AI procurement conversations tend to start with the model and end with the data. A demonstration works beautifully on a clean sample. The buyer then discovers that customer records sit in three systems, product descriptions disagree, permissions are granted by shared login, and nobody can say which report is authoritative. The model was never the binding constraint.
This guide is about buying the foundation underneath: the agreements, pipelines, definitions and controls that let a model use company data without creating a new governance problem. It is written for the manager who must justify the spend, not for the data engineer who will build it.
What AI-ready actually requires
"AI-ready" is not a product category. It describes five layers that must work together.
- Inventory and ownership: a list of systems, datasets and the person accountable for each. Without this, every project begins with archaeology.
- Quality: defined rules for completeness, accuracy, duplicates and timeliness, plus a way to measure them rather than assert them.
- Semantics: shared definitions for customer, product, order, revenue and other terms that different departments currently use differently.
- Access and governance: identity-based permissions, purpose limitation, retention and an audit trail showing who used which data and why.
- Delivery and monitoring: reliable pipelines, lineage from source to output, freshness indicators and alerting when a feed fails.
A vendor may sell one layer and describe it as the whole foundation. The buying question is which layers you already have, which you are adding, and who owns the seams.
Match the foundation to the use case
Different AI uses place different demands on the same data estate. Write the intended use case before evaluating platforms.
- Document retrieval and assistants: source freshness, chunking quality, citation, document-level permissions and deletion. Volume matters less than relevance and access control.
- Analytics and forecasting: history depth, consistent definitions, reconciliation to finance, lineage and reproducible transformations. Latency can often be daily rather than seconds.
- Agentic actions: current state, tool-level authorisation, idempotency, audit logging and a rollback path. A read-only reporting foundation is not sufficient for an agent that changes records.
If the use case is still "explore AI", do not buy a platform yet. Fund a bounded data-quality or access project that will help whichever use case is chosen later.
Foundation requirements to specify
- Named executive owner and a data owner for each domain in scope.
- A short list of decisions or workflows the foundation must support within two quarters.
- Source systems, interfaces and known data contracts documented.
- Definition of the minimum quality thresholds for each critical field.
- Permission model that reflects the organisation, not a shared service account.
- Lineage from source field to report or retrieval answer.
- Retention and deletion rules for personal and sensitive data.
- Export format and schema ownership stated in the contract.
- Support model with response times for failed pipelines and access incidents.
- Exit plan covering data, metadata, transformation logic and credentials.
The vendor landscape
- Cloud data platforms: strong storage, compute and catalogue capabilities; watch consumption pricing, egress and the cost of governance add-ons.
- Data quality and observability tools: good at monitoring rules and freshness; they do not fix unclear ownership or bad source design.
- Integration and master-data vendors: useful for customer or product identity; integration projects can become permanent if matching rules are not agreed.
- Boutique data engineering firms: practical delivery and close fit; check documentation and knowledge transfer so you are not dependent on one team.
- Internal build: best context and control, but requires sustained platform engineering capacity and on-call ownership.
Scoring the foundation vendor
Score candidates from one to five on each criterion and attach evidence.
- Use-case fit: can they demonstrate the exact access, freshness and permission pattern you need?
- Lineage and observability: can they show a field moving from source to answer, including failures?
- Identity and permissions: does access follow the employee's role and the purpose, and is it revoked when someone leaves?
- Connector coverage: are your actual systems supported natively, or will every source become custom work?
- Portability: can data and logic leave in a usable format, or does value depend on proprietary artefacts?
- Cost predictability: can you model a low, expected and high usage month without a specialist?
- Security and privacy: encryption, logging, segregation, vendor access and incident notification are documented.
- Support and knowledge transfer: your team can operate the foundation after the implementation partner leaves.
Commercial questions
Ask what unit drives the invoice: stored volume, compute time, number of users, connectors, events, queries or egress. Request a twelve-month model at low, expected and high adoption, not a single illustrative figure. Separate implementation from subscription, and ask how change requests are priced. Clarify whether monitoring, catalogue, lineage and disaster recovery are included or sold as separate modules. Ask for the renewal mechanism and any minimum commitment. If the answer requires a specialist to interpret, the pricing model is not transparent enough to buy.
Proof of value before rollout
Choose one domain and one decision. Build a narrow pipeline with real data and test completeness, freshness, schema stability, lineage, access revocation and recovery from a failed load. Measure the time it takes an authorised user to get a correct answer, and the time it takes to revoke access. Document the result. A four-week proof of value that exposes a permissions problem is more valuable than a polished demonstration that avoids one.
Ownership and exit in the contract
The contract should state who owns source data, derived data, metadata and transformation logic. Require export in an open format, a documented schema and a transition period. Agree what is deleted and when. If the platform hosts models or embeddings derived from your content, clarify whether those artefacts are yours and whether they can be removed. Exit terms are part of the architecture, not an afterthought for legal review.
When you should not buy
Do not buy a foundation when there is no named use case, no data owner, or an ERP or billing migration already in progress. Do not buy it to satisfy an "AI strategy" deliverable. And do not buy a platform whose main justification is that it is the same one a much larger company uses. Your constraints, skills and budget are the relevant comparison.
Where this sits in our practice
We build data and integration layers under Custom Software & Integrations. Engagement boundaries and commercial terms are set on the pricing page. For a longer view of sequencing this work, read the year-by-year guide to AI-ready data foundations.
Sources
- NIST, AI Risk Management Framework, accessed 17 November 2026.
- NIST, Privacy Framework, accessed 17 November 2026.
- ISO/IEC 42001, Artificial intelligence management system, accessed 17 November 2026.
- NIST SP 800-63B, Digital Identity Guidelines, accessed 17 November 2026.
Responses (0)
No comments yet. Add the first one below.
Leave a response
Your email address will not be published. Required fields are marked *.