AI adoption

The main challenges in adopting AI, and how we help with them.

Most organisations can build an AI pilot. Fewer can show a regulator, an auditor or a customer how it is controlled. We help you work out the risks, put the controls in place and, where needed, build the system with those controls included.

The challenges

Six places AI projects get stuck.

Risk, control and regulatory oversight come up in nearly every one. Each row says what goes wrong and what we do about it.

ChallengeWhat goes wrongHow we help
Wrong, biased or manipulated output

AI can produce confident errors, treat people differently because of a name or background, or be steered by instructions hidden in a document it reads. Nobody notices until a customer or a regulator does.

We decide where a person must make the decision, test for bias and hidden instructions before release, and set up monitoring after it.

Unclear regulatory obligations

The EU AI Act, GDPR, DORA and sector rules can all apply to one use case. Firms often cannot say which apply, or what evidence each one expects.

We list your AI use cases, work out which rules apply to each, and set out the evidence you need to hold.

No owner and no approval route

Pilots start in business teams without a named owner, a risk sign-off or a policy. By the time risk is asked about, the system is already in use.

We write the AI policy, roles and approval process, and set up a register of AI use cases with an owner for each.

Dependence on AI providers

Most AI features run on a third party's model. That raises questions about contracts, data protection, concentration and exit. Under DORA, the provider is generally an ICT third-party service provider.

We assess providers against DORA's third-party requirements, review contract terms, and design systems so the model can be swapped.

Personal and confidential data

Documents sent to a model may contain personal data. Where it goes, who can see it and how long it is kept all have to be known and documented.

We map the data flows, advise on lawful basis and retention, and build systems that keep each customer's data in its own database.

Moving from pilot to production

Pilots run on clean examples. In production, costs, edge cases and failures appear, and there is no record of what the AI did or who approved the result.

We set spend limits, evaluation tests and monitoring before launch, and log each AI call and decision in an audit trail that cannot be edited.

Regulation

The rules that can apply to one AI use case.

Which rules apply depends on what the AI does, who uses it and where you operate. We work this out for each use case.

EU AI Act

Regulation (EU) 2024/1689 sorts AI systems by risk. Uses such as screening job applicants or assessing creditworthiness are listed as high risk, with duties for risk management, data quality, logging, human oversight and documentation. Organisations using AI must also ensure staff have enough AI literacy. Duties phase in over several years and the timetable has been under review, so confirm current dates.

GDPR

Applies whenever personal data goes into an AI system. Article 22 limits decisions based solely on automated processing that significantly affect people, so a human review step matters.

DORA

For financial entities, an AI model supplied by a third party is generally an ICT service. It falls under the register of information, the contract requirements and exit planning in Chapter V.

Sector regulators

Regulators in sectors such as financial services, where the Central Bank of Ireland supervises, expect existing governance, outsourcing and operational-risk rules to be applied to AI as they are to other technology.

General information, not legal advice. Obligations depend on your role (provider or deployer) and your use case.

From risk to control

Each risk gets a control, and each control leaves evidence.

These are the controls we build into our own AI systems and recommend for yours.

RiskControlEvidence you keep
Biased scoring

We submit pairs of inputs that differ only in a name or another protected detail, and compare the score gap with the model's normal run-to-run variation.

Bias test report

Instructions hidden in documents

A CV or job description can contain text telling the AI what score to give. We test for this on each model before release.

Hidden-instruction test log

Decisions with no human oversight

The AI ranks and explains. A person makes every decision, such as moving a candidate to the next stage or answering a regulatory question.

Approvals recorded in the audit log

Runaway cost

Every AI call is logged with its estimated cost and blocked once the customer's budget is used.

Usage and cost log

Dependence on one model

The same test suite runs on several AI models, so a change of model is decided on measured results.

Model comparison results

Data exposure

Each customer's data is held in its own database, and code analysis, dependency audit, secret scanning and an attack scan run before every release.

Security scan reports

How we help

Four ways we work with you on AI.

AI risk assessment

List every AI use case in your organisation, including AI tools staff use without approval. Classify each under the EU AI Act and assess the controls around it.

You receive
AI use-case registerRisk classificationControl gap list

AI governance

Write the policy, assign owners, define approval and review steps, and set up the evidence you will need to show supervisors and auditors. We can map this to ISO/IEC 42001 and the NIST AI Risk Management Framework.

You receive
AI policyRoles and approval processEvidence index

Testing and assurance

Bias, consistency and hidden-instruction tests on AI you build or buy, with a report you can show a regulator or a customer.

You receive
Bias test reportHidden-instruction test logSupplier assessment

Build with the controls included

When you need a system, we build it with human decision points, logging, spend limits and testing from the start, instead of adding them afterwards. Bespoke software

You receive
Working applicationAudit logTest and scan reports

Our own example

We apply this to our own recruitment platform.

The platform ranks job applicants. The EU AI Act lists recruitment and selection tools as high risk, so a person makes every decision, we test for bias and hidden instructions, and every AI call and stage change is logged.

See the platform

Ranked matchesSynthetic CVs · fictional agency
Ranked candidate matches showing matching skills, gaps and a score breakdown for each person.

Contact

Get in touch.

Tell us about the system you want built, or the regulation you need to meet. We will reply with a few questions.

contact@echogate.ie
Open in your email app