Key Takeaways
- →Fix a trust problem after release and you unwind work that was already built, tested and shipped.
- →In February 2024 Google paused Gemini’s image generation of people three weeks after launch, and said it would improve the feature significantly, with extensive testing, before turning it back on.
- →PRIME builds five checks into the pipeline: Purposeful Impact, Respect for Humans, Interpretable by Design, Mitigated by Engineering and Enforceable Accountability.
- →Two measures show whether it works: Intelligence Velocity, with a target under 72 hours for high-severity issues, and Principles-to-Product Coverage, with a target of 80% in the first quarter.
- →IBM’s 2025 survey of 915 executives found that the top quartile of AI ethics spending, as a share of AI spend, reported 30% higher operating profit attributable to AI than the bottom quartile.
On 23 February 2024, Google apologised for a feature it had launched three weeks earlier. Gemini’s new image generator had produced pictures of people, some of them, in Google’s words, “inaccurate or even offensive”, and the company had paused image generation of people while it worked on a fix. Prabhakar Raghavan, the senior vice president who wrote the post, said the tuning meant to show a range of people “failed to account for cases that should clearly not show a range”. Google said it would improve the feature significantly before turning it back on, through a process including “extensive testing”.
Whatever review ran before that launch, the problem surfaced after release, in public, and the fix was to switch the feature off. That is the rework tax. A responsible AI review held at the end of the pipeline finds problems when they are most expensive to fix. Checks built into the pipeline find them earlier, and PRIME is the framework I use to put them there.
What Is the Rework Tax?
The rework tax is the extra cost of fixing a trust problem after a model ships, compared with catching it while the model is being built. It is paid in retraining, retesting, redeploying and rewriting documentation, and in the trust of the people who met the problem first.
Software engineering learned this long before AI. A 2002 study prepared for NIST on the cost of inadequate software testing stated the rule plainly: “regardless of when an error is introduced it is always more costly to fix it downstream in the development process.” The reason it gave was rework. To fix an early error, the software built so far “has to be unraveled and rewritten”.
AI adds a twist: some of the costliest errors are choices made before any code is written. In 2019, researchers writing in Science examined a commercial algorithm that health systems used to find patients with complex needs. It predicted health care costs as a stand-in for illness. Because less money is spent caring for Black patients than for White patients, Black patients at a given risk score were considerably sicker, and remedying the disparity would raise the share of Black patients receiving additional help from 17.7% to 46.5%. By some measures of predictive accuracy the model looked sound. The flaw was the choice of what to predict.
A review before release sees only the top of this picture. A finding after release brings the rest with it.
Why Do Reviews Happen So Late?
Because many responsible AI programmes start as documents. A policy is written, an ethics board is formed, and a review step is added at the end of the release process, where it is easiest to bolt on. The pattern I see most often is a sound policy that engineers meet only when their work is finished.
Operating the principles is the hard part, and leaders say so. In PwC’s 2025 US Responsible AI survey of 310 business leaders, half named operationalisation, turning responsible AI principles into scalable, repeatable processes, as their biggest hurdle.
Regulation is moving the same way. Under Article 9 of the EU AI Act, the risk management system for a high-risk AI system must be “a continuous iterative process planned and run throughout the entire lifecycle” of that system. For high-risk systems listed in Annex III, Article 113 applies that requirement from 2 December 2027, the date set when the Digital Omnibus amended the Act. A single review at the end of the pipeline looks nothing like that process.
What Is the PRIME Framework?
PRIME is a framework for building responsible AI checks into the software pipeline instead of reviewing for them at the end. It has five dimensions, each mapped to one of the OECD AI Principles, and two measures that show whether the checks are working.
The OECD AI Principles were adopted in 2019 and updated in May 2024. PRIME takes their five values-based principles and turns each one into requirements that live in the ticket system and the build pipeline.
The PRIME Framework
Five dimensions, each built into the pipeline
OECD principle: Inclusive growth, sustainable development and well-being. Every AI feature starts with a benefit hypothesis and a map of the groups it affects, attached to the user story in the ticket system. Who benefits, and how will you measure it? The outcome becomes part of the definition of done. This is also where the target gets chosen. The health algorithm described above predicted cost, which the study’s authors describe as a “convenient, seemingly effective” kind of proxy. A benefit hypothesis naming the patients it was meant to serve would have asked whether cost stood in for need for all of them.
OECD principle: Human rights and democratic values, including fairness and privacy. Pre-register the contexts that trigger stricter controls: employment, health, credit and anything touching protected characteristics. Write down who decides what. Build opt-outs, correction channels and plain-language disclosures into the product from the first sprint, before any audit asks for them. The GDPR guide covers the legal detail for European users.
OECD principle: Transparency and explainability. Keep model cards, data sheets and the history of prompts and versions current as the system changes. Fit the explanation to the reader: a clinician needs the source passage, a customer needs plain language, a regulator needs an audit trail. All three can draw on the same records. For retrieval-augmented generation, The Hallucination Rate Your Team Quotes Is a Guess shows how to cite the passage behind each answer.
OECD principle: Robustness, security and safety. Run automated evaluations for hallucination, toxicity, prompt injection and jailbreaks on every release, with hard thresholds that block a deploy. A team might set its jailbreak success threshold at 1% or lower, for example. Bind kill switches to live metrics. Run privacy-leakage and licence checks in the pipeline instead of a quarterly audit. If your site reliability engineers can block a deploy for downtime risk, your AI pipeline should be able to block one for trust risk.
OECD principle: Accountability. Every AI feature has a named owner, a person rather than a team. Define the evidence trail before building: datasets, prompts, model versions, evaluation results and production logs. Encode policies as code so an audit can be rerun. Extend this to vendors, with evidence and incident-sharing clauses in the contract. The Founder’s Playbook adapts this for companies where the owner is usually the founder.
The name is deliberate: to prime something is to prepare it before use. The artifacts are small by design. A benefit hypothesis is a paragraph, and a fairness evaluation is a test suite. Minimum Viable Governance covers how to start governance without building a bureaucracy, and the Governance Playbook covers the wider organisation.
What Does PRIME Look Like in the Pipeline?
Each dimension lands on infrastructure most engineering organisations already run. The table maps PRIME to five surfaces, with what gets built and what gets measured at each. NIST’s Generative AI Profile, published in July 2024 as a companion to its AI Risk Management Framework, is a useful reference for every row.
PRIME across the stack
What gets built and what gets measured at each layer
| Surface | What gets built | What gets measured |
|---|---|---|
Governance | A policy-as-code service with machine-checkable rules. No deployment without provenance tags. | P2P Coverage for each PRIME dimension |
Data | Data contracts that carry provenance, licence and consent. Automated leakage tests and licence checks before any release. | Lineage completeness; leakage incident rate |
Model and agent | An approved catalogue of base models. One evaluation suite, with unit, scenario and adversarial tests for every release. | Intelligence Velocity; adversarial test pass rate |
Experience | “Why you are seeing this” disclosures, feedback channels, and explanations fitted to each audience. | Explanation coverage; user correction rate |
Operations | An AI incident registry that uses the categories of the OECD AI Incidents and Hazards Monitor. A weekly triage where engineering, product and risk meet together. | Incidents detected; time to mitigation |
None of these surfaces is new. PRIME adds intent and a measurement column to functions you already have. The AI ROI framework connects the measures to the numbers a CFO tracks. For the incident registry, the OECD AI Incidents and Hazards Monitor is a free public example of how to classify what goes wrong.
How Do You Know PRIME Is Working?
Two measures, kept few on purpose, so that people discuss them.
Intelligence Velocity is the time from discovering a new risk, such as a jailbreak method, a bias drift or a data-leakage signal, to mitigating it everywhere it applies: guardrails updated, tests added, models redeployed. PRIME’s target is under 72 hours for high-severity issues. A responsible AI process that runs on a quarterly review cycle cannot get near it.
Principles-to-Product Coverage (P2P) is the share of shipped AI features with all five PRIME artifacts complete: a benefit hypothesis, a fairness evaluation, explanation UX, safety gates and an audit trail. PRIME’s target is 80% in the first quarter, which is reachable because the artifacts are small.
Both targets are design goals set by the framework. Neither is a benchmark drawn from client data.
How Do You Start?
Pick one revenue-adjacent generative AI use case, the one your chief executive mentions to investors, and run PRIME on it for 90 days.
PRIME in 90 days
From one use case to a decision on scale
Month 1: Baseline
Adopt the OECD AI Principles as your standard and map them to the NIST Generative AI Profile, so you start from an existing catalogue of controls. Build the minimum control book for the chosen use case and wire its evidence checks into the build pipeline. At this stage the gates warn and do not block. The habit comes before the enforcement.
Month 2: Instrument
Build or extend the evaluation suite for hallucination, toxicity and jailbreaks. Stand up the incident registry. Start a thirty-minute weekly triage with engineering, product and risk. Switch the gates from warn to block for high-severity issues, and record your first Intelligence Velocity number.
Month 3: Decide whether to scale
Track both measures. If Intelligence Velocity is under 72 hours for high-severity issues and P2P Coverage is above 80% for the pilot, extend PRIME to a second use case. Write a one-page EU AI Act readiness note: your hypothesis on the risk category, plus the evidence the pipeline already collects.
Expect the first month to be uncomfortable. Writing the control book is when a team finds the gaps nobody knew were there, often in where its data came from. Those gaps were going to surface anyway, later and at a higher price.
Apply the rework test. List the last five trust findings on your AI systems: bias, privacy, safety, explanation or ownership. For each one, note when it was found and the earliest stage that could have caught it. If most were found after release, you are paying the rework tax, and the earliest stage in your list is where the first PRIME check belongs.
Where Does PRIME Fall Short?
Research and exploration. Prototypes, experiments and spikes should not carry PRIME overhead. PRIME starts when code moves toward production. The Hidden Tax on AI Speed covers the line between exploration and production work.
Very small teams with high trust. Three senior engineers who already review for fairness, write down their reasoning and own their decisions may not need the formal structure. What they need is to notice when the team grows past the point where informal governance works.
Organisations with no AI in production. Start with the five-pillar readiness assessment. Before the first deployment, data, people and leadership matter more than pipeline gates.
How Is PRIME Different From HAX and SAIF?
Teams that know the field ask why they should not use Microsoft’s HAX Toolkit or Google’s SAIF. Both are good, and both solve different problems.
HAX is for designing AI user experiences. Its 18 Guidelines for Human-AI Interaction, a synthesis of more than 20 years of research introduced in a 2019 CHI paper, cover the moments of use: first contact, during interaction, when the AI is wrong, and over time. Guideline 6 is to mitigate social biases. A workbook helps a team choose which guidelines to apply. It guides product design and does not run in a build pipeline.
SAIF is for securing AI systems. Its risk map names 15 risks, from data poisoning to prompt injection, with controls for each, and Google extended it to SAIF 2.0 in October 2025 with a risk map for AI agents. All 15 are security risks; fairness and explanation are not on the list.
PRIME, HAX and SAIF
Three frameworks for three different jobs
| Question | PRIME | Microsoft HAX | Google SAIF |
|---|---|---|---|
What is it for? | Building trust checks into the software pipeline | Designing AI user experiences | Securing AI systems and agents |
What does it give you? | Five dimensions, pipeline gates and two measures | 18 guidelines, a design library and a workbook | A map of 15 risks with controls, and an agent risk map |
Fairness and bias | A fairness evaluation for every shipped feature | Guideline 6: mitigate social biases | Not among its 15 risks |
Security | Jailbreak and prompt-injection gates on every release | Not its subject | Its core subject |
How progress is tracked | Intelligence Velocity and P2P Coverage | The workbook’s priority list | An agent risk self-assessment |
They work best as layers. HAX should shape the user-facing decisions that PRIME’s Interpretable by Design dimension then records in the backlog. SAIF should shape the threat model that PRIME’s Mitigated by Engineering dimension turns into release gates. PRIME’s job is the connecting part: making both enforceable on every release.
Does Building Trust In Pay?
Every study below favours building trust in. None of them proves cause.
- In a survey of 915 executives published in October 2025, IBM’s Institute for Business Value found that organisations in the top quartile of AI ethics spending, as a share of AI spend, showed 30% higher operating profit attributable to AI than the bottom quartile over the previous two years.
- EY’s Responsible AI Pulse survey of 975 C-suite leaders, released in October 2025, found that respondents with real-time monitoring were 34% more likely to see improvements in revenue growth and 65% more likely to see improved cost savings. Real-time monitoring is what a pipeline check becomes once a system is live.
- PwC built a model instead of running a survey. Published in August 2025, it simulated tens of thousands of scenarios comparing companies that spend just enough to meet compliance requirements with companies that spend an extra 10% of their AI budget on responsible AI. In the simulation, the second group’s valuations were up to 4% higher and their revenues up to 3.5% higher.
Studies like these cannot separate cause from effect, since profitable firms may simply spend more on ethics. The argument does not rest on them. It rests on the old rule from software engineering: what you catch early costs less to fix. The Consortium for Information & Software Quality put the cost of poor software quality in the US in 2022 at no less than $2.41 trillion. AI systems are software, and they inherit the rule.
The rework tax is paid quietly, a sprint at a time, until a failure makes it public. The cheapest place to stop paying it is the build pipeline, where a failed trust check stops a release the same way a failed security test does.
Download: PRIME Framework Implementation Worksheet
Get the complete PRIME worksheet: dimension assessment checklists, stack mapping, Intelligence Velocity and P2P Coverage baselines, 90-day sprint planner, and existing tools audit, ready to print or save as PDF.
Enter your email to get instant access. You'll also get The Brief, every other Tuesday.
Free. No spam. Unsubscribe anytime.
Get The Brief
Original thinking on AI strategy and governance. Every other Tuesday.
Senior AI strategist helping leaders make AI real across four continents. Forbes Technology Council member, IEEE Senior Member.
Ajay's views, from fifteen-plus years in the field. Not legal or compliance advice. See full disclaimers →
Published by AI Exponent LLC
