The AI agent platform for regulated organisations and government
Compliant AI agents, deployed where your regulator says.First agent validated in weeks, not 18 months.
Ready-to-configure agents, compliance controls and validation evidence—in the environment your organisation requires.
ISO/IEC 27001 certified · 2026
- 01Choose an agent
- 02Apply controls
- 03Validate
- 04Approve and deploy
Start with a real task your team needs to complete.
Illustrative workflow · sample data
- 9–18 months to build alone→ weeks with Forge*
- 4 deployment optionscloud to air-gapped
- 12 regulated sectorsagents and packs
- Open-weight or managedmodels
* ADB’s estimate for building agents, controls and evidence in-house. “Weeks” means a first agent validated in your test environment; production timing depends on your approvals.
Our mission is to help regulated organisations ship AI faster and maintain compliance as they scale.
Why Forge
A working AI demo is only the beginning.
Getting an agent into production takes more than a working demo. Teams need permissions, data protection, testing and approval records. Forge brings these together with the agent, so you can build on the same foundation for each use case.
Build it yourselves
- 9–18 months (our estimate)
- 8–15 specialists
- You build controls, evidence and deployment from zero
General agent platforms
- Powerful tools
- Your team still designs the controls and writes the tests
- Evidence is assembled for every use case
ADB Forge
- A ready agent for the task
- Your regulator’s compliance pack
- Validation evidence, deployed where your rules require
- Less to buildYour team starts from a working agent instead of a blank project.
- Less repeated compliance workThe same control work is not rebuilt for every new use case.
- Evidence prepared alongside the workRisk and compliance reviews start with the material they need.
- Deployment that fitsThe deployment follows your requirements rather than forcing a compromise.
- Controlled improvementAgents can improve after launch without losing control of what changes.
Use Forge alongside the cloud and model providers you already work with. See the platform architecture
Compliance packs
Turn regulatory requirements into checks and evidence.
Each pack maps regulatory requirements to agent checks, records and review reports. Packs are updated as rules change and reviewed with local counsel. Ask us about coverage for your sector and country.
Controls mapped to
EU AI ActEU Artificial Intelligence Act- ISO/IEC 42001AI management systems standard
NIST AI RMFUS AI Risk Management Framework- GDPREU General Data Protection Regulation
- DORAEU Digital Operational Resilience Act
National data protection laws in our markets 13
- UAE PDPLFederal Decree-Law No. 45 of 2021
- Saudi Arabia PDPLPersonal Data Protection Law
- Qatar PDPPLLaw No. 13 of 2016
- Bahrain PDPLLaw No. 30 of 2018
- Oman PDPLRoyal Decree 6/2022
- Egypt PDPLLaw No. 151 of 2020
- Kenya Data Protection ActAct No. 24 of 2019
- Tanzania PDPAPersonal Data Protection Act, 2022
- Uganda DPPAData Protection and Privacy Act, 2019
- Rwanda DPP LawLaw No. 058/2021
- Nigeria NDPANigeria Data Protection Act, 2023
- South Africa POPIAProtection of Personal Information Act
- India DPDP ActDigital Personal Data Protection Act, 2023
Compliance packs / Retail banking
Retail banking · [Regulator]
Illustrative · sample data| Rule ID | Requirement | Check method | Severity | Evidence produced |
|---|---|---|---|---|
| BNK-DS-01 | Tell the customer the call is recorded | Transcript check, first 30 seconds | Medium | Check result with time code |
| BNK-ID-02 | Confirm identity before discussing the account | Sequence check: identity questions before account details | High | Questions asked and answer outcome |
| BNK-MS-03 | Assess needs before recommending or switching a product | Intent detection plus needs-question check; holds pending switches | High | Flagged excerpt, hold record, reviewer decision |
Use cases
Agents for the work your teams do
Review customer calls consistently, identify potential conduct issues and give reviewers a traceable record.
- The work it does
- Reviews recorded customer calls against disclosure, identity and fair-treatment checks.
- What it saves
- Reviewers start from flagged calls and the exact excerpts behind them, rather than sampling calls at random.
- What a person reviews
- A conduct reviewer decides on every flag and on any follow-up action.
- Evidence kept
- Check results, the excerpt behind each flag, and the reviewer’s decision and rationale.
All twelve sectors
- Banking
- Insurance
- Microfinance and fintech
- Wealth and capital markets
- Telecom
- Government and public services
- Healthcare
- Utilities and energy
- Mining and resources
- Transport and aviation
- Education
- Legal and professional services
Agents and compliance-pack coverage differ by sector and country. We confirm what is ready for your use case in the demo.
Human review
Try a review
Read the finding, decide what happens next and see the decision recorded.
Interactive example · sample data.
Your choices stay in this demo.
A card switch needs review
“Our premium card is a much better deal. I’ll switch you over today.”
The call contains no needs assessment before the proposed switch.
Assess the customer’s needs before recommending or switching a product.
- Recording disclosed
- Identity checked
- Complaint logged
Models
Choose the model that fits your environment
- Open-weight inside your environmentRun open-weight models inside your own environment, so prompts and data are not sent to an outside model provider.
- Managed where it fitsUse managed models from your cloud provider where they are offered in your country.
- Adapters you ownSpecialise models on your approved data with lightweight adapters you own.
- Safe upgradesSwap or upgrade models safely: a new model goes live only after it passes the same agreed tests.
See the model registry
Models
Model registry
Illustrative · sample data| Model | Type | Where it runs | Task adapter | Test result vs threshold | Status |
|---|---|---|---|---|---|
| Open-weight 8B (instruction-tuned) | Open-weight | On-premises GPUs | call-review-adapter v3 | 96.3% / 95% | Live |
| Open-weight 70B | Open-weight | National cloud | claims-intake-adapter v1 | 97.1% / 95% | Candidate |
| Managed model · in-country region | Managed | In-country cloud region | None (prompted) | 95.4% / 95% | Live |
| Open-weight 8B (instruction-tuned) | Open-weight | On-premises GPUs | call-review-adapter v2 | 93.8% / 95% | Retired |
Deployment
Your AI. Your data. Your control.
Choose where Forge runs, who operates it and who holds the keys. The right option depends on your rules and your infrastructure.
In-country cloud region
- What it means
- Forge runs in a public-cloud region located in your country.
- When it fits
- Your rules require data to be located in-country and you already use that cloud provider.
Deployments / Production
Deployment · In-country region
Illustrative · sample data- Environment
- In-country cloud region
- Operated by
- The cloud provider runs the infrastructure
- Encryption keys
- Keys held in your cloud account
- Model runtime
- A managed model offered in that region, or a model served by Forge
- Outbound connections
- In-region model endpoint only
- Updates
- Software and model updates approved by your teams before release
Run Forge on the hyperscalers—AWS, Azure, GCP and Oracle Cloud—in their in-country regions, or in a sovereign, national or on-premises environment.
Evidence
Evidence for review and sign-off
Check test results, trace actions and review the report before approving a release.
Review each update before release
Corrections become test cases. Your team approves the next version.
- 1Review outcomesReviewers see what the agent did and why.
- 2Capture correctionsCorrections are kept and become test cases.
- 3Test changesAn updated agent is run against the agreed criteria.
- 4Approve releasesYour approvers decide whether to promote it.
Questions
Common questions
What do we receive from Forge?
Agents configured for your use case, the controls and compliance pack that apply to it, validation results and reports, audit records, and the runtime for the environment you choose. The exact scope is agreed for each engagement.
Which use cases are available today?
Call review, customer service, complaints handling, claims intake, citizen helplines, case intake, know-your-customer (KYC) onboarding and document processing are examples from the catalogue. Availability depends on your sector, country, languages and deployment option, so we confirm what is ready for your case during the demo.
Can Forge work with our existing systems, models and infrastructure?
Forge connects to your systems through integrations agreed during scoping. It can use a managed model in a cloud region where one is available, or a model served by Forge in your own environment. Specific systems, models and hardware are confirmed for each deployment.
Where does our data go?
Data is processed in the environment you choose. For each deployment we document where prompts, documents, responses, logs and backups are processed and stored, which connections leave the environment, who can access it and who holds the keys.
What happens in a pilot, and who approves production?
A pilot tests the use case against agreed criteria using your approved data. Pilot scope and fees are agreed before work begins. Your organisation decides whether to move to production, after your own risk, compliance and security reviews.
Which models do you support?
Forge runs open-weight models served inside your environment, and managed models from your cloud provider where they are offered in your country. The specific models available depend on each model’s licence, your environment and the use case; we confirm them during scoping.
Do you work with open-source models?
Yes. Open-weight models are the default for on-premises, national and sovereign deployments, served with open-source inference engines. You can specialise them with adapters trained on your approved data, and you own those adapters.
How are changes reviewed after launch?
Reviewer corrections are captured and turned into test cases. An updated agent is tested against the agreed criteria, and your approvers decide whether to promote it.
Next step
See Forge on a use case like yours
See your use case in a demo. If it fits, we agree a pilot using your approved data and success criteria.
Before a pilot begins, we agree
- One initial use caseA single, well-defined task to start with.
- Data and environmentWhich data the agent may use and where the pilot runs.
- Success criteriaThe tests and thresholds that count as a pass.
- Evaluation deliverablesValidation results, audit records and a validation report.
- The decision processWho reviews the evidence and decides on production.
Pilot scope and fees are agreed before work begins.