AI readiness assessment

An AI readiness assessment tells you whether your data can answer your team's questions through AI before anything is built. You get five checks, each ending in a pass or a named fix, and ten of your team's own questions tested by hand, so you know what an AI could answer today.

Last reviewed

Readiness assessment, example

3 of 5 checks passed, and 4 of 10 questions could be answered today

Example data

The five checks

  1. Metric definitions

    Fix, priority 1

    14 metrics appear in last quarter's reports. 6 are written down with a formula and an owner.

  2. Data model

    Pass

    Orders, customers, products and sessions modelled at one row each, with tests on every key.

  3. Access and permissions

    Fix, priority 2

    Warehouse roles grant whole schemas. Ad spend and margin sit in the finance schema beside payroll.

  4. Freshness

    Pass

    Load log for 20 working days: all 12 reporting tables landed by 07:00 every day.

  5. The questions people ask

    Pass

    62 questions from last quarter's ticket queue and analytics inbox, sorted into 18 kinds.

The two fixes, in order

  1. 1. Write the 8 missing definitions, starting with net revenue, new customer and repeat purchase rate. Unblocks 3 test questions.

  2. 2. Give the AI's account read-only views of ad spend and margin, outside the payroll schema. Unblocks 1 test question.

Ranked by test questions each fix unblocks alone: definitions 3, access 1.

Ten test questions, run by hand

  1. 1

    How many orders did we take last week?

    Answered. Orders table, definition written

  2. 2

    What was net revenue by channel last month?

    Not answered. Net revenue has two versions in reports, neither written down

  3. 3

    Which products had the most returns in October?

    Answered. Returns joined to products, definition written

  4. 4

    How many new customers came from paid social last month?

    Not answered. "New customer" not defined

  5. 5

    How many sessions came from email last week?

    Answered. Sessions table, definition written

  6. 6

    What was gross margin by category in October?

    Not answered. Margin sits in the finance schema, no read-only grant possible today

  7. 7

    Which customers haven't ordered in 90 days?

    Answered. Customers table, last order date tested

  8. 8

    What is our repeat purchase rate?

    Not answered. No agreed time window, so no definition

  9. 9

    Why do customers send items back?

    Not answered. Return reason isn't captured anywhere

  10. 10

    What does a new customer cost us, by channel?

    Not answered. Ad spend sits in the finance schema

Answered 4. Not answered 6: no definition 3, access 2, not collected 1.

Answered means worked out by hand from one modelled table and a written definition.
Example data: an invented online retailer's assessment. Question 9 fails for a reason outside the five checks: the data was never collected, so it goes to the data work list. Question 10 needs both fixes, because it also needs a definition of a new customer. The tally counts it once, under access, and neither fix counts it alone.

What this fixes

The board wants AI by springSomeone has promised an AI reporting pilot without checking whether the numbers it would read agree.
Finance and marketing both say revenueBoth teams have a figure called revenue, and the AI will pick one of them.
Security asked for a listThe project needs a list of every table the AI could read, and the warehouse only grants whole schemas, payroll included.
The pilot that guessedA trial chatbot got churn wrong, and the project lost its sponsor.

What you get from an AI readiness assessment

The five checksEach check shows pass or fix and one line of evidence, as in the exhibit. We give no overall score, because a score doesn't tell you what to fix first.
The fix listEach failed check becomes a named piece of work, ranked by how many test questions it unblocks. Pipeline or modelling work goes to data infrastructure, and data never collected goes to tracking.
The question logIt answers ten of your team's questions by hand, or says why it couldn't. The answered ones become the acceptance test for a later build.
The draft access mapIt lists the tables an AI would need, each read-only, and the ones it must never reach. If the build goes ahead, it becomes the MCP server's list of named tools.
The definitions gapIt lists every metric in your reports, whether it's written down and who could own it. The finance dashboard in the KPI library shows a finished entry.

How the assessment runs, before any AI implementation

  1. 1

    Collect

    Last quarter's questions from the ticket queue and analytics inbox are sorted by kind, and the ten chosen cover the commonest kinds.

  2. 2

    Check

    The checks test definitions against your reports, the model against the warehouse, access against role grants, freshness against load logs, and the fifth, a real question, against the question list.

  3. 3

    Test by hand

    Each of the ten gets an answer in SQL against the modelled tables, using only written definitions, or the reason it couldn't.

  4. 4

    Handover

    The checks, the question log, the draft access map and the fix list go into your drive. If you go ahead, the AI-led analytics build starts from the fix list.

How an AI strategy fails before the build

Questions made up for the testTidy questions written for the test pass easily, so the test uses real ones from the ticket queue, typos included, each noted with its source.
A pass on paperA definition in a slide deck gets marked as passed. We pass a check only on something you can open and read today.
A score where a list should beA maturity rating of three out of five tells no one what to do next, so every check gets a pass or a named fix, and the fixes go in order.

The assessment installs nothing and no AI touches your data, because the ten test questions are answered by hand.

Readiness by industry, and a worked example

EcommerceNet revenue has to say whether a returned order comes off on the day the parcel arrives back or the day the refund clears.
Financial servicesThe draft access map names which tables an AI may read and which model provider would see the results.

To see a question once all five checks pass, try the AI-led analytics demo. It runs three scripted questions on invented data and shows the query behind each answer.

Questions about the assessment

Do we need an AI strategy consultant before this?

Only if nobody can yet name the questions the AI should answer. The assessment starts from the questions your team already asks, and if there aren't any yet, collecting them comes before the five checks.

What does the assessment need from us?

It needs read access to the warehouse, its load logs and your current reports, plus last quarter's analytics questions and one person who can say what each metric is meant to mean.

What happens if every check fails?

You get five named fixes in order, each tied to the test questions it would unblock. In the example the first, writing down the 8 missing definitions, needs no new software and unblocks 3 of the 10 test questions on its own.

Can an AI implementation consultant skip the assessment?

They can, but then the checks happen during the build, after a wrong figure has reached someone. An AI build needs all five: a written definition, a modelled table, a read-only grant, a fresh load and a real question.

Is our data too messy for AI reporting?

The test questions answer that for your data. A question passes when one modelled table and one written definition give the answer by hand. If four of ten pass, those four can go live first, and the fix list says what the other six need.

Has someone promised the board an AI pilot?

We take ten of last quarter's questions on the call and say which checks could sink them and how long the assessment needs.

Book a call

Free 30 min call