Responsible AI humans decide, always.

AwardScience uses AI to remove administrative work from award and grant programmes, never to decide who wins. A person makes every decision that affects an applicant, and no customer data is ever used to train an AI model. Here is exactly how our AI works, and the limits we place on it.

Principle
01
01

Humans decide

AI output is advisory. Human review is mandatory and cannot be switched off. No configuration lets an AI result determine an applicant’s outcome without a person confirming it.

Principle
02
02

No AI model training

Your data is never used to train, fine-tune or improve any AI model, ours or a provider’s. That covers submissions, applicant data, uploaded documents, scores and judge comments. Inference runs under enterprise terms that prohibit training on inputs.

Principle
03
03

You stay in control

Every AI feature can be switched off, individually or entirely, per programme. With AI off, no submission data reaches any model.

Every AI feature, and what it sees.

All AI inference runs through Amazon Bedrock in the customer’s own tenant region, so submission content processed by an AI feature stays in that region. Each feature below states what it does, what data it receives, and where a person stays in control. None of them trains a model on your data.

Build
01
01

Programme building

Drafts a complete award or grant structure, stages and workflow from a short brief, with a quality review pass before it is shown.

  • Sees administrator text only
  • No applicant or submission data
  • Administrator reviews and publishes
  • No AI model training
Build
02
02

Form building

Generates application form fields, question order, help text and conditional logic from a description.

  • Sees form design text only
  • No applicant or submission data
  • Administrator reviews and publishes
  • No AI model training
Build
03
03

Campaign site assistant

Helps structure the public programme site, its pages and navigation, from the administrator’s instructions.

  • Sees configuration text only
  • No applicant or submission data
  • Administrator approves every change
  • No AI model training
Process
04
04

Document extraction

Reads uploaded PDFs and images and pulls structured fields into the application, so applicants and reviewers retype less.

  • Sees the uploaded document
  • Output visible and editable
  • Optional, per field
  • No AI model training
Process
05
05

Translation

Translates form content, submission text and email campaigns, so multilingual programmes run on one platform.

  • Sees the content being translated
  • Original always retained
  • Optional
  • No AI model training
Integrity
06
06

Duplicate detection

Compares submissions by meaning rather than exact wording, and surfaces likely duplicates with a similarity score before judging begins.

  • Embeddings stored in your tenant region
  • Deleted with the submission
  • Flags only; a person decides
  • No AI model training
Assess
07
07

Scoring assistance

Drafts an assessment against your published criteria, with written reasoning a judge can check, to support rather than replace the judge.

  • Sees submission and criteria
  • The judge scores; AI only assists
  • Off by default for regulated customers
  • No AI model training
Automate
08
08

Workflow automation

Helps configure workflow stages and conditions, and powers the messaging assistant that answers applicant questions about status and deadlines.

  • Configuration changes need approval
  • Cannot change an outcome
  • Optional
  • No AI model training
Insight
09
09

Ask AI and insights

Answers questions across your programme data, analyses charts, and suggests content for generative fields, for authorised administrators.

  • Respects role permissions
  • Advisory only
  • Optional
  • No AI model training

No automated decisions, and no model training. No AwardScience AI feature makes an automated decision producing a legal or similarly significant effect on an applicant, within the meaning of Article 22 GDPR, and no customer data is used to train, fine-tune or improve any AI model. Human confirmation is required before any AI output affects an outcome, and this cannot be configured away. Before an AI feature is released or a model changes, we evaluate it for extraction accuracy, translation quality, fabrication, scoring consistency and fairness, and enterprise customers receive 30 days’ notice of any change of model, version or provider. AwardScience uses Anthropic Claude and Google Gemma models, served through Amazon Bedrock, so the model providers do not receive customer data.

Scroll to Top