AUGUST 28, 2026  |  9 minute read

Impact Investors Are Already Using AI. The Harder Part Is Learning How to Steward It Responsibly

How we mobilize philanthropic and investment capital is shaped by leaders working at its edges and pushing the boundaries – practitioners, thinkers, and provocateurs who are reckoning with what the field is getting right (or wrong) and where it’s headed. 

 

To spark reflection and dialogue, we are making space for members of Impact United’s community to share what is on their mind through ‘op-ed’ style guest blogs. We are interested in showcasing a range of ideas, opinions, and perspectives, and invite you to share yours at hello@impactunited.ca. 

Article by Gigi Alsaadi (CRANE Fellow, Strategic Growth at Prime Coalition) and Michelle Baldwin (Co-Executive Director, Impact United Academy)

(Guest blogs reflect the views of its author, not Impact United Academy.)

Ask an impact investor whether they use AI, and most will say yes: for due diligence, memos, portfolio monitoring, data that used to take days to find or process. Ask whether they have had any real say in how that AI was built, whose knowledge trained it, or who absorbs the environmental and human costs of running it, and the answers become more challenging.

That gap, between using AI and shaping its direction, was the subject of a July conversation hosted by Michelle Baldwin at Impact United Academy. She was joined by Keegan Leahy, Anishinaabe founder of the Turtle Island Foundation and co-founder of the International Indigenous Innovation Initiative (i4); Ryan Clancy, who leads impact management and AI strategy at Rally Assets; and Beth Richardson, founder of Management of the Good, who helps social-sector organizations put AI into practice. 

In this blog, we are sharing their perspectives on where AI is already reshaping investment practice, where it may reproduce the biases and power imbalances that impact capital exists to correct, and where investors, foundations and communities still have real influence. 

“Think about AI not just as a tool, but as a system that you have the power to shape” — Michelle Baldwin

Impact investors climb the AI maturity curve

About half of impact investors are using AI tools. As adoption grows to match the AI use in financial services firms (at about 80%), it is becoming more of an imperative to understand how AI gets used, and how far it should go.

Clancy pointed to how deeply embedded AI already is at Rally Assets, running through due diligence, portfolio monitoring, and stewardship work. Richardson sees that as one end of a broader curve: a five-stage maturity model most organizations move through, from scattered individual experimentation (roughly 12% of organizations) to what she calls the “delegate” phase, where teams connect data sources and share automated workflows across the organization (22%) — the stage she thinks most impact organizations should aim for. 

For example, BlueMark has automated part of its own fund-manager scoring: instead of a full survey, managers submit documents, AI generates a first-pass analysis, and BlueMark IQ gives investors a faster read on a manager’s practices. In another scenario, Richardson’s client team replaced hours of manual chart-building in Tableau with an AI workflow, built on Airtable and Claude, that generates on-brand infographics from survey data, with a layer of human QA (quality assurance).

Richardson doubts the friction around doing this responsibly will last. “This is a product flaw,” she said, pointing to how much manual effort responsible AI still demands, a difficulty she sees as a business opportunity waiting to be solved.

Bias lies in the training data

“The machine has a master, so we need to make sure that the master understands Indigenous values.” — Keegan Leahy

AI doesn’t invent its biases; it inherits them from the people who build it. That conviction drove Leahy to write an essay, “The Machine Has a Master,” after repeatedly encountering AI outputs stating disputed or simply inaccurate claims about Indigenous history and culture as settled fact. 

Pushing back usually worked, which was itself revealing: “People that program AI, their own internal biases tend to come out in that,” he said. The bias was not sitting on top of a neutral system but was built into what the system had learned to treat as authoritative.

That mechanism showed up in smaller, everyday forms too: one participant watched AI strip every mention of diversity from a LinkedIn post summary, which was exactly the post’s subject, and a colleague saw AI meeting notes do the same with U.S. speakers. Richardson tied this to something structural: so much AI training data is in English that entire countries and languages barely register, meaning an emerging-market investor’s prompt may return thinner results for the very geographies impact capital exists to serve. And more recently, a UN Women review found measurable gender bias in 44% of 133 AI systems assessed, a pattern researchers link to decades of unequal representation in training data. 

For investors, Richardson suggests treating bias mitigation as ongoing testing: for example, rewording prompts that might exclude data-limited investments, and checking if a “founders” search returns only men. She also suggested that AI can potentially cut bias by acting as one more voice at the investment committee table. 

Is AI then truly a bias machine to avoid, or can it act as a bias reduction tool? Both realities may be true, depending on who’s checking, how often, and for what. 

Whose knowledge, and whose rules?

For Indigenous communities, questions of consent and control aren’t abstract. CARE principles are the frameworks formalized by the Global Indigenous Data Alliance in 2019 to assert Indigenous peoples’ collective authority over data about their own communities (as opposed to treating that knowledge as a public resource). Those principles predate generative AI by years; Leahy’s point was that they already describe the consent and governance model AI development still lacks.

He described cases where national organizations, acting in good faith to preserve Indigenous knowledge, uploaded culturally sensitive material into AI systems that aren’t “ring-fenced” — meaning it doesn’t stay contained within the community that shared it. A key issue is that with more than 500 distinct Indigenous communities in Canada, each with its own protocols, there is no single “Indigenous perspective” an AI system could responsibly represent. 

Leahy’s advice was to use free, prior and informed consent standards used on resource projects: don’t assume but ask directly, meet with elders and economic development leadership, and let the community decide what’s shared and what stays out entirely. 

Clancy described what that looks like for investors: Rally classifies Indigenous or culturally sensitive data as one of its highest-risk tiers, off-limits to casual use. They also guard against the opposite failure of under-including communities because their data is treated as too sensitive to touch. That way, the tension between protection and erasure is negotiated, repeatedly, with the people involved. 

AI’s price tag isn’t evenly shared

Data centers today consume about 1.5% of global electricity, which is expected to double by 2030. Clancy shared that data center demand rose 17% in 2025 alone, more than five times the 3% growth in overall global electricity demand. Rally mitigates this via tracking efficiency approaches, such as neuromorphic computing and quantization, and practicing “model routing” — sending routine tasks to smaller, less energy-intensive models instead of defaulting to the largest one available. 

But efficiency gains don’t settle who absorbs the prevailing costs. One participant shared that AI infrastructure is sometimes built in communities with little power to object. As they lose access to clean water or face steep electricity costs for facilities that create few local jobs, AI companies often route psychologically difficult, content-screening work to low-wage workers in the Global South. 

The picture isn’t one-sided: community opposition has already blocked or delayed more than $64 billion in proposed projects. But who bears AI’s physical costs versus who captures its financial upside is a gap investors are only beginning to price in. 

The jobs picture was less clear-cut. Clancy cited a World Economic Forum forecast of 170 million jobs created and 92 million displaced by 2030, a net gain of 78 million, though there is no real consensus on those figures. What concerned him more was distribution: genuine job growth can still concentrate among people and places already positioned to capture it. That is what he thinks investors should ask portfolio companies about, not simply whether they use AI, but what their human capital strategy looks like alongside it.

What levers do investors actually have?

“Capital markets too have the ability to accelerate climate change and inequality and have done that very thing.”

— Ryan Clancy

One practical message came from Clancy’s framework for where impact investors actually hold influence:

  • Internal adoption: using AI responsibly in-house; 
  • Investment activity: due diligence and stewardship conversations with fund managers and portfolio companies; and 
  • Capital allocation: funding AI applications that close real gaps, or the infrastructure that makes AI itself more sustainable. 

Framed that way, “will AI advance our purpose or will it erode it?”, the question Rally asked at the start of building its AI strategy, is not answered only once. It gets asked again at every lever, on every deal.

Richardson believes the biggest opportunities lie beyond any single organization. She pointed to a newer B Corp requirement that member companies engage in collective action, and wondered whether impact investing networks could do something similar, advocating jointly on AI policy rather than each institution navigating risk alone. 

As an example of collective effort, a $20 million philanthropic grant recently helped education-sector practitioners identify where AI’s underlying data falls short: voice-recognition tools that don’t work for children under five, and datasets missing the demographic details needed to check if education outcomes are equitable. Rather than just documenting those gaps, the grant funded the creation of the missing data itself. That, Richardson said, is what shaping AI looks like in practice, rather than simply adopting it.

Other attendees described similar bets already underway: an early grant to an Indigenous AI platform to help it become investment-ready, with equity to follow. The same attendee also made a separate equity investment in a Canadian company working on political disinformation protection, which later drew government matching grants worth three times the original stake. The lesson was about sequencing: philanthropic and concessionary capital can absorb the early risk conventional venture money won’t touch, clearing a path toward the equity checks that follow. 

The better question: Who gets to shape AI?

Using AI is quickly becoming the baseline, not the differentiator. The more useful distinction is between an organization that has adopted AI and one that is actively shaping the systems, incentives, data and governance AI depends on. It is the difference between keeping pace and, in Baldwin’s words, understanding you have “the power to shape” the system rather than simply operate inside it. 

For impact investors, foundations and asset owners, that distinction points to specific next questions rather than a general resolution to do better. Where does your organization’s data, or your portfolio companies’ data, actually come from, and whose consent was asked before it went into a model? Where could a grant, a stewardship conversation or an equity check do more to close a real gap in whose knowledge counts as data, rather than simply making an existing workflow faster?

While none of the practitioners had a complete playbook, this conversation offered a better way of asking the question that keeps the sector’s actual purpose in view: not whether to use AI, but on whose terms, and for whose benefit.

Watch the session recording here. Thanks to our guest speakers for generously sharing their time and insights.

Resource Links