AI governance is a business control, not a slogan
Artificial intelligence can summarise records, screen job applicants, help underwrite finance, inspect industrial processes and produce marketing copy. Those uses raise very different legal risks. An internal writing assistant may chiefly threaten confidentiality, while an automated credit or employment decision may affect a person's rights or livelihood. A useful governance programme must therefore identify actual systems, users, datasets, suppliers and decisions. It should not represent voluntary good practice as already compulsory under a law that has not been enacted. Equally, the absence of a comprehensive AI-specific code does not suspend established obligations under existing Pakistani law.
Present law versus emerging policy
The National AI Policy 2025 establishes objectives for adoption, capability-building, responsible innovation and human oversight. The Digital Nation Pakistan Act, 2025, establishes the Pakistan Digital Authority within a wider digital-governance structure. The Islamabad AI Declaration announced in 2026 describes sovereign and accountable implementation. Each is important, but their legal forms differ. A business should check whether a particular requirement arises from an enacted Act, valid regulation, regulator direction, licence or contract rather than treating all policy announcements as directly enforceable. Regulated financial, communications and healthcare entities must continue to consider their sector-specific obligations.
Inventory every use case and appoint an owner
The first control is a complete internal inventory of AI usage. Record the supplier, model or service, department, purpose, data categories, output and person responsible for reviewing it. Identify any connections to customer systems, payment controls, employee records or confidential repositories. The inventory should include external AI tools purchased by individual teams, not only centrally procured enterprise systems. Each significant use case needs a business owner and someone responsible for legal and security review. A board committee may set the policy and escalation criteria, but accountability for a particular deployment should be explicit rather than dispersed among several unnamed departments.
Classify risk by impact rather than marketing labels
Risk assessment should examine who may be harmed, how reversible the outcome is, whether sensitive data is involved and whether a human can independently check the model's reasoning. A low-risk productivity task should not require the same validation as automated denial of insurance, employment screening, medical recommendations or fraud reports to law enforcement. Useful questions include whether outputs may be biased, fabricated, insecure, manipulated or inconsistent with the applicable contractual purpose. A documented assessment provides an evidence-based reason for chosen controls. It should be updated when a model, dataset or connected workflow changes materially.
Protect client and employee data
Prompts may disclose trade secrets, customer histories, source code, patient records or privileged communications to an outside provider. Before using the service, determine lawful access, contractual restrictions, permitted processing purposes, security safeguards, retention, data location and onward transfers. A privacy notice is not a universal permission to reuse confidential records for model training. Although Pakistan's comprehensive personal-data protection proposals should not be mistaken for an enacted statute, duties may arise from existing constitutional, contractual and sectoral frameworks. Cross-border businesses should also assess overseas data rules where those laws apply. Access limitation and approved tools should be supported by technical settings.
Negotiate AI supplier agreements carefully
The contract should specify what the product actually performs, ownership of input and output, confidentiality, authorised training uses, security standards, incident notice, sub-processors and disposal of customer material. Model changes can affect outputs without changing the visible interface, so the agreement should address version management, testing and notice of material modifications. Where the system is essential, the customer needs business-continuity commitments and a practical exit route. Unlimited assurances of perfect AI accuracy may be unrealistic, but vendors can still give meaningful promises concerning security, permitted uses, service availability, audit cooperation and remedial steps after failure.
Address intellectual property and open-source models
Commercial AI work can involve proprietary source code, copyrighted reference materials, licensed datasets and open-source model weights. The rights attached to each component may differ. A generated image or paragraph should not automatically be described as free of infringement risk, and the customer's ownership of prompts does not always determine ownership of every output. Contracts should allocate rights in customer data, improvements, deliverables, fine-tuning arrangements and third-party components. Developers should preserve licence notices and maintain a record of dependency versions. Before selling or licensing a product, assess whether model and software licences permit the proposed distribution and commercial use.
Human oversight and evidence of decisions
A human reviewer should understand both the purpose and the limitations of the system. Merely having a person click approve does not establish meaningful oversight where they lack time, competence or information to question the outcome. For consequential recommendations, set clear review criteria and escalation mechanisms. Record the model version, relevant input, reasons for departure and material corrections where lawful and proportionate. Systems should be tested on representative cases for factual error, bias, harmful recommendations and prompt-injection vulnerabilities. An unexplained output should not be transformed into an accusation of misconduct or a denial of a customer right without adequate review.
Employee policy and professional standards
An employee policy should name authorised tools, prohibited data, access controls, review rules, client-disclosure requirements where appropriate and procedures for suspected errors. Staff using AI for legal drafting, financial reporting or other professional outputs must verify important facts and authorities. A chatbot can invent a case citation or plausible financial figure; fluency is not proof of reliability. Employee misconduct should be assessed against actual instructions, contracts, intent and consequences rather than treating every use of an unapproved tool as an automatic cybercrime. Training is most effective when employees have a safe, practical alternative to consumer services.
Cybersecurity and incident response
AI services introduce new attack surfaces including stolen API credentials, prompt injection, malicious files, exfiltration through connected plugins and synthetic impersonation. Security policies should define authentication, rate limits, logging, vendor approvals and access revocation. If an incident occurs, preserve original data and logs, contain access, establish affected systems and assess any relevant contractual, sectoral or criminal reporting duty. PECA and NCCIA may become relevant where a specific electronic crime is alleged, but a routine model error does not automatically constitute a criminal offence. Legal, technical, insurance and communications responses should be coordinated.
Regulated sectors and board oversight
Banks, securities intermediaries, insurers, telecom companies and healthcare providers may be subject to specific rules on outsourcing, recordkeeping, confidentiality, operational resilience and customer treatment. The existence of the Pakistan Digital Authority does not erase those responsibilities. Directors and senior managers should ensure important decisions are not delegated to a model without adequate accountability. For public-sector procurement, contracts should cover data governance, measurable quality, security, audit, lawful administrative decision-making and exit arrangements. Board reporting should focus on actual material exposures, incidents, vendor dependence and overdue remediation rather than merely the number of AI pilots.
A ninety-day implementation timetable
During the first month, identify tools and owners, restrict high-risk unapproved uses and map sensitive data. During the second month, review vendor terms, adopt a proportionate risk classification, define human approval gates and establish staff rules. During the third month, test significant applications, document acceptance criteria, train personnel and rehearse incident response. Records should include the use-case register, procurement assessments, incident log, policy exceptions and review decisions. Governance is a continuing process: update it whenever a model materially changes, a new regulator instrument takes effect or the business changes how outputs are used.
Legal sources and continued reading
The National AI Policy, Digital Nation Pakistan Act and PDA's official instruments page are essential starting points. See our related AI legal framework guide, deepfakes and digital evidence analysis and NCCIA and PECA guide. This is general information, not a legal opinion for an individual system.
General information only. No advice is given on any particular matter.