Capability moved faster than intervention.
AI systems are becoming more capable. Organisations are becoming more dependent on automation. Human oversight only matters if intervention remains possible.
Organisations can evidence governance. Showing that they can intervene in time is a different test.
I founded Intervene because I have watched the same pattern repeat throughout my life, in institutions I had to trust, long before I encountered it in organisations governing technology.
Harm becomes visible. Concerns are raised. Responsibility disperses. The system continues.
Three decades across technology, systems architecture, digital products, payments and organisational risk, including co-founding Financial Innovation in Transactions and Security (FITS), brought me back to that pattern. Controls, reporting and governance structures existed, yet effective intervention repeatedly failed. Problems were detected too late, escalation stalled, decisions drifted and ownership became unclear.
It took time to stop treating these as isolated experiences. The question connecting them was whether anyone could still act once the evidence changed, while there was still time to prevent harm.
AI gives that question greater urgency. Systems act faster, affect more people and produce outcomes that are harder to reverse. Regulators press for effective human oversight, and boards need confidence that intervention remains possible before the window closes.
That is why we developed AGDA™, our proprietary assessment of Intervention Readiness. It assesses whether an organisation can detect, escalate, decide and intervene within the time available, and issues an Intervention Readiness verdict with its evidence basis.
This work is personal, but its purpose is practical: to make the capacity to intervene something organisations can test rigorously, without reducing failure to individual blame. The same discipline matters beyond organisations, in how any of us reads a signal, takes responsibility and acts while there is still a choice.
For those accountable for consequential systems, AGDA™ provides a defensible answer to a question governance papers alone cannot settle.
Can you stop it in time?
The operating commitments behind the assessment.
Operating constraints, codified into how we accept work and what we will and will not deliver.
-
Independent judgement.
No vendor affiliation, no co-marketing and no commercial relationship may reach an AGDA™ assessment. We do not implement what we recommend.
-
Documented methodology.
Theses, catalogues, corpus governance and rater protocols are the documented basis of the method. They are not published here, and are provided under NDA during diligence. AGDA™ is not validated against real-world outcomes.
-
Verifiable records.
Signed SEDI result records, where supplied, carry an Ed25519 attestation over specified records. Inputs and outputs are hashed. An audit committee can check them with the verifier and trust material supplied under the engagement terms.
-
Refusal of false control.
If the evidence does not support the conclusion, the assessment does not reach it. Findings the client did not want, in writing, before the regulator does.
Where the category is most urgent.
Our initial commercial focus is large UK and European regulated financial institutions. The sectors below are where the question is most urgent; some we deliberately do not serve.
-
UK-regulated financial services.
Banks, asset managers, insurers under PRA / FCA jurisdiction with material automated decision systems.
-
EU high-risk AI operators.
Annex III high-risk systems under the EU AI Act, where human oversight must be capable of intervention.
-
Critical national infrastructure.
Energy, water, telecoms, transport. Operators where intervention failure compounds across dependency chains.
-
Healthcare systems.
Clinical decision support and diagnostic AI. By invitation, where independence from device vendors can be guaranteed.
-
Defence & intelligence.
Autonomous decision systems and human-on-the-loop frameworks. With appropriate clearances and engagement letters.
Consumer internet, enterprise productivity, B2B SaaS, advertising tech. Not because the work is unimportant, but because Intervene's methodology, corpus governance and rater protocol are designed for regulated, board-accountable systems.
What Intervene does.
The practice in four statements. For the product's own boundaries, see what AGDA™ is not.
-
Assesses one system at a time.
A named consequential AI-enabled system, a defined decision or action pathway and a specified intervention scenario. The conclusion applies to those conditions and says so.
-
Works from evidence, not assertion.
Records, interviews and walkthroughs across Detect, Escalate, Decide and Intervene. Documented arrangements are distinguished from exercised capability; a policy fallback is not a measurement.
-
Owns the judgement.
Facts can be corrected and evidence can change the conclusion. The conclusion is not negotiated, and issuance requires approval separate from the assessor's own judgement.
-
Stops at the finding.
Remediation, control build, integration and policy drafting are the client's or another provider's work. Keeping them separate is what lets the finding be trusted.
The question returns to the same place.
Can you stop it in time? Name the system you are accountable for and we will tell you whether an assessment would answer it. Twenty minutes, at no charge. The sample verdict is there if you would rather see the evidence format first.