Most enterprises now use AI somewhere, but few can show what it adds to the bottom line. McKinsey’s 2025 State of AI survey of 1,993 respondents found that 88% of organizations use AI in at least one business function, yet only about one-third have begun scaling it, and just 39% report any impact on EBIT.
An AI maturity model shows which capabilities you have, which one is holding you back, and what to fund next. This blog covers five stages, six dimensions with a scoring rubric, 2025 results from DBS and Bank of America, and a 90-day roadmap for each move up the scale.
What Is an AI Maturity Model?
An AI maturity model is a scoring framework that places an organization on a defined scale of AI capability. Each level describes what the organization can reliably do with AI today and what it must build before it can do more. Used well, the model answers three questions: where you stand, what is holding you back, and where the next dollar should go.
An artificial intelligence capability model looks at the same problem through the capabilities themselves: data, platforms, talent, governance, and delivery. The two views belong together. A single stage label hides how uneven most companies are. A retailer can run demand forecasting in production while its customer data still sits in four systems with no shared customer ID.
Maturity measures repeatability. Counting tools, licenses, or pilots measures activity. A company with 40 pilots and none in production is less mature than a company with three models driving daily decisions and a tested process for shipping the fourth.
If you need a primer on scope before scoring, start with what enterprise AI covers in practice.
The Five AI Maturity Stages
Most enterprises sit at stage 3 or below: in McKinsey’s 2025 survey, nearly two-thirds of respondents said their organizations had not yet begun scaling AI across the enterprise. The table shows what each stage looks like in practice, the blocker that usually keeps a company there, and the signal that it has earned the next stage.
| Stage | What it looks like | Typical blocker | Ready to move up when |
| 1. Exploring | Teams test public AI tools on their own. No AI policy, no budget line, no inventory of use. | No executive owner and no acceptable-use policy | A named sponsor, an approved use policy, and a shortlist of use cases with estimated value |
| 2. Piloting | Several funded pilots with owners. Results are reported as stories with no measured baseline. | Data sits in silos that pilots cannot reach in production | One pilot has a measured baseline, a measured result, and a funded path to production |
| 3. Operationalizing | The first models run in production with monitoring, retraining, and a review process. | Every deployment is custom, so the second use case costs as much as the first | A shared platform and deployment process used by at least two teams |
| 4. Scaling | A shared AI platform serves several business units. Value is tracked on business dashboards. | Governance and change management lag deployment speed | AI outcomes appear in business unit targets as well as IT targets |
| 5. AI-Native | AI sits inside core decisions and products. Agents run end-to-end processes, and people review exceptions. | Keeping cost, control, and accountability clear as autonomy grows | An ongoing state, measured by the share of decisions and revenue that run through AI |
The split between stages 3 and 4 is deliberate. Running a first production model and running a shared platform for many teams need different budgets, different owners, and different skills. Treating them as one stage hides the most common place programs stall.
McKinsey’s 2025 data shows what the climb is worth. Only about 6% of respondents qualify as AI high performers: organizations that attribute 5% or more of EBIT to AI and report significant value. These companies behave like stage 4 and 5 organizations in this model:
- About three-quarters are scaling or have scaled AI, compared with one-third of other organizations.
- They are nearly three times as likely to have fundamentally redesigned workflows around AI.
- They are three times as likely to report that senior leaders show strong ownership of AI initiatives.
- More than one-third commit over 20% of their digital budgets to AI.
Workflow redesign stands out. McKinsey found it among the strongest contributors to meaningful business impact of the 31 factors it tested, which is why the exit signal for stage 4 requires AI outcomes in business unit targets.
The AI Maturity Framework: Six Dimensions to Score
A stage label tells you where you are. The dimensions tell you why. Score these six separately, because they rarely move together.
| Dimension | What it covers | Evidence to ask for |
| Strategy and value | Which business outcomes AI is funded to change, who owns them, how value is measured | A ranked use case portfolio with an owner, a baseline, and a target for each item |
| Data foundation | Quality, access, lineage, and security of the data AI depends on | Time it takes a new use case to get production data access; share of critical data with a named owner |
| Platform and technology | Shared infrastructure for building, deploying, and reusing models and agents | Number of teams using the same deployment path; components reused across use cases |
| Talent and operating model | Skills, team structure, and how business and technical staff work together | Who owns a use case after launch; AI literacy coverage for leaders and frontline staff |
| Governance and risk | Policies, review, security, compliance, and accountability for AI decisions | An inventory of AI systems in use, with a risk rating and a reviewer for each |
| Delivery and operations | Monitoring, retraining, incident response, and cost control in production | Drift and performance alerts in place; time from a detected issue to a fix |
Data is the dimension most companies overrate. A useful test: ask how many weeks it took your last pilot to get access to the production data it needed, and who approved it. If the answer involves a spreadsheet export, data is your ceiling. Our guide to data engineering services covers the pipelines and ownership models that close this gap, and data lakehouse architecture explains one common way to give analytics and AI the same governed source.
Delivery and operations are the dimension most companies skip. A model that nobody monitors loses accuracy as the data around it changes, and nobody notices until a business metric moves. Mature teams treat MLOps practices as part of the build, with versioning, automated retraining, and rollback planned before launch.
How to Run an AI Readiness Assessment
Most of an AI readiness assessment is evidence gathering; with the evidence ready, one business unit can be scored in a single working session. Run it in this order:
- Set the scope. Assess one business unit or one value stream first. A company-wide average hides the units that are ahead and the ones that are stuck.
- Build the panel. Bring together senior technical and data leaders. Add the business owners whose metrics AI is meant to change, plus someone from risk or compliance.
- Collect evidence first. Inventory every AI system and pilot, its owner, its data sources, and its measured result. Scores based on opinion drift upward.
- Score each dimension from 1 to 5 using the rubric below. Where the panel disagrees, take the lower score and record the evidence that would raise it.
- Set the stage. Your stage is the median of the six scores, rounded down, and never more than one level above your lowest score.
- Choose two dimensions to fund for the next 90 days. Usually that means the lowest score plus the one blocking your highest-value use case.
- Re-score every six months against the same evidence list.
| Dimension | Score 1 | Score 3 | Score 5 |
| Strategy and value | AI interest with no owner or budget | Funded use cases with owners; value measured for some | AI targets sit in business unit plans; value reported quarterly |
| Data foundation | Data copied by hand for each project | Shared pipelines for priority domains; access takes weeks | Governed data products; new use cases get access in days |
| Platform and technology | Tools chosen team by team | Standard stack for one or two teams | Shared platform with reusable components across units |
| Talent and operating model | A few enthusiasts working alone | Central team supports business units on request | Business and technical owners run use cases together; leaders trained |
| Governance and risk | No policy or inventory | Policy and review board; reviews happen after launch | Risk-tiered review before launch; full inventory; audit trail |
| Delivery and operations | Models run without monitoring | Monitoring for production models; manual retraining | Automated monitoring, retraining, rollback, and cost tracking |
Scores of 2 and 4 sit between the descriptions on either side.
The stage rule matters more than any single score. Take a company that scores 4 on strategy, data, platform, talent, and delivery, but 2 on governance. The median is 4. The cap is 3, one level above the governance score, so its stage is 3. That result is accurate: a company cannot scale AI across business units faster than it can review the risk of each deployment.
AI Maturity Examples
Two banks published detailed 2025 results that show what the upper stages look like in practice: the DBS 2025 annual report and Bank of America’s August 2025 Erica update.
| Organization | What they built | Published 2025 result | Stage in this model |
| DBS Bank | More than 2,000 models across 430+ use cases; DBS-GPT assistant made available bank-wide; DBS Joy, a generative AI chatbot for corporate and SME clients | About S$1 billion in economic value from data analytics and AI in 2025; DBS-GPT supports two-thirds of employees; 20,000+ corporate and SME customers used DBS Joy after its July 2025 launch; coding time cut by up to 20% on some data science tasks | 5 |
| Bank of America | Erica, an AI assistant built to be reused across consumer banking, Merrill, Benefits OnLine, and CashPro | More than 58 million client interactions per month; Erica handles over 40% of CashPro Chat client interactions; Erica for Employees used by 90%+ of staff and cut IT service desk calls by 50% | 4 |
Three patterns run through both examples.
Value is measured at the enterprise level. DBS reports a single economic value figure for AI across the bank, which is the stage 5 test in the stages table: leadership tracks and reviews AI’s share of value.
Reuse drives scale. Bank of America says its AI capabilities are designed to scale and reuse across business areas, so one assistant serves retail clients, wealth clients, corporate treasury users, and employees.
Employee adoption runs alongside customer-facing AI. Both banks report that most staff use an internal AI assistant, which builds the AI literacy the talent dimension measures. For a regional view, our look at DFW banks deploying AI agents covers where they start.
Building Your AI Transformation Roadmap
Plan the roadmap in 90-day cycles, each aimed at the exit signal for your current stage. The table lists what to fund in the next cycle and the gate that proves it worked.
| Move | Fund in the next 90 days | Exit gate |
| Stage 1 to 2 | Name an executive sponsor. Publish an acceptable-use policy. Train the leadership team. Rank 10 to 20 use cases by value and feasibility, and fund the top three as pilots with a measured baseline. | Three funded pilots, each with an owner, a baseline, and a target |
| Stage 2 to 3 | Pick the pilot with the clearest measured result and take it to production. Build the data pipeline it needs as a reusable asset. Stand up monitoring and a review process before launch. | One model in production, monitored, with a business owner reporting its result |
| Stage 3 to 4 | Turn the components from your first production use cases into a shared platform: deployment pipeline, data access, evaluation, guardrails. Move the second and third business units onto it. | Two or more business units shipping on the same platform; second use case delivered faster than the first |
| Stage 4 to 5 | Redesign core processes around AI decisions. Introduce agents with approval gates where governance scores allow. Put AI outcomes into business unit targets. | AI-driven decisions and revenue tracked as a share of the total and reviewed by leadership |
Three decisions shape every roadmap.
Buy or build at each stage. Early stages rarely justify custom platforms. AI as a service lets stage 1 and 2 companies test value without infrastructure spend, and our guide to choosing enterprise AI solutions covers how to compare vendors against your use case list.
Where the platform runs. Stage 3 usually forces the question of whether data and models move to the cloud. The cloud migration strategies most CTOs use apply directly to AI workloads.
Who builds the foundation. Data pipelines and model operations are the work most often underestimated. A dedicated data and ML engineering team, internal or external, keeps that work from landing on the same people running pilots. At stage 5, the design question shifts to AI-native architecture, where systems are built around AI decisions from the start.
Where AI Programs Stall and How to Get Past Each Point
Most stalls happen at stage 2, and they share a cause: pilots prove a model works, while production needs proof that a business process works with the model inside it. Six patterns account for most of them.
- Pilot sprawl. Dozens of pilots compete for the same data engineers. Cap active pilots at the number your team can take to production, and stop any pilot without a baseline after 60 days.
- No owner for the metric. IT owns the model, and nobody in the business owns the result. Assign every use case a business owner whose targets include its outcome.
- Data access by exception. Each pilot negotiates data access from scratch. Build the first production pipeline as a reusable data product with a named owner.
- Governance after launch. Reviews happen once a system is live, so teams either skip them or wait months. Tier use cases by risk and set review time limits for each tier.
- The second use case costs as much as the first. Nothing from the first deployment gets reused. Before funding use case two, list which components it will reuse from use case one.
- No way to test quality. Teams cannot show that a new version performs better than the last one. Build test sets and scoring before launch; our guide to AI agent evaluation methods covers accuracy, reliability, and cost metrics.
Where AI Agents Fit in the Maturity Model
You can pilot agents at stage 2, but they run safely in production only at stage 4 or above. An agent takes actions in real systems: it updates records, sends messages, and moves money. Every weakness in your governance and operations scores becomes a live risk the moment it acts.
The 2025 data points the same way. In McKinsey’s survey, 23% of respondents said their organizations are scaling an agentic AI system somewhere, but no more than 10% are scaling agents in any single business function. By industry, agent use is most widely reported in technology, media and telecommunications, and healthcare. Teams working on AI in healthcare face the strictest version of the governance requirement below.
The forecast is a warning. Gartner’s June 2025 agentic AI forecast predicts that over 40% of agentic AI projects will be canceled by the end of 2027 because of escalating costs, unclear business value, or inadequate risk controls. Gartner also predicts that by 2028, at least 15% of day-to-day work decisions will be made autonomously through agentic AI. The companies that reach that point will be the ones whose governance and operations scores support it.
Match agent autonomy to your governance score with four levels:
| Autonomy level | What the agent does | Minimum governance score |
| Assist | Answers questions and suggests next steps | 2 |
| Draft | Prepares outputs a person approves before anything happens | 3 |
| Act with gates | Takes defined actions, with approval required above set thresholds | 4 |
| Act and escalate | Runs a full process and routes exceptions to people | 5 |
Three capabilities separate companies that run agents from those that demo them. The first is an AgentOps discipline for monitoring agent behavior, cost, and failures in production. The second is identity and access management for AI agents, so each agent has its own credentials, scoped permissions, and audit trail. The third is a contained first domain. IT operations is common, and agentic AIOps shows how agents triage incidents before they touch customer-facing work.
Companies at stage 5 design new processes around agents from day one. The difference between AI-first and AI-native companies explains why that design choice changes how the business operates.
Conclusion
An AI maturity model earns its keep when it changes where money goes. Score the six dimensions against evidence, apply the stage rule, and fund the two dimensions that cap your progress for the next 90 days. In most companies, that means data access and governance before more pilots.
McKinsey’s 2025 data shows why the effort pays: about three-quarters of AI high performers are scaling AI, compared with one-third of other organizations, and they are nearly three times as likely to have redesigned workflows around it. The first production model with an owner, a baseline, and monitoring is the most important milestone. Re-score in six months and compare against the same evidence list.
Frequently Asked Questions
What are the stages of an AI maturity model?
This model uses five: Exploring, Piloting, Operationalizing, Scaling, and AI-Native. Other frameworks, including Gartner’s AI Maturity Model, also use five-level scales. The labels differ, but each one tracks the same progression of capability, repeatability, and business impact.
What is the difference between an AI readiness assessment and an AI maturity assessment?
A readiness assessment checks whether you can support a specific AI initiative now: data, skills, budget, and governance for that use case. A maturity assessment measures your overall capability across the organization and is repeated over time to track progress. Most companies run a readiness check before each major investment and a maturity assessment once or twice a year.
How long does it take to move up a stage?
Plan in 90-day cycles and expect several cycles per stage. The technical work is rarely the slow part. Assigning business owners, changing review processes, and building shared data access take longer, because they change how teams work.
Can a company skip a stage?
A company can move faster through a stage by buying capabilities it would otherwise build, such as a managed AI platform or an outside data engineering team. It cannot skip the capabilities themselves. A company that deploys agents without stage 3 monitoring and governance is running stage 2 pilots with production risk.
Which AI maturity framework should we use?
Use one with stages tied to observable evidence, and keep using it. Gartner’s AI Maturity Model and the six dimensions McKinsey uses to explain AI value (strategy, talent, operating model, technology, data, and adoption and scaling) are common references. The six-dimension rubric in this guide adds the scoring detail needed to decide what to fund. Consistency matters more than the choice: a score only means something when you can compare it with the last one.
How does generative AI change AI maturity?
Generative AI lowers the cost of stages 1 and 2, because teams can test useful tools in days without training models. It does not lower the bar for stage 3. Production generative AI still needs governed data, evaluation, monitoring, and clear ownership, and those take the same organizational work as before.
Why Choose [x]cube LABS?
[x]cube LABS works with enterprise teams to design and deploy AI agents across complex, regulated environments.
We help enterprises become AI-native, not by adding AI on top of existing systems, but by rebuilding the intelligence layer from the ground up. With 950+ products shipped and $5B+ in value created for clients across 15+ industries, here is what we bring to the table:
1. Autonomous AI Agents
We design and deploy agentic AI systems that sense, decide, and act without human bottlenecks, handling complex, multi-step workflows end-to-end with measurable resolution rates and no manual intervention.
2. Enterprise Voice AI
Our voice AI platform, Ello, puts production-ready voice agents in front of your customers in minutes. Zero-latency conversations across 30+ languages, with no call centers and no wait times.
3. AI-Powered Process Automation
We replace manual, error-prone workflows with intelligent automation across invoicing, compliance, customer service, and operations, freeing your teams to focus on work that requires human judgment.
4. Predictive Intelligence and Decision Support
Using machine learning and real-time data pipelines, we build systems that forecast demand, flag risk, optimize inventory, and surface strategic insights before your teams need to ask for them.
5. Connected Products and IoT
We design and build IoT platforms that turn physical devices into intelligent, connected systems with built-in real-time monitoring, remote management, and condition-based automation.
6. Data Engineering and AI Infrastructure
From data lakes and ETL pipelines to AI-ready cloud architecture, we build the foundation that makes everything else possible, scalable, reliable, and designed to grow with your business.
If you are looking to move from AI experimentation to AI-native operations, let’s talk.