AI Policy

PCV Clients, Customers, Colleagues, Industry Partners, and Supporters, 

Human civilization has entered a new era, the 4th industrial revolution, of technological innovation as artificial intelligence (AI) and machine learning (ML) bring the power and potential to streamline our business operations, generate ideas and content, and process vast information at speeds beyond human capacity. 

PCV has long been a data-driven, “high tech and high-touch” organization founded as one of the country’s first impact investing organizations. PCV is dedicated to investing in small business entrepreneurs’ passion and resilience, helping them create good quality jobs and advance economic mobility, climate resilience, and financial wellbeing outcomes, while uplifting communities that have been historically underserved. PCV is also a federally and state-certified community development financial institution (CDFI), representing a field born out of the Civil Rights movement as its economic justice pillar, more than 30 years ago following a federal acknowledgment of historical redlining in the formal financial industry. 

We acknowledge the potential ethical issues that AI poses, and the impact, both positive and negative, that it can have on our employees, clients, customers, and communities. We are committed to continually researching and understanding the algorithmic bias perpetuated by AI as its capabilities advance, particularly in sensitive industries related to ours – such as financial decisions, hiring, and criminal justice. 

Therefore, we are developing, implementing, and regularly updating a comprehensive policy around the ethical use of AI and ML that aligns with our company’s values, mission, and goals. We believe that decisions to leverage any AI or ML application, tool, methodology, or system must uphold individual and community consent, data autonomy and privacy, and respect for human agency and dignity, while minimizing adverse climate impacts. 

Therefore, in the use of AI and ML technologies, PCV strictly adheres to the following six core principles: 

1. PCV must ensure that the benefits of the AI used minimize the harm created, and it must not harm any particular group or individuals disproportionately.
2. PCV must select or develop AI that has been trained on demographically disaggregated data to prevent discrimination or disproportionate harm on any person or group of persons by the use of that AI system. We acknowledge that algorithms trained without proportionally balanced demographics nearly always perform worse for low-income, underrepresented, or underserved groups.
3. The AI system must never be used to discriminate against or exclude individual clients, customers, or groups of people during the course of business. AI systems include, but are not limited to, predictive models, matching algorithms, support desk chat bots, recommendation engines, auto-generated content for marketing and communications, tools, and scripts to automate administrative tasks (e.g., like finance, loan processing, and reporting) and others.
4. AI systems used at PCV must at all times prioritize and safeguard clients’ and customers’ privacy and data rights, both, to protect data from abuse or monopolization and to ensure we are abiding by loan-level privacy legislation as we are obligated as a CDFI. Clients include small business entrepreneurs and their workers. Customers include other CDFIs or mission-driven organizations that PCV supports in its role as a capacity builder in data, analytics, ML, and AI. PCV commits to strict protection of customer and client data, at all times, maintaining this data within a firewalled environment separated from other program data. Clients and customers shall maintain the right to request access to their data and remove it from PCV’s systems. PCV will undertake the best efforts to fulfill such a request. 
5. The AI system should be used to inform and support human critical analysis and decision-making and should never replace human decision-making, nor be unsupervised in any scenario, nor act as a fully autonomous system.  An AI system must always be designed to keep a human in the decision-making loop.  TAs AI and ML systems become increasingly autonomous and agentic; the following examples are not exhaustive but intend to guide PCV staff: 
   • Coding agents (Claude Code, Copilot, Antigravity, Jules, etc.): heavy and unsupervised reliance on coding agents is forbidden at PCV. Coding agents may be leveraged during algorithm-building, assembling data engineering workflows, software development, testing, and deployment, only as checks on human-authored code, as tools to assist troubleshooting, as aids to generate test scripts, and as guides to advise on efficient deployment. Coding agents must never be activated on tasks that generate new, unsupervised code, access data,  folders, databases, or other repositories without explicit human permission, scrape any preexisting content without explicit human permission, or be allowed to autonomously run scripts or actions of any kind without a two-person review, testing, and formal move of the codebase into production. At any point of a technical project lifecycle, a developer must be able to explain one or more lines of code when questioned. An inability to provide a rationale or appropriate levels of transparency by a technical human resource will be assumed as an indication of unauthorized AI use.   
   • LLM-assisted writing: fully autonomous writing (for example, leveraging the outputs of OpenAI, Claude, and other LLMs, and representing them as one’s own writing) is strictly forbidden at PCV. Staff may use such tools to support the writing process – such as to brainstorm new program ideas, conduct research on primary sources, assist with an outline for a well-structured proposal, conduct grammar and spell-checking on final content, and similar light editing functions, but the actual written content must be one’s own writing and represent one’s own original thinking. When presenting a final deliverable to one’s manager or other stakeholder, PCV staff must disclose AI tool use either verbally, in email, or as a footnote on the writeup. The disclosure should include the AI tool used, the approximate dates and activities for which it was used. 
   • AI-generated collateral: fully autonomous content creation (for example, websites, slides, logos, graphics, social media posts, photos, videos, landing pages, style guides, user interfaces, user experience workflows, and branding collateral, as a few examples) without thorough human review and without proper attribution, is strictly forbidden at PCV. Staff may use tools such as Base44, Miro, Figma, Canva, Framer, Tailwind – and others – to generate prototypes whose purpose is to communicate their ideas and vision, assist them and their colleagues with brainstorming, and communicate final specifications. Such AI-generated collateral, however, should never be considered final deliverables, nor be pushed into production, without a thorough code review (for example, new landing pages, web-based tools), and management sign-off to ensure brand consistency, security, and intellectual property considerations.  
6. PCV will not use autonomous AI to make lending decisions or other decisions of consequence without a change to this policy and CEO approval. Consequential decisions include but are not limited to the following: 
   • Determining who will receive and not receive a loan. 
   • Determining who will receive and not receive business advisory support.
   • Determining who we will inform of our products, services, and new opportunities, such as participating in a study, a new pilot program, or incentives of any kind. 
   • If a particular AI-supported decision poses any doubt, PCV will ensure that no group is barred from a PCV product or service without citing specific programmatic or loan eligibility criteria or legal rationale. Always ensure that no information is withheld from a particular group or groups, again, without a previously communicated programmatic or loan eligibility criteria or legal rationale. 

Because algorithms are trained on large data sets of lending, hiring, and criminal justice decisions that often employ a definition of “risk” that has historically excluded low-income and underserved communities, PCV has to be particularly intentional about the use cases and guardrails puts in place to test them within PCV and our community finance industry. We recognize that these new technologies can easily reverse the gains made to date with PCV’s high-tech high-touch approach and risk accelerating bias in community finance field at a scale and speed that we may never be able to address before adverse impacts get deeply reinforced. 

PCV recognizes that what we do as a CDFI, an impact investor, a community development organization – in terms of how we deploy AI now –may set a precedent for the sector for years to come. As we increase our technology and AI capacities within PCV, we commit to sharing learnings with the CDFI and impact ecosystem, advocate for evidence-backed practices that center clients and community voices, protect against bias, curtail market drift, prevent exclusion, and ensure we seek the explicit consent of those we and our peers serve. 

When collaborating with peer CDFIs and other mission driven organizations, we must share our values and the parameters that we consider to be acceptable uses of AI (detailed below). If during the course of the collaboration, we observe divergence in these values and parameters, we will exercise the right to dissolve the collaboration. For those organizations we secure alignment with, we will work to build a community of practice that can help propagate AI standards that advance human dignity for all, especially underserved populations. 

More specific guidelines follow:

1. Always credit AI use: Whenever AI assists in the creation of content to co-author a piece of writing or to wholly create the content, the employee must ensure that proper credit is given to the AI system or tool used. This must be done by adding a footnote or other indications in the document or publication.
2. Ensure data privacy: PCV recognizes the importance of protecting sensitive customer data and confidential company information when feeding material into AI applications. Before using any AI tool, the employee must review the accompanying language on data security and obtain approval from the Chief Data Officer and CEO before initiating use (see below). PCV will ensure that appropriate security measures are in place, such as data anonymization, encryption, plug-in tools, and access controls (to mitigate third-party tools), to prevent unauthorized access or use of such information.
3. Exercise caution in decision-making: PCV recognizes that AI can automate standard processes that lead to decision-making. However, we must exercise caution and ensure that human oversight is an intentional part of the process to avoid biased or discriminatory decisions. PCV and its employees commit to regularly reviewing the AI systems and tools used to ensure their accuracy, fairness, and transparency. PCV will provide training to employees on the ethical use of AI – to ensure it is leveraged to inform and support human decision-making, not displace it.
4. Partner with trusted institutions: PCV’s senior leadership is working to establish partnerships with trusted institutions, such as non-profit organizations, industry associations, and research institutions, who have expertise in the proper and ethical use of AI. We will regularly seek their guidance on ethical AI practices and incorporate their recommendations into our policies and procedures. (For example, Code for America, Black Wealth Data Center, Data Equity, etc.)
5. Uphold transparency: PCV and its employees must ensure that the use of AI is transparent to our stakeholders, clients, customers, and communities. We will clearly communicate the intended use of AI systems and tools, their limitations, and any potential risks to stakeholders. We will also provide an avenue for feedback and concerns related to AI use.
6. Commit to continuous improvement: PCV will regularly review and update our policy on the ethical use of AI to reflect changing trends, best practices, and emerging risks. We will ensure that our policy aligns with industry standards, regulatory requirements, and our company’s values and mission. 
7. Measuring Bias: PCV will use household income, ZIP code designations, and demographic data to measure and mitigate potential bias. 
8. Data Revamp Working Group:  PCV will apply the same data governance practices to machine learning methodologies and generative AI tools that it does with vendor platforms like Lenderfit, Downhome, and/or Qooper. Any concerns regarding bias, security, privacy, or mission drift will be surfaced in regularly scheduled internal Data Governance Working Group meetings to discuss and decide with internal stakeholders. Any PCV staff may also raise concerns directly with PCV’s Chief Data Officer and/or CEO at any time.
9. Approvals for New Tools: Staff field-testing new AI and ML tools need to submit a request for permission to the Chief Data Officer and CEO, clarify the use case they want to test, what the intended result or value-add will be to PCV, and how they are adhering to the above guidelines. They also need to clarify this policy with any external consultants, incorporate into external MOUs, and receive a signature of confirmation. If new tools carry additional cost, the staff must research all cost implications, identify the tier or package that is most appropriate for tool use, and obtain explicit permission from the CFO, CDO, and/or CEO before initiating use. Every additional user or license that is requested requires approval through the same channels. The staff that proposes the new tool is also responsible for identifying any changes to this AI Policy that may arise as a result of using the tool. The new policy guidelines must be discussed preferably at the Data Governance Working Group forum, or at the very least with the Chief Data Officer for approval and inclusion in the policy. 
10. AI tools for meeting transcription, recording of virtual conference calls:PCV permits its staff to use a company-approved virtual call transcriber/recorder only under the following conditions:
    • Staff must only use PCV-approved applications (for example: Fathom, as of January 2026). Fathom, at the time of writing, is the one service that stands out in terms of its privacy and security features. It is Soc 2 Type II, HIPAA, CDPR, and CCPA compliant. It also has a feature that collects electronic consent from meeting attendees prior to the meeting before a recording is authorized. If staff wish to propose an alternate solution, it must meet these standards at a minimum. The solution must also be compatible with a Microsoft environment and must receive explicit approval from the Chief Data Officer and CEO. 
    • Staff must collect explicit consent from every participant on the virtual call, including all internal or external stakeholders prior to the start of recording and/or transcribing. If someone objects, staff will turn the recorder off and discuss other means to collect the information needed. 
    • Staff must disable all automated features in the recording application that a) involve a recorder entering a meeting without explicit permission provided by a human at its start, and /or b) automatically sending the recording and transcripts to all meeting participants at the close of the meeting. The recording must be sent manually to participants who request it, and only those who attended the original meeting can receive access to the recording by PCV staff.  
    • Staff must maintain the highest level of security of each recording output and treat each recording as a PII data artifact. This means that the staff member who initiated the recording is by default the only person who has access to the recording, and s/he provides access to others based on the principle of least privilege, determined first and foremost by internal staff who were present at the meeting at the time of recording, and secondly by external stakeholders who were present at the meeting. If any other person requests access to the meeting recording, PCV staff must obtain permission from their manager and/or the Chief Data Officer prior to sharing. 
    • By default, no recorders or transcribers will be permitted in internal PCV calls unless determined ahead of time and agreed upon by all participants. 
    • PCV staff must exercise caution and vigilance in calls with all external parties. If there is a chance – however small – that sensitive information will be shared during the call, PCV staff must request that the counterparty deactivate its recorders and notetakers if a mutual NDA (MNDA) is not already in place. If an MNDA is in place, the counterparty can proceed with the recording, provided that the PCV team clarifies that the recording will only be accessible to participants in the call. If an MNDA is not in place, and if the counterparty insists on recording, the meeting can proceed only with the caveat provided by PCV staff that no sensitive information will be shared, and the discussion will be limited to generalities. Sensitive information includes any client and/or customer PII, details on PCV’s programs that are not publicly available, and any proprietary methodology, tool, application, technology, and/or process developed by PCV. 
11. Usage of AI tools for Auto-Completion: AI tools are increasingly becoming integrated into day-to-day applications and software such as email, Microsoft Office applications, and coding platforms. Auto-completion features, or “intelligent recommendations,” now accompany most applications and are intended to create efficiency. PCV employees are permitted to utilize auto-complete functionality in Word, Excel, PowerPoint, email, and other frequently used applications but must review the auto-completed content for accuracy. Auto-completed code (for example: while writing Python or R while performing data analysis or building algorithms) is discouraged, as the auto-generated code often suggests shortcuts, does not fully comprehend mission-based use cases, and creates a series of steps whose logic can be difficult to interpret during code reviews. If code recommendation tools are used, the developer must check all logic, syntax, and outputs for accuracy, AND provide comments in the code where auto-complete was used. In cases where inaccurate, misleading, or “hallucinated” content is delivered – whether to internal or external stakeholders – blaming auto-completion tools or AI recommendations will not be tolerated. All PCV employees take personal responsibility for the content they as individuals, with or without auto-completion, produce. 
12. Balancing AI utilization with climate impacts: PCV staff must use AI judiciously, always being mindful of its climate impact. A good rule of thumb is to assume that a typical working session with ChatGPT (involving between 20-50 prompts) uses 16 ounces of water, or the amount in a standard water bottle one buys from a convenience store. Before a tool like ChatGPT, Claude, or other, is used, staff must always consider whether it is worth essentially pouring a bottle of water down the drain. A few other considerations: 
    • When PCV employees deploy their AI-supported tools (e.g., AIKKA) and build machine learning algorithms, they must analyze the trade-offs of performing multiple runs versus optimizing for accuracy. For example, if multiple additional runs (for example, for hyperparameter tuning) are needed to increase predictive accuracy by 3% from 92% to 95% for a machine learning algorithm, employees should not undertake the additional processing for the energy it would expend for limited benefit. 
    • When PCV selects systems and technology vendors, it will choose those who have demonstrated a commitment to sustainability in their data center infrastructure. 

By adopting this policy, PCV and its employees aim to promote the ethical use of AI and to build trust among all our stakeholders. We believe that responsible and ethical AI use generate value for our customers, bring efficiencies to our workflows, and contribute to the innovation, growth, and outcomes for underserved communities. 

PCV staff referenced these frameworks, articles, and experts, in the creation of this policy: 

• “AI Is Fundamentally Incompatible with Civil Rights”, Dr. Vivienne Ming, Socos Labs from The Urban Institute, 2017, https://www.youtube.com/watch?v=Cm7IJiokqz8
• “This Is Not the Industrial Revolution”, Dr. Vivienne Ming, Socos Academy, 2018, https://academy.socos.org/not-the-ir/
• “We don’t need an AI manifesto — we need a constitution”, Dr. Vivienne Ming, Financial Times, https://www.ft.com/content/b16fab3e-7f19-49ab-9bbb-9bfeccbaf063
• “Technologist Vivienne Ming: AI is a human right”, Dr. Vivienne Ming, The Guardian, 2018, https://www.theguardian.com/technology/2018/dec/07/technologist-vivienne-ming-ai-inequality-silicon-valley
• “Human at the Helm: Build Trust in AI with a Human Touch”, Salesforce Office of Ethical and Humane Use, 2024, https://humanatthehelm.splashthat.com/
• “The Data Equity Framework”, We All Count, 2022, https://weallcount.com/the-data-process/
• “Responsible Use of Technology”, World Economic Forum, 2019, https://www3.weforum.org/docs/WEF_Responsible_Use_of_Technology.pdf
• “The Fourth Industrial Revolution: what it means, how to respond”, Klaus Schwab, World Economic Forum, https://www.weforum.org/stories/2016/01/the-fourth-industrial-revolution-what-it-means-and-how-to-respond/
• “Blueprint for an AI Bill of Rights”, The White House, 2023, https://www.whitehouse.gov/ostp/ai-bill-of-rights/
• “AI Risk Management Framework”, National Institute of Standards and Technology, 2023, https://www.nist.gov/itl/ai-risk-management-framework
• “Voluntary Commitments by Microsoft to Advance Responsible AI Innovation”, Microsoft, 2023, https://blogs.microsoft.com/on-the-issues/2023/07/21/commitment-safe-secure-ai/
• https://www.ie.edu/insights/articles/from-cloud-to-cup-how-much-water-does-your-chatgpt-drink/