to prompt injection hidden in inbound email.
  2. 02A third of GenAI use cases missing from any register.
  3. 03No named owner for third party model risk.

Each finding maps to a governance action.

Section 03 / 05

From findings
to operations.

Turning the assessment into a governance design that runs.

Governance design

Every finding becomes an owned, operational action.

Assessment findingLeads toGovernance design actionOwner
Use cases missing from the registerMandatory AI intake, inventory and risk tiering at idea stageAI office
Agents open to prompt injectionRelease gate needs red team sign off for tier 1 and 2 systemsCISO
Bias untested on live populationsFairness thresholds and monitoring KPIs per use caseModel risk
No owner for third party model riskVendor and model due diligence with contractual controlsProcurement
Policies not mapped to regulationOne control set mapped to EU AI Act, ISO/IEC 42001 and NIST AI RMFCompliance

Operating model

Governance built into the AI lifecycle.

  1. IdeateIntake and risk tier
  2. BuildData and design review
  3. ValidateEval and red team gate
  4. DeployApproval and sign off
  5. MonitorDrift, incidents, retest

Decide

RAI council with clear decision rights and escalation by risk tier.

Operate

Product teams own risk day to day, with controls built into tooling.

Assure

Independent challenge plus continuous red teaming and evaluation.

Section 04 / 05

Make it real.

Implementing PRC, governance and continuous assurance.

Implementation / PRC = Policy, Risk & Controls

PRC: the backbone that makes governance enforceable.

P

Policy

  • AI policy and acceptable use
  • Model, data and agent standards
  • Mapped to EU AI Act, ISO 42001, NIST
R

Risk

  • AI risk taxonomy and tiering
  • Living risk register tied to the inventory
  • Risk appetite and key risk indicators
C

Controls

  • Control library tied to lifecycle gates
  • Automated tests and evidence
  • Continuous monitoring

Policy sets the rules Risk sizes the exposure Controls enforce and evidence

Security by design

  • Prompt injection and jailbreak defences
  • Least privilege agent and tool permissions
  • Vendor and model due diligence

Red teaming and monitoring

Attack it first, then keep watching.

Faculty led red teams pair automated attack generation with expert adversaries who know how frontier models fail, then monitor in production.

Your AI systemmodel + tools + dataPrompt injectionHarmful outputAgent tool misuseData leakageJailbreaks
  • Prompt injection
  • Harmful output
  • Agent tool misuse
  • Data leakage
  • Jailbreaks
  1. Before launchFull red team, tier 1 and 2 systems.
  2. Every releaseAutomated regression attack suite.
  3. ContinuousLive monitoring, drift, new threat intel.
  4. AnnuallyIndependent assurance to the board.

Model agnostic across GAO (Gemini, Claude, GPT) and open weight models.

Implementation roadmap

Six months from assessment to steady state.

Indicative timeline, calibrated to scope in week 1.

Section 05 / 05

Next steps.

Five actions to start in the next four weeks.

Next steps

Five actions, four weeks.

Week 1

01Confirm sponsor and use casesClient sponsor
02Agree scope, risk appetite, PRCJoint team

Week 2

03Share models and policiesClient AI office
04Launch the Fable rapid evalAccenture, Faculty

Week 3

Rapid eval in flightTests generated, run and graded; findings verified by Faculty

Week 4

05Executive readout and governance design workshopJoint team

Where to start

AI you can trust,
and prove it.

Thirty minutes with the executive accountable for AI risk or deployment. We agree what a first evaluation should prove. One way to begin: a two week Claude Fable rapid evaluation.

Accenture
[Name, title, email]
Faculty
[Name, title, email]