Applied AI research that asks what current systems can and cannot do.

KindLogic AI Labs is a separate applied‑R&D division. We review new research, test current tools and models, and build practical systems around real problems.

Talk to KindLogic

Currently exploring — retrieval on ambiguous source material Currently testing — workflow agents with human checkpoints Currently building — a lead-gen engine for job seekers

A division that keeps the institution close to the work.

We care about real impact — what works, what does not, and how it changes the organizations, people, and populations AI reaches. The Lab is where we test it.

Labs is a separate KindLogic division. We review new research, test what current systems can and cannot do, and build practical systems around real problems.

What we learn guides how the rest of KindLogic frames leadership conversations and which AI claims hold up under engineering scrutiny.

View KindLogic Sessions

Current Explorations

What Labs is working on now.

Current work, not a finished portfolio. Status changes; this index reflects what is active.

Workflow · Lead Qualification

AI workflow and lead-qualification systems Workflow systems that qualify inbound interest, enrich context, and route what matters.

What we tried — Workflows that combine model calls, scoring rules, and existing CRM and email systems, orchestrated with tools designed for both engineers and operators.

What broke — Most failures were not model failures. They were data hygiene, routing edge cases, and the silent ways humans had been compensating for system gaps.¹

What it raises — Once a workflow runs without supervision, the question is no longer "does it work" — it is "who owns the failure when it doesn't." That is governance, not engineering.

AI workflow and lead qualification system screenshot

RAG · Knowledge Assistant

Knowledge-based AI assistant Retrieval-augmented systems that read structured and unstructured source material to support task work.

What we tried — Pipelines combining document chunking, embeddings, retrieval, and small generation calls — both for internal task work and as a research test bed.

What broke — Retrieval quality is brittle in the places it has to be best — ambiguous queries, stale source material, and content that should not be returned to the wrong audience.

What it raises — Most "AI assistant" deployments are really information-access deployments. That means existing data permissions, review workflows, and accountability structures matter more than model selection.

Retrieval-augmented knowledge assistant screenshot

Applied Product · Job Seekers

Lead-generation engine adapted for job seekers A public-facing tool using the lead-generation engine to identify relevant contacts for job seekers.

What we tried — Adapting infrastructure designed for internal sales use into a consumer-facing surface, with new consent, privacy, and reliability constraints.

What broke — Internal-tool assumptions about data freshness, opt-in, and tolerance for false positives did not hold once real end-users were in the loop.

What it raises — What is acceptable for an internal sales workflow can be inappropriate the moment a system is public. Most teams discover that constraint after launch. We try to discover it earlier.

1 Source: internal Lab post-mortem on the lead-qualification workflow build. Specific failure modes catalogued during run-up to v0.4. Not externally published.

What this work raises.

KindLogic's operating idea is that AI deployed without care creates real impact at scale. The Lab is where we try to keep that impact concrete instead of abstract.

In the lead-qualification workflow, most failures were not model failures. They were data hygiene and the silent ways humans had been covering for system gaps.

When an unsupervised workflow gets it wrong, who in your org is named as the owner of the failure?

Retrieval quality breaks down on the exact queries the system is supposed to handle best — ambiguous wording, stale documents, content that should not reach the wrong audience.

Are your existing data permissions and review workflows ready for AI-augmented information access — or about to find out at scale?

Adapting the lead-gen engine for a public-facing job-seeker tool exposed assumptions about consent, freshness, and false-positive tolerance that were not validated against real users.

Where in your roadmap does an internal AI system become a public one — and what review happens before that line gets crossed?

Two of three Lab projects raised governance questions before they raised engineering ones.

Is governance review running in front of your AI deployment, or behind it?

Have a leadership team working through these decisions.

Talk to KindLogic about how the questions in this Lab show up inside real institutions.

Talk to KindLogic