What is actually inside an AI readiness assessment

"Assessment" is a word that has been used to sell everything from a two-hour workshop to a six-figure transformation programme. So here is the scope of ours, in enough detail that you could brief a competitor with it. If the document you receive does not contain these seven things, ask why.

1. A workflow inventory, scored

We sit with the people who do the work and write down the processes as they actually run, not as the process document claims. Each one gets scored on four axes: volume, variability, cost of an error, and whether the inputs are already digital and retrievable. Most organisations we look at have somewhere between ten and twenty candidate workflows and between two and five that are genuinely ready. Naming the ones that are not ready is the more valuable half.

2. A data and permissions audit

Where the documents live, who is allowed to see them, how they change, and whether that access model can be reproduced in an index. This section is where most projects find their real blocker, and it is the reason the assessment happens before the build rather than during it.

3. A cost model at your projected volume

Not a per-token price list. A model with the line items that actually decide viability:

• Inference cost at today's volume, the twelve-month plan, and 5× the plan
• Embedding and re-embedding cost as a function of how often the corpus changes
• Retrieval and vector infrastructure, including what has to stay warm
• Evaluation runs — the cost of knowing whether a change helped
• Human review for the share of outputs that will not ship unchecked
• The engineering cost of the build itself, and of running it for a year

The output is a number with a sensitivity range, which is what lets someone approve or decline it.

4. A governance and obligations map

Which regimes apply to this specific use case — EU AI Act risk classification, GDPR, India's DPDP Act, sector rules — and for each obligation, the control that satisfies it and the person who owns it. For organisations heading toward ISO 42001, this section is the gap analysis.

5. A target architecture, at the level that matters

Components, trust boundaries, data flows, and the two or three decisions that will be expensive to reverse. Deliberately not a vendor selection: the point is to make the shape clear enough that several vendors could implement it, which is also what keeps you free to change your mind.

6. Risks, named, with owners

Each risk gets a likelihood, an impact, a mitigation, and a name against it. A risk register without owners is a list of things nobody will do.

7. A sequenced 90-day roadmap

Not a wish list. An ordered sequence where each step is small enough to finish, has an owner, and produces something the next step needs. If the honest recommendation is that step one is to fix a data problem before writing any AI code, that is what it says.

What it costs and how long it takes

Two to three weeks. ₹1.5–5 lakh for Indian clients, $5,000–25,000 internationally, quoted as a fixed number before anything starts. Roughly 20–35 pages, plus the working artefacts — the cost model as a spreadsheet you can change the assumptions in, and the architecture diagrams as editable files.

The test of whether an assessment was worth the money is whether it is still useful to you after you decide not to hire us.

The outcome we will not hide from

A meaningful share of these end with a recommendation not to build. Sometimes an existing product does the job for a fraction of the cost. Sometimes the process should be simplified before it is automated, because automating a bad process just makes it fast. Sometimes the data will not be ready for two quarters and the honest sequence starts there. Those conclusions go in the document in the same plain language as everything else.