Tips for evaluating AI proposals
| Who | Procurement teams |
| What | Best practices for evaluation |
Your evaluation methods will vary depending on what and how you are buying. As you evaluate proposals, keep your end user and/or beneficiary in mind based on your organization’s user research. Make sure that your evaluation criteria and selection process speak to their needs. Here are some tips to consider.
General good practices
Lifecycle and vendor management
- Plan for ongoing model maintenance, updates, and version control
- Confirm vendor commitments to transparency about changes and potential impacts
- Include rollback and incident response procedures in contracts
Risk and compliance
- Ensure adherence to applicable AI regulations and standards in your jurisdiction
- Establish explainability requirements aligned with use case needs
- Define data governance policies covering data sharing, retention, and deletion
- Assess risks related to third-party dependencies and supply chains
Documentation and governance
- Maintain detailed records of evaluation criteria, scoring, and decisions
- Require vendors to disclose model limitations and ethical considerations
- Embed responsible AI requirements into procurement contracts
Additional considerations
- Compatibility with open source AI models or platforms, where appropriate
- Support for data interoperability and integration with existing public sector systems, tools, and data repositories
- Transparency and communication plans for public-facing AI applications
- Preparedness for public records requests or transparency laws impacting AI outputs
- Strategies to mitigate risks related to hallucinations, misinformation, or misuse
- Clearly define support and escalation pathways for incidents
- If appropriate and legally feasible given your context, consider involving end-users/beneficiaries themselves in the evaluation process.
Best practices by contract method
| Method | Best practices |
|---|---|
| Open solicitation | Evaluate AI-specific technical criteria
Understand technology and vendor transparency
Review vendor and solution track record
Pricing and commercial terms
Security, privacy, and compliance
Responsible AI and ethical considerations
|
| Innovation procurement | Include all above points, plus:
|
| Framework agreement (multiple awards) or under-threshold purchase | Focus on essential due diligence:
|
Evaluation criteria
As you create your evaluation criteria, align around your metrics, what the terms really mean, and the implications. For example, “accuracy” often comes up during evaluation. But other metrics matter as well, such as precision (which is also sometimes confused with accuracy by non-experts) and recall. In some use cases, precision and recall are trade-offs as parts of the overall accuracy of the solution. Speak to technical experts to understand the implications of these metrics in terms of false positives/negatives, and discuss relevant operational and legal implications.
Your evaluation criteria will be case-dependent. Some evaluation criteria for open solicitations you may consider include:
- Problem fit, including demonstrating a strong understanding of your need and the appropriateness of the solution.
- AI approach and explainability, such as model algorithm and training data, model transparency, algorithmic clarity, audit readiness.
- Environmental impact, such as energy use and source.
- Data governance and bias handling, including strategies for fairness, data quality, and addressing known model limitations.
- Integration, training and operations, such as fit with existing infrastructure, plans for training and upskilling internal staff, and ongoing support services.
- Risk and accountability, like liability mechanisms, governance structures, and human-in-the-loop system design.
- Financial value, such as cost over the lifecycle and real ROI expectations.
- Vendor experience, including relevant experience with similar AI solutions.