By Bonnie Chiu, Managing Director, The Social Investment Consultancy

This is a two part blog, find the first part here.

What next?

The choice between AI as a scaffold or substitute is ultimately a choice about what kinds of knowledge, we believe, get us closer to solving the world’s greates social and environmental challenges.

We propose a 2×2 quadrant to help organisations think about the type of knowledge they actually need. This framework maps different forms of knowledge production in the age of AI across two dimensions: speed of cognition (fast vs slow) and mode of knowing (explicit/analytical vs tacit/intuitive).

The horizontal axis distinguishes between explicit, analytical knowing—forms of knowledge that can be articulated, written down, and formalised (such as models, metrics, reports, and causal explanations)—and tacit, intuitive knowing, which is experiential, context-dependent, and often difficult to fully articulate, such as pattern recognition, judgement, and systems intuition.

The vertical axis distinguishes between fast thinking, which is abundant, scalable, and increasingly accelerated by AI, and slow thinking, which is scarce, effortful, and dependent on sustained attention, reflection, and the integration of complex information over time.

	Explicit / Analytical Knowing	Tacit / Intuitive Knowing
Fast (abundant)	Automated Analysis (AI-generated synthesis, reporting, dashboards)	Patterned Intuition at Scale (heuristic outputs, “sounds right” answers)
Slow (scarce)	Structured Sensemaking (evaluation, causal inference, theory-building)	Judgement under uncertainty (strategic insight, systems understanding, wisdom)

To explain each of these quadrants:

Fast + Explicit (Automated Analysis)

This quadrant captures the rapid production of analytical outputs such as summaries, dashboards, reports, literature syntheses, and basic causal explanations. It is the domain where AI is most powerful: high-volume, low-cost recombination of existing explicit knowledge. While these outputs are useful for communication and coordination, they are increasingly commoditised and interchangeable. Their value lies in speed and accessibility rather than depth of understanding, and they primarily compress information rather than interrogate it.

Fast + Tacit (Patterned Intuition at Scale)

This quadrant reflects rapid, heuristic-like interpretations that feel intuitive or insight-rich but are generated without sustained reflection. In the AI context, it includes fluent narratives about “what is going on” produced from incomplete or noisy inputs, often with a high degree of coherence but limited causal grounding. It resembles intuition, but is structurally closer to pattern completion at speed. The risk in this quadrant is that it produces a strong sense of understanding without the underlying discipline of interrogation, creating plausible stories that may not withstand scrutiny.

Slow + Explicit (Structured Sensemaking)

This quadrant represents deliberate analytical work: evaluation, causal inference, theory-building, and systems modelling. It involves explicitly testing assumptions, integrating multiple sources of evidence, and constructing explanations that can be examined and refined over time. This is the traditional domain of rigorous research and evaluation. It is slow because it requires iteration, contradiction, and sustained attention to complexity, but it produces forms of knowledge that are transparent, contestable, and capable of improving collective understanding.

Slow + Tacit (Judgement under Uncertainty)

This quadrant captures the most scarce and consequential form of knowing: judgement formed under conditions of ambiguity, incomplete evidence, and long causal chains. It includes strategic insight, systems intuition grounded in experience, and the ability to navigate trade-offs where there is no clear analytical resolution. Unlike explicit knowledge, it cannot be fully codified; and unlike fast intuition, it cannot be reliably generated without sustained exposure to feedback and complexity. It is the domain of wisdom—where decisions are made not on the basis of complete information, but on the capacity to interpret what matters when certainty is unavailable.

Recommendations to protect slow thinking

This article argues that we need more slow thinking. The central question is therefore how organisations can deliberately protect and design for it. This is fundamentally an institutional design challenge—how organisations structure time, define value, and reward different kinds of cognitive work. Without intervention, there is a risk that we drift toward a substitution future by default, as AI makes fast, explicit knowledge abundant and organisations respond to pressure by accelerating output rather than deepening understanding.

Social impact systems, which already privilege explicit knowledge, risk becoming over-optimised for what can be written down and under-optimised for what must be understood. If organisations continue to reward only what can be produced quickly and documented easily, AI will amplify an already dominant fast-thinking system. If, however, institutions deliberately design space for slow, interrogative, and judgement-forming work, AI may instead become a mechanism for expanding—not collapsing—the depth of human understanding.

We therefore propose recommendations at two levels: for knowledge organisations and for funders and commissioners.

For knowledge organisations

1. Treat thinking time as a designed asset, not residual capacity
Slow thinking must be explicitly planned, protected, and resourced. It cannot be assumed to emerge in gaps between delivery tasks. This means creating structured space for sensemaking, reflection, and interrogation of assumptions—not as “non-billable overhead,” but as core cognitive infrastructure

2. Separate fast production from slow interpretation
Organisations should distinguish between AI-accelerated production work (analysis, synthesis, drafting) and slower interpretive work (meaning-making, causal reasoning, systems thinking). These should not be collapsed into the same workflow or time pressure, otherwise fast outputs will continuously displace reflective judgement.

3. Redesign performance expectations away from throughput alone
Output volume is an increasingly weak proxy for value. Organisations should explicitly recognise and reward contributions that improve the quality of understanding over time: reframing problems, identifying faulty assumptions, and improving the robustness of theories of change.

But their ability to do so fundamentally hinge upon funders and commissioners for knowledge work.

For funders and commissioners

1. Stop treating speed and cost efficiency as primary signals of value
While efficiency matters, it is not equivalent to understanding. Commissioning systems should avoid reinforcing a logic where faster delivery is automatically interpreted as better performance, particularly in complex and long-term contexts.

2. Commission for judgement, not only deliverables
Much of the highest-value work in social impact is not the production of reports, but the development of better decision-making under uncertainty. Funders should explicitly resource processes that improve interpretive quality over time, not just outputs at fixed points.

3. Balance accountability with epistemic diversity
Social impact systems over-rely on explicit, measurable forms of knowledge because they are easier to audit. However, this systematically under-values tacit, experiential, and slow forms of knowing where much strategic judgement resides. Funding models should deliberately protect space for these less legible but essential forms of cognition.

The future of AI in social impact work will not be determined by the capabilities of the technology, but by whether we design institutions that protect the time required for judgement in a world optimised for speed. The challenge for organisations is ensuring that the abundance of answers does not crowd out the slower processes through which understanding, judgement and wisdom emerge.