Service
AI that recommends. People that decide.
Kerah applies AI where it can reduce repetitive work or improve access to information, with the boundaries, evaluation and oversight defined before it becomes part of an operation.
Who this is for
- Operations leaders with a measurable volume of repetitive work
- Organizations that must be able to explain a decision after the fact
- Teams whose documents arrive in more than one language
- Organizations that have run an AI pilot and stalled before production
The problems this addresses
- Staff spending hours extracting the same fields from the same document types
- Knowledge that exists in documents nobody can search effectively
- A triage queue where the delay costs more than the decision
- An agent demonstration that acted on a live system before anyone had scoped its permissions
- A pilot that impressed everyone and cannot be put into production responsibly
- No agreed answer to what happens when the model is wrong
What changes
- A defined use case with a measurable before and after
- Known accuracy on representative cases, including the awkward ones
- Consequential decisions reviewed by a person with the authority to make them
- A documented answer to what happens when the model is unavailable or unsure
- Cost and value reassessed after launch rather than assumed
What Kerah can deliver
Capability, not a claim of past work. Final scope is established through discovery.
- Document extraction and classification
- Knowledge retrieval with source citations
- Internal and customer assistants scoped to a task and a permitted data set
- Assistants that use tools and act on systems within permissions granted for the task
- Request triage and recommendations
- Drafting correspondence and structured content
- Workflow suggestions
- Human-approved automation
- Evaluation, monitoring and fallback design
What you receive
Artefacts, not activities. Each one is something you can hold, read or run after the engagement.
- Use-case definition with success and failure criteria
- Data boundary decision — permitted and prohibited data
- Tool and permission specification for anything the system may act on
- Representative evaluation set and measured results
- Human review workflow
- Monitoring, logging and traceability design
- Fallback behaviour specification
- Post-launch review of cost, risk and value
How the engagement runs
- Bound the use case
- Scope, permitted data and success criteria are agreed before implementation. An open-ended assistant is not a use case.
- Evaluate against real cases
- Accuracy is measured on a representative set drawn from your own work, including malformed and multilingual inputs, not on a curated demo.
- Keep the human where it counts
- The system proposes; a person with the relevant authority decides. Which decisions require review is defined explicitly and recorded.
Commercial shape. A bounded pilot with defined evaluation criteria, followed by production work only if the pilot earns it.
Raised from the start, not before launch
Architecture, security and data decisions that are cheap to make early and expensive to retrofit.
How Kerah approaches security- A defined use case with success and failure criteria agreed in advance
- Data classification, and an explicit list of data the system may and may not receive
- Which tools an assistant may call, under whose permissions, and which of its actions can be reversed
- Evaluation against representative scenarios before production use
- Human review for consequential outputs
- Logging and traceability appropriate to the impact of the decision
- Failure and fallback paths when a model is unavailable or low-confidence
- Reassessment of cost, risk and value after launch
Where this stops
- Kerah makes no claim of guaranteed accuracy. Measured performance on a defined evaluation set is what can honestly be offered.
- An agent acts only within the permissions held by the person on whose behalf it runs, and any consequential action requires that person's approval. Kerah does not build an agent that grants itself access, spends money or contacts a third party unattended.
- Autonomous medical, employment, credit, eligibility and similarly high-impact decisions are out of scope without specialist governance and explicit written approval.
- Confidential, personal or production data is not submitted to a public AI service unless the contract, data processing agreement, configuration and security review all permit it.
- Where a deterministic rule performs the task at lower cost and risk, Kerah will recommend the rule.
Other services
Discovery and Architecture
A documented blueprint with a target architecture, a permission and data-role map, a phased roadmap and a scope you can approve, price or decline on evidence.
Enterprise Systems
One operational system where work is assigned, decisions are recorded, permissions are explicit and reporting comes from the same data people work in.
Digital Product Engineering
A designed, accessible product with a secure back end, built around the tasks people actually perform and measured after release.
Cloud, Data and Integrations
Defined interfaces, a documented data model and monitored flows, so information moves once and reporting can be trusted.
Managed Evolution
A defined support and release model with security maintenance, an owned backlog and a roadmap reviewed on a set cadence.
Security and Privacy Engineering
A system whose access model, data inventory, encryption, logging and deletion behaviour were designed together — and the documentation an assessor asks for, produced as delivery output rather than reconstructed afterwards.
Quality, Testing and Accessibility
A system tested in both languages by people who read both, measured against WCAG 2.2 AA, and exercised under realistic load and failure conditions before launch rather than after it.
Discuss an AI use case.
Tell us what is not working today, who it affects and what a better outcome would change.