# Data and AI foundations that give AI the right context.

> Connect is veridive’s data and AI foundations service. The team connects approved data sources through permission-aware retrieval and pipelines, selects models by testing them on a client’s own examples, builds evaluation sets with acceptance thresholds, and deploys in the client’s cloud or on-premises. It runs inside a pilot or as a four- to eight-week foundation project.

Good answers need the right data, the right permissions and the right model. veridive connects approved sources, chooses models on evidence and builds the evaluation that keeps quality visible.

## What are data and AI foundations?

Data and AI foundations are the layers under a useful AI system: approved data sources, permission-aware retrieval, the right model for each task, and the evaluation that proves quality. veridive connects these layers to your existing systems and deploys them in your cloud or on-premises, so answers stay grounded, governed and measurable.

## When should you talk to us?

- You are not sure whether your data is good enough, or reachable, for AI.
- Answers must respect who is allowed to see which document.
- The team argues about which model to use, with no evidence either way.
- Data has to stay in your own cloud, in a specific region or on your own servers.
- Quality changes from week to week and nobody can say why.

## What do you get?

- **Permission-aware retrieval.** Search and retrieval that respect the access rights already attached to each document and record.
- **Data pipelines.** Approved sources connected, cleaned and kept current, from ERP records to shared drives.
- **Model selection on your examples.** Candidate models compared on quality, cost and speed using your own cases, not public benchmarks.
- **Evaluation sets and thresholds.** Real examples with expert-approved answers, and the acceptance thresholds a system must meet before it goes live.
- **Deployment where your data lives.** Your cloud, a managed environment or on-premises, with processing locations and access roles mapped against your KVKK and GDPR requirements.

## How does a foundation project run?

1. **Map sources and access** (Week 1). List the approved data, where it lives, who may see it and how current it needs to be.
2. **Build the evaluation set** (Weeks 1–2). Collect real examples with expert-approved answers and agree the acceptance thresholds.
3. **Connect and compare** (Weeks 2–6). Build retrieval and pipelines, then compare candidate models on the evaluation set.
4. **Deploy and monitor** (Final 1–2 weeks). Deploy in the agreed environment with quality, cost and access checks running.

## How can you start?

- **Pilot to Production** · ≈6–10 weeks. The foundations a first workflow needs are built inside the pilot: retrieval, model choice and evaluation, on approved data.
- **Embedded AI Partnership** · Monthly. As more workflows come on board, a named team extends the shared foundations, with monthly capacity and service levels agreed together.

## What changes?

- **Grounded answers.** Every answer traces back to an approved source that someone can check.
- **Access that matches your rules.** People see only what they are already allowed to see.
- **Model choice on evidence.** You can switch models when prices or quality change, because the evaluation set shows what you would gain or lose.
- **No lock-in.** The design is model-agnostic, with Anthropic, OpenAI, Google and open-weight models all on the table.

## Questions about data and models

### Is our data good enough for AI?

Usually it is good enough to start, as long as it is reachable. What matters is whether the information a workflow needs lives in approved systems we can access with the right permissions, and whether it is current. We check this early, on your real examples, and if the data is not ready we say so and scope that work first.

### Which AI model should we use?

The one that performs best on your examples, at a cost you can sustain. We work with Anthropic, OpenAI, Google and open-weight models such as Llama, Mistral and Qwen, and compare candidates on the same evaluation set for quality, cost and speed. Often a smaller model handles most cases and a larger model or a person handles the rest.

### Can AI run on our own servers?

Yes. We deploy in your cloud, in a managed environment or on-premises, depending on where your data needs to live. For regulated data that cannot leave your infrastructure, open-weight models can run on your own servers. We map processing locations and access roles up front and align with your KVKK and GDPR requirements.

## Related

- Where it works: [Knowledge & document intelligence](https://veridive.com/solutions/knowledge-document-intelligence/) · [ERP & enterprise workflows](https://veridive.com/solutions/erp-enterprise-workflows/) · [Voice & meeting intelligence](https://veridive.com/solutions/voice-meeting-intelligence/)
- Field notes: [Evaluation sets are the new requirements document](https://veridive.com/insights/evaluation-sets-are-the-new-requirements/) · [Answers with receipts: why citations matter at work](https://veridive.com/insights/answers-with-receipts/) · [The cost of a token vs. the cost of a mistake](https://veridive.com/insights/cost-of-a-token-vs-cost-of-a-mistake/)
- See also: [Guardrails](https://veridive.com/approach/#guardrails) · [Improve: Reliability & continuous improvement](https://veridive.com/services/ai-reliability/) · [Ways in](https://veridive.com/services/#ways-in) · [Start a project](https://veridive.com/contact/)

## Wondering whether your data is ready?

Tell us which workflow you have in mind. We will reply within one business day with a recommended first step. [Start a project](https://veridive.com/contact/) or write to [hello@veridive.com](mailto:hello@veridive.com).