What an AI readiness assessment framework is
An AI readiness assessment framework is a structured way to judge whether an organization can put AI to work. It names the areas that matter, says what to check in each, and gives you a way to score the answers, so the result is more than an opinion. Most frameworks cover the same ground: strategy, data, infrastructure, governance and people.
This guide compares four public frameworks, shows where they stop short for AI agents, and sets out the framework behind our free AI readiness assessment, with the scoring in full so you can run it yourself or change it.
Four public frameworks compared
| Framework | Structure | What it’s for |
|---|---|---|
| Microsoft AI Readiness Assessment | Seven pillars, listed at 45 minutes | A self-assessment with recommendations |
| TDWI AI Readiness Assessment | Five dimensions and about 75 questions | A detailed score, overall and per dimension |
| NIST AI RMF | Four functions: Govern, Map, Measure, Manage | Managing AI risk, not measuring readiness |
| ISO/IEC 42001 | Requirements for an AI management system | Governance you can have independently certified |
The first two measure readiness. The last two are often listed beside them, but they do a different job:
- Microsoft
- Its pillars are Business Strategy, AI Strategy & Experience, Organization & Culture, Data Foundations, Infrastructure for AI, Model Management, and AI Governance & Security. Unlike TDWI’s, it has a pillar for managing models.
- TDWI
- Its dimensions are Organizational, Data, Skills, Operational and Governance readiness. The assessment asks for your work email and company details before you start.
- NIST AI RMF
- A voluntary framework published in January 2023 and now being revised, with a core that runs from building a culture of risk management to acting on the risks you’ve found. It doesn’t tell you whether you’re ready to start. It’s most useful once you’ve chosen what to build, and for writing the governance part of your plan.
- ISO/IEC 42001
- Published in December 2023 for organizations that develop, provide or use AI systems. Certification is voluntary and comes from independent certification bodies, not from ISO. It’s worth a look when customers or regulators ask for proof of governance. For a first pilot, it’s more than you need.
Microsoft sells AI infrastructure and services, and we sell a paid assessment, so weigh any recommendation with that in mind. We’ve compared only what can be read openly, which leaves out analyst models behind a subscription.
Where they stop short for AI agents
These frameworks ask whether an organization is ready for AI in general. An AI agent raises two more questions.
What is it allowed to do? A chatbot answers. An agent signs in to your systems and acts in them, so identity, per-action permissions, approvals and logs decide whether it’s safe to run. In the pillar names above, access sits inside governance or security rather than standing alone. It deserves a closer look: shared logins and all-or-nothing admin accounts can’t be limited for an agent.
Which work should it do? A readiness score describes the organization, but a pilot is one workflow. Two companies with the same score should start in different places, depending on the work they repeat and where its data lives.
A six-area framework for AI agents
Our framework has two parts: readiness, scored across six areas, and workflow fit, scored for each job you might hand over. Here’s how the six areas line up with the two readiness frameworks. The mapping is our own reading of their pillar names, not theirs:
| Area | Microsoft | TDWI |
|---|---|---|
| Goals and ownership | Business Strategy, AI Strategy & Experience | Organizational |
| Data | Data Foundations | Data |
| Systems and integration | Infrastructure for AI | Operational |
| Security and access | AI Governance & Security | Part of Governance |
| People and process | Organization & Culture | Skills |
| Governance and risk | AI Governance & Security | Governance |
Two choices stand out. Security and access is an area of its own, for the reason above. And there’s no row for Microsoft’s Model Management: most companies starting with agents use a provider’s models, so that’s a decision for the build rather than a test of readiness. If you train your own models, add it as a seventh area.
The scoring rubric
Each area has two questions, and each question scores 0 to 3. The answers between the two shown here score 1 and 2, and you can read them all in the scorecard.
| Question | Scores 0 | Scores 3 |
|---|---|---|
| Goals and ownership | ||
| How clear is what you want AI to achieve? | No specific goal yet | Named problems, each with a measure of success |
| Who owns decisions about AI? | Nobody yet | A named owner with budget and time |
| Data | ||
| Where does the information for your key work live? | Paper, email and people’s heads | Business systems with APIs and clear owners |
| How reliable is that information? | Often wrong or missing | Trusted and kept up to date |
| Systems and integration | ||
| Can software reach the tools your team uses? | No: desktop apps or no APIs | APIs, plus an integration tool already in use |
| What automation do you run today? | None | Automations with owners and monitoring |
| Security and access | ||
| How is access to your systems managed? | Shared passwords | Single sign-on with roles, reviewed regularly |
| Could you limit what an AI agent can see and do? | No | Per action, with approvals and logs |
| People and process | ||
| Are your processes written down? | No, people just know | Documented, with examples of good and bad outcomes |
| How does your team use AI today? | Worried about it, or against it | Used daily, under agreed rules |
| Governance and risk | ||
| Do you have rules for using AI? | None | A policy, plus review of AI output that reaches customers |
| Do you know which data must never leave your control? | No | Yes, classified and enforced in our systems |
That gives each area up to 6 points and the whole assessment up to 36, shown as a percentage. All six areas count equally. Some frameworks weight their pillars, but weights imply a precision that a self-assessment doesn’t have, and any area at zero can stop a project on its own. That’s also why you should read the area scores before the total: an average can hide the one gap that matters.
- Under 40%
- Exploring. Fix the weakest area before building anything.
- 40% to 69%
- Ready to pilot. Prove value on one workflow, with a fixed scope and one measure.
- 70% and up
- Ready to scale. Run several agents, with shared monitoring and access rules.
These bands are our judgment, not a benchmark drawn from survey data. Treat your score as a guide to what to fix first, and don’t compare it with another framework’s number.
Scoring workflow fit
Each job you might hand to an agent gets a fit score from 0 to 100, made of three parts:
- Value, 45%: hours spent each month, from how often the job runs and how long it takes. It’s on a log scale that tops out at 160 hours, so one huge job doesn’t drown out the rest.
- Feasibility, 40%: whether the inputs are in systems software can reach, and whether the rules are clear.
- Safety, 15%: how easily a mistake can be undone.
The score ranks the jobs. A separate set of rules, applied in this order, says how much freedom to give the agent:
- Under 2 hours a month: not worth it yet.
- Inputs on paper, in people’s heads, or scattered in email: fix the inputs first.
- Judgment every time: assistant for the person doing it.
- Inputs in systems, clear rules, and mistakes that stay internal: automate end to end.
- Anything else: agent drafts, a person approves.
Here it is on three illustrative jobs, scored with the scorecard’s own code:
| Job | Answers | Fit | Result |
|---|---|---|---|
| Invoice matching | Daily, 5 to 30 minutes each. Inputs in systems with APIs. Clear rules, a few exceptions. Cost of a mistake: moderate, internal rework. | 60 | Automate end to end |
| Replying to refund requests | Many times a day, 5 to 30 minutes each. Inputs in systems, exported by hand. Rules with many exceptions. Cost of a mistake: high, customers see it. | 54 | Agent drafts, a person approves |
| Reviewing contracts | Weekly, 30 minutes to 2 hours each. Inputs scattered in email and documents. Judgment every time. Cost of a mistake: severe, legal, financial or safety. | 21 | Fix the inputs first |
The refund replies take the most time, yet invoice matching ranks first, because its inputs and rules are ready and its mistakes stay internal. For more candidates, see AI agent use cases, by team.
How to run the assessment
- Set the scope. Choose the teams, workflows and systems to cover. A whole-company score is too vague to act on.
- Talk to the owners. Ask the people who own the goals, data, systems and day-to-day work. One person filling in a form gets you one person’s view.
- Check the evidence. Score what’s true today, not what’s planned. If someone says a tool has an API, look at the docs. If they say roles are reviewed, ask when that last happened.
- Score each area, then find the lowest. That’s your first piece of work, whatever the total says.
- Rank your workflows. List the work your team repeats and score each job for fit.
- Plan one pilot. Take the top job, give it a fixed scope and one measure, such as hours saved a month.
- Score again after the pilot. You’ll have learned which answers were too generous.
Three ways to do it: the free scorecard takes twelve answers and scores them in your browser, the AI readiness checklist gives you 30 checks to tick with your team, and our AI readiness assessment service does the interviews and systems review with you.
Which framework should you use?
- You want a detailed self-assessment
- Microsoft’s or TDWI’s.
- You’re planning AI agents
- The six areas above, because they test access and rank your workflows.
- You need to manage AI risk
- NIST AI RMF, alongside any of the above.
- You need to prove governance to others
- ISO/IEC 42001.
Common mistakes
- Scoring plans instead of facts. “We’re rolling out single sign-on” isn’t single sign-on.
- Asking only leadership. The people who do the work know where the data really lives.
- Reading only the total. A good average can hide shared passwords.
- Treating readiness as a gate. You don’t need a perfect score. A pilot teaches you more than another round of questions.
- Stopping at the score. If it doesn’t end in a named first fix or a named pilot, it was only a survey.