Ethics | Bespoked AI

What we build

What we build and don't build

Bespoked AI builds knowledge architectures from documents created by people and organisations: research papers, strategy documents, programme frameworks, institutional knowledge, practice wisdom. It makes existing knowledge findable, usable, and conversational through AI.

It does not process personal data or sensitive data, build case management systems, or make decisions about people. The risks that are most relevant here are specific to knowledge retrieval and synthesis, not those associated with classification or decision-making systems.

The principles

Eight design constraints built into what we build. The implementation details will change as we learn - these are what we're trying to hold steady.

P · 01
The knowledge architecture is an asset, not a subscription

It belongs to the client. Not to us, not to whatever AI provider sits underneath. If the relationship ends, the client walks away with everything of value - not just raw files, but the intelligence layer that enables it. We do not claim intellectual property rights over the intelligence layer built from your documents.
P · 02
Provider independence

Knowledge architectures designed to be provider-agnostic. The intelligence layer works independently of which model sits underneath. You're never locked in.
P · 03
Transparent sourcing

Every output shows where it came from. Users can see which documents the AI drew from and make their own judgments about reliability. Outputs are framed as starting points, not conclusions.
P · 04
Humans are kept at the centre of the work

These systems are designed to augment a thinking human, not replace the thinking. The workflow keeps the domain expert central because the person working with the knowledge is the one who can assess whether a connection is real, whether a synthesis holds up, whether something important has been missed.
P · 05
Income-banded pricing

Access to useful tools should not be gated by ability to pay commercial rates. Pricing is banded by organisational income, reflecting the financial realities of the people and organisations we work with.
P · 06
Per-organisation, not per-seat

The knowledge base becomes more valuable when more people use it. Pricing should not create incentives to restrict access.
P · 07
Anonymisation by design

Content is curated and redacted before it reaches any AI model. Human-led review is a standard part of the process, not an afterthought.
P · 08
No training on client data

Client documents are never used to improve AI models or shared across tenants. Complete isolation of each organisation's knowledge base.

Unresolved

What we haven't figured out yet

AI systems consume energy and water at scale. The aggregate picture is serious. We make real choices about this - smaller models where adequate, efficient retrieval design, UK-based hosting on renewable energy - but the fundamental tension between building AI-dependent tools and environmental responsibility is not resolved.

Well-designed knowledge architecture can also change where the intelligence sits. When real human thinking has gone into how knowledge is organised and related, you don't always need the most powerful model to work with it. The quality of the structure compensates for the capability of the model. Some of our deployments run effectively on smaller, cheaper, less resource-intensive models precisely because the way we structure the knowledge encodes the intelligence required for a model to understand it efficiently.

Unresolved

Accountability and oversight

When the system gets something subtly wrong, what is the accountability structure? Showing sources is necessary but probably not sufficient. The broader question of shared accountability between platform, model, and user is genuinely unresolved across the industry.

Knowledge retrieval systems have their own bias risks that standard auditing frameworks don't cover: consistently surfacing certain documents over others, making invisible whatever is not in the knowledge base, smoothing out contradictions that should be preserved. We are developing approaches to monitor these patterns, but the work is ongoing.

Whether a Community Interest Company or social enterprise form would better align with these principles than a standard limited company is an open question. This connects to what happens if the business scales or gets acquired, and whether structural safeguards matter more than stated principles.

The implementation details will change as we learn. The principles are what we're trying to hold steady.