Generative AI & LLMs
AI Readiness Assessment: A Practical Framework Before You Build
Written by Tehreem FatimaReviewed by Umaid Asim
Published 18 min read

Red marks the three gates (problem, data and risk) and the not-yet decision. Blue marks the four checks that can only add conditions and the two decisions that let the project proceed.
An AI readiness assessment checks whether your organization can build and run one specific AI use case with the data, systems, people, controls and budget it has today. It ends in one of three decisions (ready, ready with conditions, or not yet) and a short list of gaps to close before anyone writes code.
This guide is for the people who have to say yes or no to an AI project: a founder choosing a first AI feature, an operations lead with a process worth automating, or a product or IT lead whose last pilot stalled. It gives you a seven-check AI readiness framework, the evidence to ask for in each check, a template you can copy, and a scoring rule that turns the results into a decision.
The sections follow the order you would use in practice: what the assessment is, why readiness gaps sink projects, the seven checks, how to run and score them, and the extra checks for generative AI, AI agents and larger organizations.
What Is an AI Readiness Assessment?
An AI readiness assessment is a structured review of whether the conditions for one AI use case are in place: a measured problem, usable data, a way to test results, connected systems, accountable people, risk controls and a budget for running it. It collects evidence, not opinions, and ends in a decision on whether to proceed.
The term covers two different exercises. An organization-wide assessment scores general maturity. Microsoft's AI Readiness Assessment, for example, measures preparedness across seven pillars: business strategy, AI governance and security, data foundations, AI strategy and experience, organization and culture, infrastructure for AI, and model management Source [1]. A use-case assessment asks a narrower question: can we do this one thing well, now?
Both are useful, but only the second tells you whether to fund a specific project. A company can score well on general maturity and still lack the data for the use case on the table. Gartner makes the same point about data, saying AI-ready data must be "representative of the use case" it serves Source [2]. This guide focuses on the use-case assessment and borrows from organization-wide frameworks where they help.
| Aspect | Organization-wide assessment | Use-case assessment |
|---|---|---|
| Question it answers | How prepared are we for AI in general? | Can we build and run this one use case now? |
| Scope | Strategy, culture, skills, platforms and governance across the company | One process, one set of data, one group of users |
| Typical evidence | Survey answers and interviews | Data samples, access tests, process measurements and named owners |
| Output | A maturity score and a roadmap | Ready, ready with conditions, or not yet, with a list of gaps |
| Best used | When planning an AI program or annual budget | Before funding a build or a proof of concept |
Question it answers
- Organization-wide assessment
- How prepared are we for AI in general?
- Use-case assessment
- Can we build and run this one use case now?
Scope
- Organization-wide assessment
- Strategy, culture, skills, platforms and governance across the company
- Use-case assessment
- One process, one set of data, one group of users
Typical evidence
- Organization-wide assessment
- Survey answers and interviews
- Use-case assessment
- Data samples, access tests, process measurements and named owners
Output
- Organization-wide assessment
- A maturity score and a roadmap
- Use-case assessment
- Ready, ready with conditions, or not yet, with a list of gaps
Best used
- Organization-wide assessment
- When planning an AI program or annual budget
- Use-case assessment
- Before funding a build or a proof of concept
Why Readiness Gaps Sink AI Projects
AI projects often fail for reasons a readiness check can catch early: an unclear problem, missing data, technology chosen before the need, weak infrastructure, or a task too hard for current AI. Finding these gaps before a build costs a few meetings and some data pulls. Finding them after the build costs the build.
That list comes from RAND, which interviewed 65 experienced data scientists and engineers about why AI projects fail and identified five leading root causes Source [3]. The most common, in RAND's account, is a misunderstanding or miscommunication about the purpose of the project Source [3].
Data is the second recurring theme. In a July 2024 Gartner survey of 1,203 data management leaders, 63% of organizations either did not have, or were unsure whether they had, the right data management practices for AI. In February 2025, Gartner also predicted that through 2026 organizations will abandon 60% of AI projects unsupported by AI-ready data Source [2]. That is a forecast rather than a measured outcome, but it points to the same weakness RAND found.
Each of the seven checks below is aimed at one of these failure points, so a gap shows up as a named item on a list instead of a surprise halfway through a build.
The AI Readiness Assessment Framework: Seven Checks
This AI readiness assessment framework reviews one use case against seven checks: the problem and its baseline, data, evaluation, systems and integration, people and ownership, risk and controls, and running cost. Each check is marked ready, conditional or blocked based on evidence, and three of them (problem, data and risk) can stop a project on their own.
The checks are written for a single use case. If you are assessing several, run the framework once per use case and compare the results side by side, rather than averaging them into one company score.
1. The Problem and Its Baseline
Start with the problem, not the technology. Microsoft's AI adoption guidance opens use case identification with "Start with business problems," and asks teams to confirm the activity happens often enough to justify the investment Source [4]. This check needs three things in writing: the task the AI would help with, who does it today, and how it performs today in a number you can measure again later, such as hours per week, error rate or time to first response.
Ask, too, whether the task needs AI at all. A fixed, repetitive job may be done more cheaply and predictably by rules or robotic process automation, and our comparison of agentic AI vs RPA covers where each fits. If several tasks are competing for the first slot, our guide to agentic AI use cases sorts examples by how much autonomy each needs.
Red flag: the use case is described by its technology ("we want a chatbot") rather than by a task and a number. RAND lists a focus on the latest technology over the users' real problem as one of its five root causes Source [3].
2. Data
This check asks whether the data the use case needs exists, can be reached and can be trusted. Gartner's guidance is direct: "if the data has issues, then the data is not ready for AI" Source [2]. Ask for evidence rather than assurances:
- A sample of real records, pulled by the person who would supply them in production.
- The source of truth for each field, its owner and how often it changes.
- Who is allowed to see the data, and whether the AI system can be held to the same limits.
- Real examples of unusual and difficult cases, not only the common ones.
For document-based use cases, confirm that the documents are current and that outdated versions and duplicates can be excluded. Our guide to RAG architecture and retrieval quality explains why retrieval quality depends on this groundwork, and our generative AI integration article shows how to confirm data access for one workflow before choosing an architecture.
Red flag: the only evidence of data quality is someone's word that the data is fine.
3. Evaluation
A use case is only ready to build if you can tell whether the result is good. OpenAI's evaluation guidance starts from the point that "Generative AI is variable": the same input can produce different outputs, which makes traditional software testing insufficient on its own Source [5]. Before a build, you need a set of real cases with known good answers and a person qualified to judge the borderline ones.
This check is ready when a domain expert can describe a correct output for a typical case and for an awkward one, and has time set aside to review samples. Our guide to LLM testing before launch covers how to turn those cases into a test set.
Red flag: success is described as "it should feel accurate," with no example of a right or wrong answer.
4. Systems and Integration
An AI output has to land somewhere: a support ticket, a CRM record, an email draft or a dashboard. This check maps where the input comes from, where the output goes, and whether those systems can be reached through an interface your team controls. RAND counts inadequate infrastructure for managing data and deploying finished models among its five root causes Source [3].
Ask for the list of systems involved, whether each has an API (a defined interface other software can call) or an export, who can grant test access, and which systems would need to change for the output to be used. If an AI agent would take actions in those systems, add the agent checks later in this guide.
Red flag: someone would copy the output by hand from one screen to another, with no plan to change that.
5. People and Ownership
Every AI system needs a named owner who decides what good looks like, users who will actually work with it, and people with the skills to build and maintain it, in-house or through a partner. Microsoft's adoption guidance pairs skills with data in its maturity levels, from basic prompt and integration skills up to model training and orchestration, and matches each level to the kinds of use cases that are feasible Source [4].
Ask for the owner's name, the reviewers' available hours, the users who will test the system and the person who supports it after launch.
Red flag: the project has a sponsor but nobody whose daily work changes when it ships.
6. Risk and Controls
This check asks what could go wrong, how much it would matter and who signs off. The NIST AI Risk Management Framework calls this mapping: its map function establishes the context for an AI system and informs "an initial decision about appropriateness or the need for an AI solution" Source [6]. A valid outcome of the risk check is that AI is not the right tool for the job.
Ask how sensitive the data is, what happens if the AI is wrong, which outputs need a person's approval, which laws or contracts apply in your jurisdiction, and what your model provider stores. Our AI proof of concept guide sets out what to check in provider retention and training terms. If staff already use AI tools with no written rules, Fossilite's AI acceptable use policy template is a practical starting point, covering which tools are approved, what data can go in and who reviews outputs.
Red flag: nobody has asked legal or security, or the answer is "we'll sort it out after the pilot."
7. Running Cost and Operations
A system that works in a test still has to be paid for, monitored and maintained. This check estimates the cost per task at expected volume, names who watches quality after launch, and sets aside time for fixes when data, models or prices change. Our AI agent development cost guide shows how to turn a per-task cost into a monthly budget with a worked example.
Ask for the expected monthly volume, the most you would pay per task, who owns monitoring, and the budget for changes in the first year.
Red flag: the budget covers the build and nothing after it.
AI Readiness Assessment Template
Use the template below as a working AI readiness checklist. For each check, answer the question, attach the evidence and mark the check ready, conditional or blocked. A check without evidence cannot be marked ready, however confident the team feels, and the problem, data and risk checks can stop a project on their own.
| Check | Question to answer | Evidence to attach | Can block on its own? |
|---|---|---|---|
| Problem and baseline | Which task, for whom, and how does it perform today? | One-sentence use case and today's number | Yes |
| Data | Does the data exist, can we reach it, and can we trust it? | Real sample, source and owner per field, access rules | Yes |
| Evaluation | How will we know an output is good? | Real cases with agreed answers, named reviewer | No |
| Systems and integration | Where does input come from and output go? | System list, API or export per system, test access | No |
| People and ownership | Who owns it, uses it and supports it? | Named owner, reviewer hours, test users, support owner | No |
| Risk and controls | What could go wrong, and who approves? | Data sensitivity, approval points, provider terms, legal and security sign-off | Yes |
| Running cost and operations | Can we afford to run and maintain it? | Volume, maximum cost per task, monitoring owner, first-year budget | No |
Check
Problem and baseline
- Question to answer
- Which task, for whom, and how does it perform today?
- Evidence to attach
- One-sentence use case and today's number
- Can block on its own?
- Yes
Check
Data
- Question to answer
- Does the data exist, can we reach it, and can we trust it?
- Evidence to attach
- Real sample, source and owner per field, access rules
- Can block on its own?
- Yes
Check
Evaluation
- Question to answer
- How will we know an output is good?
- Evidence to attach
- Real cases with agreed answers, named reviewer
- Can block on its own?
- No
Check
Systems and integration
- Question to answer
- Where does input come from and output go?
- Evidence to attach
- System list, API or export per system, test access
- Can block on its own?
- No
Check
People and ownership
- Question to answer
- Who owns it, uses it and supports it?
- Evidence to attach
- Named owner, reviewer hours, test users, support owner
- Can block on its own?
- No
Check
Risk and controls
- Question to answer
- What could go wrong, and who approves?
- Evidence to attach
- Data sensitivity, approval points, provider terms, legal and security sign-off
- Can block on its own?
- Yes
Check
Running cost and operations
- Question to answer
- Can we afford to run and maintain it?
- Evidence to attach
- Volume, maximum cost per task, monitoring owner, first-year budget
- Can block on its own?
- No
How to Run an AI Readiness Assessment in Five Steps
Run an AI readiness assessment in five steps: write the use case down, name the people, collect evidence, score each check and decide. A focused assessment of one use case involves a business owner, a data owner, someone from IT or security and a future user, and it produces a written decision with a list of gaps.

Blue marks the steps that gather evidence. Red marks the step that makes the decision.
Step 1: Write the Use Case as One Sentence
Name the task, the user and the result you expect, in plain words. Microsoft suggests turning each business problem into a short statement that names the activity and the expected result Source [4]. An illustrative example: help support staff draft replies to warranty questions from the product manuals, so first responses go out the same day.
Step 2: Name the People
Bring in the people who can produce evidence, not only the people who want the project: the process owner, whoever controls the data, someone from IT or security, and at least one person who would use the output.
Step 3: Collect Evidence, Not Opinions
Ask for the items in the template: a data sample, an access test, today's numbers and the list of systems. A meeting where everyone agrees the data is fine is not evidence. A spreadsheet of real records with the awkward ones included is.
Step 4: Score Each Check
Mark each check ready, conditional or blocked using the rule in the next section, and write one line on why. Score with the people from step 2 in the room, so every conditional mark leaves with an owner and a date.
Step 5: Decide and Write the Gap Plan
Turn the scores into one of three decisions, list the gaps with owners and dates, and agree when you will reassess. The written decision is what you take to whoever funds the next step.
How to Score AI Readiness and Make the Decision
Score each check as ready (evidence in hand), conditional (a named gap with an owner and a date) or blocked (no evidence or no workable path). Any blocked mark on the problem, data or risk check means not yet. All seven ready means proceed. Anything in between is ready with conditions.
This rule is stricter than averaging on purpose. An average lets a strong score in one area hide a missing dataset in another, and a project with no usable data is not partly ready. NIST's generative AI profile takes a similar approach to deployment approval, recommending minimum thresholds for performance or assurance criteria in go or no-go decisions Source [7]. Readiness benefits from the same discipline: agree what blocks a project before you score it.
| Decision | Rule | What happens next |
|---|---|---|
| Ready | All seven checks marked ready | Plan a proof of concept with pass marks set in advance |
| Ready with conditions | No block on problem, data or risk, and at least one check conditional or blocked elsewhere | Close the named gaps first, or build closing them into the proof of concept plan |
| Not yet | Any block on problem, data or risk | Fix the foundation, choose a different use case, or drop the idea |
Decision
Ready
- Rule
- All seven checks marked ready
- What happens next
- Plan a proof of concept with pass marks set in advance
Decision
Ready with conditions
- Rule
- No block on problem, data or risk, and at least one check conditional or blocked elsewhere
- What happens next
- Close the named gaps first, or build closing them into the proof of concept plan
Decision
Not yet
- Rule
- Any block on problem, data or risk
- What happens next
- Fix the foundation, choose a different use case, or drop the idea
Figure 3 applies the rule to the illustrative warranty-reply example from step 1.

Blue marks ready checks. Gray marks conditions to close before or during the proof of concept. No check is blocked, so the decision is ready with conditions.
Extra Checks for Generative AI and AI Agents
A generative AI readiness assessment adds checks for outputs that can be confidently wrong, data that may leave your systems, and content that can manipulate the model. An AI agent readiness check adds two more: what the agent is allowed to do, and which of its actions need a person's approval first.
Generative AI Checks
NIST's Generative AI Profile lists twelve risks that are unique to or made worse by generative AI Source [7]. Three of them map directly onto the checks above:
- Confabulation, which NIST describes as "confidently stated but erroneous or false content" Source [7]. This raises the bar for the evaluation check: include cases that test whether the system says it does not know.
- Data privacy, covering leakage and unauthorized disclosure of personal or sensitive information. This sharpens the data and risk checks.
- Information security, including prompt injection, where instructions hidden in a document or web page change what the system does. This matters as soon as the system reads content you do not control.
NIST also advises against extrapolating generative AI performance from narrow, anecdotal assessments Source [7]. A few impressive demo answers are not readiness evidence.
AI Agent Checks
An AI agent chooses and takes actions, so readiness also covers what it may touch. Anthropic recommends finding the simplest solution possible and adds: "This might mean not building agentic systems at all" Source [8]. Before assessing an agent, confirm that a fixed workflow would not do the job. Our explainer on AI agents vs agentic AI vs LLMs helps confirm which of the three a use case needs.
If an agent is justified, add four items to the systems and risk checks: every action the agent may take, the permission each action needs, which actions need a person's approval, and how the agent stops when it is unsure. Our guide to building agentic AI systems covers permissions and security in detail.
Enterprise AI Readiness Assessment: What Changes at Scale
An enterprise AI readiness assessment uses the same seven checks but runs them across many candidate use cases, with one governance process deciding which go first. The main additions are a consistent scoring method across business units, a shared record of AI systems and their owners, and approval rules written once and applied to every project.
Microsoft's guidance suggests ranking use cases by scoring each on business impact, technical complexity, resource requirements and alignment with organizational goals Source [4]. Run the seven checks on the shortlist rather than the long list, so the effort of gathering evidence goes where it can change a decision.
NIST treats governance as cross-cutting: in the AI RMF, the govern function is designed to inform and run through the other three functions Source [6]. In practice, the risk check should not be reinvented for each project. Write the approval rules, data classes and review requirements once, and let each assessment apply them. Our guide to enterprise AI agents covers what happens after the first use case, from an owner for every agent to staged rollouts.
What to Do When the Answer Is "Not Yet"
A not-yet result tells you which foundation to fix and keeps money out of a build that would stall. The usual fixes are measuring the current process, cleaning and connecting the data, naming an owner, reducing the risk of the first use case, or accepting that a simpler tool would do the job better than AI.
- Blocked on the problem: measure the current process and rewrite the use case around the number you want to move.
- Blocked on data: start with data work, such as consistent definitions, tested pipelines and an owner for each dataset. SensViz's data analytics service starts from that foundation.
- Blocked on risk: choose an internal, lower-risk use case first, as Microsoft suggests for early proof of concept projects Source [4], or narrow the scope so a person approves every output.
- The task does not need AI: a rules-based workflow may be the better build. Our AI automation service covers workflows where triggers, rules and integrations do most of the work.
From AI Readiness to a Proof of Concept
A readiness assessment answers whether you can attempt a use case. A proof of concept answers whether it works on your data. When the assessment says ready, or ready with conditions, the next step is a short, time-boxed proof of concept with pass marks written before any results, using the evaluation cases the assessment identified.
Carry the outputs straight across. The one-sentence use case becomes the proof of concept's objective, today's baseline becomes the number to beat, and the evaluation cases become its test set. Our AI proof of concept guide covers how to scope, run and judge one, and how to decide whether to go, pivot or stop.
How SensViz Can Help
SensViz is an AI and software development company that builds custom AI, agentic and generative AI systems. Its custom AI development process includes a feasibility and data assessment that reviews data access and quality, integrations, risk, privacy and whether custom AI is the right approach, before any solution is designed Source [9].
The same page sets a commercial principle: discovery should produce a clear recommendation, not a commitment to use AI, and if a standard product, automation, analytics workflow or conventional feature is the better answer, SensViz should say so Source [9]. You can read more about how that work runs on the custom AI development page.
If you have a use case in mind and want a second view on its readiness, tell us about it.
Frequently Asked Questions
Short answers to questions that come up when planning an AI readiness assessment.
What Are the Pillars of AI Readiness?
AI readiness has no single standard set of pillars. Microsoft's assessment uses seven, from business strategy to model management Source [1], while Cisco's AI Readiness Index uses six: strategy, infrastructure, data, governance, talent and culture Source [10]. Both cover strategy, data, infrastructure and governance. A use-case assessment adds evaluation and running cost, because those decide whether one project can ship.
Is an AI Readiness Audit Different From an Assessment?
An AI readiness audit and an AI readiness assessment overlap, and neither term has a fixed definition. In practice, "audit" usually signals checking evidence, such as access logs, data samples and written policies, while "assessment" can also mean a self-scored survey. The evidence-based method in this guide works as an audit for one use case.
How Long Does an AI Readiness Assessment Take?
An AI readiness assessment has no standard length. Microsoft's online version is a set of multiple-choice and multiple-response questions you answer yourself Source [1]. An evidence-based assessment of one use case takes longer, because data samples, access tests and baseline numbers have to be gathered, so its pace usually depends on how quickly data owners respond.
Can a Small Business Run an AI Readiness Assessment Without a Consultant?
Yes. A small business can run an AI readiness assessment itself if it limits the scope to one use case and involves the person who does the work today, the person who controls the data and someone who understands the systems. Outside help is most useful for the evaluation, systems and risk checks, where technical judgment decides the score.
Should You Repeat an AI Readiness Assessment?
Yes, whenever something the assessment relied on changes: a new data source, a different model provider, new regulation, or a move from internal users to customers. NIST describes AI risk management as continuous and carried out throughout the AI system lifecycle Source [6], so treat the scorecard as a living record that is reviewed again before each expansion.
Ready to assess a use case?
Share your use case with the SensViz team and get a second view on its readiness before you commit to a build.
Sources
- [1] Microsoft. Microsoft Learn: AI Readiness Assessment (opens in a new tab).
- [2] Gartner. Gartner: Lack of AI-Ready Data Puts AI Projects at Risk (Q&A with Roxane Edjlali, 26 February 2025) (opens in a new tab).
- [3] RAND. RAND: The Root Causes of Failure for Artificial Intelligence Projects and How They Can Succeed (RR-A2680-1, August 2024) (opens in a new tab).
- [4] Microsoft. Microsoft Learn: Plan for AI adoption, Cloud Adoption Framework (opens in a new tab).
- [5] OpenAI. OpenAI: Evaluation best practices (opens in a new tab).
- [6] NIST. NIST Trustworthy and Responsible AI Resource Center: AI RMF Core, from the AI Risk Management Framework 1.0 (opens in a new tab).
- [7] NIST. NIST: Artificial Intelligence Risk Management Framework: Generative Artificial Intelligence Profile (NIST AI 600-1, July 2024) (opens in a new tab).
- [8] Anthropic. Anthropic: Building effective AI agents (opens in a new tab).
- [9] SensViz. SensViz: Custom AI development service page (opens in a new tab).
- [10] Cisco. Cisco: Cisco AI Readiness Index methodology (opens in a new tab).





