AI Readiness Audit
Outcomes
You stop guessing which use case to fund first
Most organisations have between eight and twenty plausible AI ideas circulating. The audit scores them against a consistent framework — data availability, technical feasibility, process fit, measurable value, change cost — and ranks them. Disagreement about priorities usually dissolves once everyone is looking at the same scoring.
You find out what is actually blocking you, early
The blockers are rarely the model. They are data sitting in three systems that disagree with each other, a process nobody has documented, a compliance question nobody has asked, or a workflow whose exceptions outnumber its rules. Finding these in week two costs a fraction of finding them in month five.
You get a defensible business case
Each recommended use case comes with a build estimate, a run cost, an expected effect on a metric you already track, and the assumptions behind that estimate written down. That is what a CFO needs, and it is what most AI proposals do not have.
You avoid spending on the wrong thing
Industry surveys consistently find the majority of AI pilots never reach production. The dominant causes are not technical: unclear success criteria, poor data foundations, and no owner for the deployed system. An audit is inexpensive insurance against that pattern.
What we deliver
Six artifacts, all yours to keep, all written to be read by executives and engineers alike.
Use case register. Every candidate identified during discovery, scored and ranked, with the ones we recommend against included and the reasoning stated. The rejected list is often the most useful page in the document.
Data readiness assessment. What data you hold, where it lives, its quality and completeness, who owns it, and specifically what would need to change for each recommended use case to be feasible. Includes lineage gaps and consent or retention issues where they exist.
Systems and integration map. Your current architecture, the integration points any AI system would need, and the constraints they impose — API limits, latency, authentication, whether the system of record can even accept writes.
Reference architecture. For the top-ranked use cases: a concrete proposed design, named models and services, build-versus-buy recommendation, and a run-cost estimate at your projected volume. Not a generic diagram.
Governance and risk review. Applicable regulation, data residency requirements, model risk, human-in-the-loop requirements, and an audit-trail approach. Covers GDPR, India's DPDP Act, and sector rules such as HIPAA or RBI guidance where relevant.
Roadmap and first-engagement scope. A sequenced 6–12 month plan, and a fully scoped statement of work for the first build — enough to start immediately without a second discovery.
How it works
Week 1 — Discovery. Structured interviews with process owners, technical leads and at least one person who does the work daily. That last group is non-negotiable; the gap between how a process is documented and how it runs is where AI projects die. We also inventory systems and request data samples under NDA.
Week 2 — Technical assessment. Hands-on examination of data quality and volume, integration feasibility, and existing infrastructure. Where a use case hinges on an uncertain assumption, we test it — a retrieval spike against a document sample, a classification baseline on historical records — rather than assuming it holds.
Week 3 — Analysis and modelling. Scoring, cost modelling, architecture design, risk review. A mid-week checkpoint with your team to challenge the emerging ranking before it hardens.
Week 4 — Delivery. Written report, a working session with your leadership, and the scoped first engagement. We walk through the recommendations we rejected as carefully as the ones we made.
Four weeks is typical. Two to three is achievable for a single department with clean systems. Enterprises with heavy regulatory scope or many disconnected systems run longer, and we say so before starting rather than discovering it midway.
Technology we assess against
Model providers: OpenAI, Anthropic, Google, plus open-weight options including Llama, Mistral and Qwen where data residency, cost at volume, or fine-tuning control favour self-hosting.
Infrastructure: AWS Bedrock and SageMaker, Google Vertex AI, Azure AI Foundry, and self-hosted inference where economics justify it.
Data and retrieval: Postgres with pgvector, Pinecone, Weaviate, Qdrant; Snowflake, BigQuery, Databricks; dbt and Airflow for pipelines.
Orchestration and evaluation: LangGraph, LlamaIndex, and the emerging agent frameworks — assessed sceptically, since this layer moves fast and much of it is not production-ready.
We are not a reseller for any of these and hold no incentive to recommend one over another. Where the honest answer is that a spreadsheet, a rules engine, or better process design beats a model, we say that.
Where this applies
Most valuable for organisations with meaningful operational data and repetitive knowledge work.
Least valuable for companies with a single obvious, well-understood use case and clean data. If you already know exactly what you want built and why, skip the audit and start building. We will tell you that on the intro call rather than sell you an assessment you do not need.
What it costs
Fixed price, fixed scope, agreed before we start. If the audit concludes you should not proceed with AI right now, that is a legitimate outcome and the fee is unchanged — a recommendation against spending several hundred thousand on the wrong build is worth more than the audit.
- OpenAI
- Anthropic
- Mistral
- AWS Bedrock
- SageMaker
- Google Vertex AI
- Azure AI Foundry
- Postgres with pgvector
- Pinecone
- Weaviate
- Qdrant
- Snowflake
- BigQuery
- Databricks
- LangGraph
- LlamaIndex
Frequently asked questions
Three to four weeks for most organisations. Two to three weeks for a single department with clean, accessible systems. Six weeks or more for enterprises with heavy regulatory scope or many disconnected systems, which we scope and state up front.
Access to the people who own the relevant processes, roughly six to ten hours of their time across the engagement, read access to representative data samples under NDA, and documentation of your current architecture if it exists. If it does not exist, we produce it as part of the audit.
No. Representative samples under NDA are sufficient, and can be anonymised or synthetic where the data is sensitive. We do not require production write access at any point during an audit.
That is a valid and reasonably common result. The report then focuses on what would need to change — usually data foundations, process documentation, or ownership — and sequences that work. It costs the same. Knowing this before committing a build budget is the point.
A consultation gives you an opinion. An audit gives you evidence: examined data, tested assumptions, costed architecture, and a written business case you can take to a board. Free assessments are typically sales instruments designed to arrive at a predetermined recommendation.
Yes, and you are under no obligation to use us. The report is deliberately written so another firm could execute from it. An audit that only makes sense if you hire the auditor is not an audit.
Yes. A common engagement is assessing why an existing deployment underperforms — usually evaluation gaps, retrieval quality, or a process that was never redesigned around the system rather than the model itself.
More AI services
AI Strategy Consulting
Turn scattered AI ambition into a sequenced, costed plan. We decide what to build, what to buy, what to ignore, and in what order.
Agentic AI Automation
We build AI agents that complete multi-step work inside your systems — with defined scope, human checkpoints, and evaluation. Deployed to production, not demos.
Generative AI Development
Custom generative AI applications built for production — content, code, image and document generation with evaluation, guardrails and cost control.