Two AI systems that scale a firm's expertise without replacing its judgment
A political and security risk consultancy holds nearly two decades of West Africa expertise, most of it locked in documents and in the heads of its analysts. Two AI systems turn that into a capability the firm can actually scale: a source-grounded assistant over its own knowledge, and a tender intelligence platform that drafts but never decides.
The problem
Expertise that lives in documents and in people does not scale. Analysts spend their time locating what the firm already knows instead of using it, and every answer depends on who happens to be available. The same constraint applies to business development: relevant public and private tenders appear across Europe and Africa faster than a small team can read them, and each one demands a tailored response.
Approach
Two systems, one principle: the AI handles retrieval, synthesis and first drafts, and the expert keeps every judgment that carries weight. The assistant lets analysts interrogate the firm's own corpus and produce insights that cite their sources. The tender platform watches the market, reads requirements and assembles a first application package, which an expert then reviews, corrects and personalises before anything leaves the firm.
- Domain-specific assistant grounded in the firm's internal knowledge
- Retrieval, document synthesis and source-attributed intelligence insights
- Tender discovery and requirement analysis across public and private calls
- First-draft application packages, always expert-reviewed before delivery
Architecture
The assistant follows a retrieval-augmented design: the firm's documents are ingested, segmented and indexed for semantic search, the relevant passages are assembled into context, and generation is constrained to that context so every claim can be traced back to a source. The tender platform adds a discovery stage ahead of the same pattern, extracting the requirements of each call before any drafting begins.
- Ingestion, segmentation and semantic indexing of the internal corpus
- Retrieval and context construction ahead of every generation
- Source attribution, so an analyst can verify rather than trust
- Requirement extraction as a distinct stage before drafting
Deployment and control
The expert review gate is the deployment constraint, not an afterthought. The system produces drafts and insights; it does not decide and it does not submit. For a firm whose credibility rests on the judgment of its analysts, an AI that acted on its own would be a liability rather than a gain, so the workflow is designed to put a human at every point where something leaves the firm.
Business value
Knowledge that was effectively inaccessible becomes queryable by any analyst, which shortens the path from question to a sourced answer. On the business development side, the firm can consider a far wider field of tenders than it could read manually, while the quality of each submission still rests where it belongs: with the expert who signs it.
Lessons learned
Generative AI is at its strongest where unstructured knowledge is the bottleneck, not where a model is the novelty
Source attribution is what makes an assistant usable in analysis: an answer you cannot verify is an answer you cannot cite
Keeping humans on the decisions that carry weight is a design requirement, not a limitation to engineer away