RAG
Answers grounded in your documents, with citations. Best when the knowledge exists but is hard to find.
AI knowledge base
Solutions
AI use cases that reach production, not just the demo
Fourteen business problems we solve repeatedly, each mapped to the AI pattern that fits, the guardrails it needs and a rollout you can budget for.
- 4 patterns, 14 use cases
- Human-in-the-loop by default
- 8–14 weeks to production
Four patterns
Most AI use cases are one of four mechanisms
Naming the pattern early tells you what data you need, what can go wrong and how to test it. We label every solution below with the pattern it uses.
Agent
A model that plans, calls tools and takes actions in your systems, with approvals where it matters.
AI agents and automationExtraction
Structured fields pulled from PDFs, scans, emails and recordings, validated before they touch a system.
Document processingPrediction
Scores, forecasts and anomaly flags from your historical data, with thresholds you can tune.
Predictive analytics
Use cases
Fourteen problems we solve with AI
Each page explains the problem, the workflow with a diagram, the guardrails, a typical rollout and an illustrative example. No invented clients.
RAG + agent
Customer support AI
Tickets pile up while the answers sit unread in help centres, past tickets and internal wikis.
Deflect, assist, escalate RAG + agentSales copilot
Reps spend hours on call prep and CRM notes instead of selling, and pipeline data goes stale.
Prep, notes, next steps ExtractionDocument processing
Forms, scans and PDFs are keyed by hand, slowly and with errors that surface downstream.
Classify and extract at scale Extraction + RAGContract analysis
Clause review is slow and inconsistent, and renewal dates or risky terms get missed.
Clauses, risks, obligations Extraction + RAGRecruiting AI
Hundreds of CVs per role make fair, fast screening impossible by hand.
Screen, summarise, schedule ExtractionInvoice processing
Accounts payable keys invoices manually and chases mismatches against purchase orders.
Capture, match, post PredictionFraud and anomaly detection
Unusual transactions and behaviour are found too late, by rules that generate noise.
Score, flag, review Prediction + embeddingsRecommendation engine
Customers cannot find what they would buy, and merchandising rules do not scale.
Relevant, explainable picks PredictionPredictive analytics
Forecasts for demand, churn and capacity live in spreadsheets and gut feel.
Forecast with confidence ExtractionMeeting intelligence
Decisions and actions agreed in meetings are lost before anyone writes them down.
Transcribe, summarise, act AgentAI QA and testing
Regression testing cannot keep pace with releases, so bugs reach customers.
Generate, run, triage AgentLegacy modernisation
Old code that nobody dares to touch blocks every new feature and integration.
Understand, migrate, verify RAGAI onboarding and training
New hires take months to become productive because the knowledge sits with a few people.
Guided, measurable ramp-up RAGKnowledge management
Company knowledge is trapped in silos, duplicated, outdated and impossible to search.
One answer, with sources By sectorBrowse by industry
The same patterns applied to healthcare, fintech, legal, HR tech, logistics, events and more.
See industries
Choosing
Which pattern fits which problem
A quick reference we use in scoping calls. Most solutions combine two rows.
| Pattern | Use it when | Needs | Main risk | Example solutions |
|---|---|---|---|---|
| RAG | People ask questions the documents already answer | Clean, permissioned sources | Stale or wrong sources | Support AI, knowledge management, onboarding |
| Agent | Work needs several steps across systems | APIs, approval rules, audit log | Unintended actions | Sales copilot, QA testing, modernisation |
| Extraction | Fields must come out of unstructured files | Labelled samples, validation rules | Silent errors at scale | Documents, invoices, contracts, CVs |
| Prediction | You need a score or forecast | Historical outcomes, feature pipeline | Drift and bias | Fraud, recommendations, forecasting |
Still unsure? The build vs buy guide and our AI strategy service exist for exactly this decision.
Free tools
Interactive
Interactive
Two tools to use before you call us
Both run in your browser, store nothing on our servers and show their working.
AI readiness assessment
Sixteen questions across strategy, data, technology, skills and governance. Get a score per dimension, a maturity label and tailored next steps in about ten minutes.
Start the assessmentLLM cost calculator
Enter requests per day, token sizes, cache hit rate and a model tier. See monthly volume, cost per request and what moves the number most.
Estimate monthly costAlso useful: the RAG cost estimator, the vector database chooser and the AI security checklist.
Included in every solution
What does not change from use case to use case
The pattern changes; the engineering discipline does not. Every solution we ship includes the following, because the ones that fail in production usually skipped one of them.
- Accuracy thresholds agreed up front. We define what "good enough" means per field or per answer before writing a prompt, then measure it with evals on your real data.
- Human-in-the-loop where the cost of error is high. Low-confidence outputs route to a person; confident ones flow through. The threshold is yours to tune.
- Audit trail. Every input, model output, reviewer decision and downstream write is logged, so you can answer "why did it do that" months later.
- PII handling by design. Redaction, retention limits and the right deployment option, following our AI security and governance practice.
- Monitoring and cost control. Drift, latency, token spend and failure rates on a dashboard, covered by LLMOps and monitoring.
- Handover. Your team can run it. Documentation, runbooks and a training session are part of the fixed scope, as described in how we work.
FAQ
Frequently asked questions
How do you choose between RAG, agents, extraction and prediction?
By the shape of the problem. If people need grounded answers from documents, it is RAG. If software must take multi-step actions in your systems, it is an agent. If structured fields must come out of unstructured files, it is extraction. If you need a number or a flag about the future, it is prediction. Most real solutions combine two patterns, and we say which on each page.
Do these solutions need our data to leave our environment?
Not necessarily. We deploy to your cloud account, on-premises or to a hosted model API depending on the sensitivity of the data. Our AI data privacy page covers the options, and this guide compares them.
How long does a first solution take to reach production?
Typically eight to fourteen weeks from kickoff: two for discovery and data access, four to eight for a measured pilot with real users, and two to four for hardening, monitoring and handover. Extraction use cases tend to be faster; agents that write to systems of record take longer because of the approval design.
Can we start with the free tools before talking to you?
Yes. The AI readiness assessment scores your organisation across five dimensions in about ten minutes, and the LLM cost calculator gives a transparent monthly estimate. Both run entirely in your browser.
What if our use case is not on this list?
The fourteen listed are the ones we see most often, not the limit. If the problem fits one of the four patterns, we can usually scope it in a single call. Start with AI strategy if you are still deciding where AI pays off.