Enterprise AI works best as a staged programme: confirm readiness, name an owner, pick a modest first use case, put governance in place, then measure return. Skipping a stage is the usual reason enterprise ai efforts stall.
| Phase | What the evidence marks | Planning use |
|---|---|---|
| Awareness | Moving from Unaware to Aware means the organisation sets a budget for its first AI project and identifies AI champions. | Fund a modest first project and name champions before buying platforms. |
| Readiness | AIRI v3.1 assesses organisations on a 0-5 scale across five readiness levels. AI Ready (Level 2) is the recommended target for most organisations. | Treat Level 2 as a sensible goal, not a ceiling to outgrow quickly. |
| Team | Aware to Ready means assembling an AI project team that can work with stakeholders. Ready to Competent means expanding to a full-fledged AI Engineering team. | Grow the team only as the portfolio of projects justifies it. |
| Return | Deloitte reports satisfactory ROI on a typical AI use case within two to four years. | Set patient review points rather than judging a project at its first budget cycle. |
Phase markers come from AI Singapore’s AIRI readiness framework and the payback range from Deloitte’s 2025 survey, which together set a realistic pace for enterprise ai programmes.
Each phase below connects to a deeper guide. For the techniques that sit underneath, start with machine learning basics. For the products on the market, see ai tools for business.
What enterprise AI means in practice
Enterprise AI is the use of machine learning and related techniques across a large organisation’s operations, backed by shared budget, data, governance and people. The word that matters is shared. A team that trials a writing assistant on its own is experimenting. An organisation that decides which functions get AI, who is accountable for it and how results are judged is running a programme.
That distinction shapes every later decision. Experiments can be cheap and informal because a failure does little damage. A programme needs consistent rules, because the same vendor or model may end up touching finance, HR and customer data, and a mistake in one place can spread to others.
Useful questions to settle at the outset are simple. Who owns the programme? What is the organisation actually ready to do? How will anyone know it worked? The phases that follow answer those questions in order. Real programmes overlap them, though. Governance work often starts while the first project is being scoped, and the way success will be measured belongs at the beginning rather than the end.
Where an organisation starts depends on where it already is:
- No budget and no owner yet. Begin with the awareness phase. Spending on tools before ownership is settled makes the later phases harder.
- Pilots already running without shared rules. Go straight to the governance and return phases. Retrofitting rules and baselines onto live pilots is untidy, but it is cheaper than discovering the gaps after wider rollout.
- An established team and several live projects. Concentrate on architecture and measurement, and use the readiness phase as a periodic check rather than a starting gate.
Awareness phase: budget, champions and sponsorship
AI Singapore’s AIRI framework describes the step from Unaware to Aware in practical terms: the organisation sets a budget for its first AI project and identifies AI champions. Both are concrete acts rather than a mood of enthusiasm, which makes them easy to check. Either the budget exists and the champions are named, or they are not.
Sponsorship matters just as much. Deloitte’s report on AI returns finds that in ten percent of organizations the CEO is the primary leader of the AI agenda. Leadership therefore sits elsewhere in most organisations. That is not a problem in itself, but it makes a named sponsor with authority across functions important, because AI work regularly crosses reporting lines.
Actions for the awareness phase
- Name a sponsor. Choose someone able to settle disputes between functions and to protect the project when priorities shift.
- Set a budget for a single modest project. A sum sized for an organisation-wide rollout invites scrutiny that an early project cannot yet answer.
- Identify champions. Look for people in the functions most likely to benefit, and release some of their time. A champion with no time is a title, not a role.
- Describe the change a good outcome would bring. Use terms the sponsor would recognise, such as work removed from a process or errors caught earlier.
This phase suits organisations still debating whether to begin. It also tells an organisation when to wait. If nobody is willing to own the work, hold off on procurement, because tools bought without an owner are hard to defend later.
Readiness phase: checking where the organisation stands
Running an ai readiness assessment turns a general feeling of being ready or not into evidence that leaders can discuss. AI Singapore’s AIRI is one structured option. Its versions 1.0 and 2.0 (2019-2025) distilled the critical success factors, and AIRI v3.1 consists of five pillars that map to fifteen dimensions, as set out on the AI Singapore AIRI page.
Reading the scale
AIRI v3.1 assesses organisations on a 0-5 scale across five readiness levels. Its stated recommendation is that AI Ready (Level 2) is the target for most organisations. That is a conditional recommendation for most organisations, not a requirement and not a maximum. Reaching Level 2 is a sensible objective for an organisation planning real projects, and a higher level is worth pursuing only when the portfolio of work demands it.
The spread of five pillars across fifteen dimensions carries a practical lesson. Readiness is wider than the technology function. A review that involves only IT will miss the dimensions that sit with business leaders, and it will probably flatter the result.
Running a useful review
- Gather a mixed group. Include the sponsor, the champions and representatives from the functions that would use the output.
- Score candidly. A generous score feels pleasant and hides the gap that will later cause delay.
- Find the weakest dimensions. Those, rather than the average, determine where the first project is likely to struggle.
- Choose a target level and a date to revisit it. Readiness changes as people and systems change, so a single review loses value with time.
The review also guides project choice. If data access looks weak, favour a first project that works with data the organisation already controls. If skills look thin, favour a use case where a ready-made product can do most of the work.
Strategy phase: choosing where to start
An enterprise ai strategy is a short list of decisions rather than a long document: which problems to tackle first, who is accountable, what the organisation will build or buy, and how progress will be reported. A strategy that cannot be summarised in a few paragraphs is unlikely to guide anyone’s day-to-day choices.
On where to begin, practitioner guidance on creating an enterprise AI strategy suggests that focusing on IT, HR, finance and procurement is often a good starting point. A plausible reason is that these are internal functions with defined processes and users who can give prompt feedback, so the cost of an early mistake stays contained. That is a reason to consider them, not a rule, and an organisation with a pressing problem elsewhere should weigh it on its merits.
Screening candidate use cases
A candidate project is stronger when it meets most of the following tests:
- A named owner. Someone in the business, not only in the technical team, wants the result and will accept accountability for using it.
- Accessible data. The information the project needs already exists and can be used lawfully and without a long preparation effort.
- A measurable outcome. The change can be observed in a number the organisation already tracks, such as handling time or error rate.
- Tolerable downside. If the output is wrong, a person can catch it before harm results.
- A realistic route to adoption. The people who must change how they work have been consulted.
Projects that fail the first test rarely recover, however strong the technology. Projects that fail the fourth need stronger governance before they proceed, which is the subject of the next phase.
Build, buy or blend
Early projects usually favour bought or configured products over custom builds, because they reduce the skills the organisation must hold at the start. The ai tools for business guide explains what categories of product exist and how to compare them. Custom development becomes more attractive when the problem is specific to the organisation and the readiness review shows the skills and data are in place. For concrete examples of where organisations apply AI, see enterprise ai use cases.
Governance phase: rules before scale
Governance should arrive before AI spreads, not after. It sets who can approve a use, what risks must be considered and how problems are reported. Each organisation needs its own ai governance framework, and published frameworks provide the raw material. They differ in purpose, which is what the comparison below shows.
| Framework | What it is | Planning use |
|---|---|---|
| OECD principles | Non-binding but influential principles that emphasize a rights-based approach. | Use as a values baseline when drafting policy wording, not as a compliance checklist. |
| NIST AI Risk Management Framework | Released in January 2023 as a voluntary set of guidelines. It has four core functions: Govern, Map, Measure and Manage. It is not subject to formal certification schemes. | Use as a working structure for internal risk practice. There is nothing to be certified against. |
| ISO 42001 | Designed for compliance certification, unlike the AI RMF. | Consider when customers, partners or auditors need external assurance. |
| UNESCO | Calls for comprehensive AI impact assessments to identify risks and benefits. | Run an impact assessment before significant uses go live. |
These descriptions follow Bradley’s overview of global AI governance frameworks.
Choosing between guidance and certification
The frameworks do not compete so much as serve different needs. The NIST AI RMF suits an organisation that wants a practical way to organise its own risk work. Its four functions can be read as a sequence of questions: who decides, where is AI used, how is performance and risk checked, and what happens when something goes wrong. ISO 42001 suits an organisation that must show an outside party it has met a standard. Many organisations begin with the guidance and consider certification later, once their practice is stable enough to be audited.
Turning principles into practice
Principles only help if they change decisions. The broader guide to responsible ai in business covers how values such as fairness and transparency become review steps and approval gates. A practical minimum is an inventory of where AI is used, an approval route for new uses and a method for assessing impact that matches the importance of the decision. UNESCO’s call for impact assessments is a useful prompt for the last of these, because it makes teams state risks and benefits before a launch rather than after.
Build phase: team and architecture
AIRI ties growth in capability to growth in the team. Moving from Aware to Ready is marked by assembling an AI project team that can work with stakeholders. Moving from Ready to Competent is marked by expanding that team to a full-fledged AI Engineering team, according to the AIRI framework.
The order is the lesson. A project team that works closely with the business comes first, and an engineering function follows once there is enough sustained work to occupy it. Hiring a large engineering team before there is a portfolio of projects creates cost without direction. Remaining with a single overstretched group after demand has grown creates delay and dependence on a few people.
What the architecture needs to do
An enterprise ai architecture is the set of building blocks that lets projects share what they need instead of rebuilding it each time. Typical blocks include access to data, an environment for developing and deploying models, connections into the business systems people already use, monitoring of how models behave over time, and security controls. The detailed guide covers each in turn.
The practical rule for this phase is to standardise only what several projects already share. Designing a grand platform before the first project has shipped usually produces components nobody asked for. Letting each project choose its own approach, on the other hand, multiplies maintenance and makes governance harder. A reasonable middle path is to let the first projects reveal their common needs, then convert those needs into shared services.
Return phase: measuring what AI delivers
Money is flowing into AI faster than clear evidence of payback. In Deloitte’s 2025 survey of 1,854 executives across Europe and the Middle East, 85 percent of organizations increased their investment in the past 12 months, and 91 percent plan to increase it again this year. The same Deloitte analysis of AI returns reports achieving satisfactory ROI on a typical AI use case within two to four years.
The planning consequence is about patience and honesty. A payback horizon of two to four years is longer than many annual budget reviews. A project judged only at its first annual review may look like a failure when it is behaving normally. The opposite error is equally costly: rising budgets can make a programme feel successful simply because spending is growing.
Setting up measurement
The guide on how to measure ai roi sets out methods in detail. These habits make the biggest difference:
- Record a baseline before launch. Without a figure for how the process performed beforehand, any later improvement is a guess.
- Agree the measures with the business owner. A result the owner does not accept will not secure further funding, however sound the analysis.
- Track early indicators as well as final returns. Adoption and error rates show whether a project is on course long before financial payback appears.
- Count the full cost. Include people, data preparation, integration and ongoing monitoring, not only licences.
- Schedule reviews at points that match the expected horizon. Then a slow start does not trigger a premature cancellation.
Measurement also protects the programme politically. A sponsor who can show a clear baseline and a credible path to return is in a much stronger position when budgets are contested.
Troubleshooting stalled programmes
Programmes usually stall for ordinary organisational reasons rather than technical ones. The wider patterns are explored in enterprise ai adoption challenges. The checks below map common symptoms back to the phase where the cause usually lies.
- Enthusiasm without progress. Return to the awareness phase. Check that funding is real and that the champions have time and a sponsor who will back them.
- Pilots that never reach daily use. The cause is often that nobody in the business owns the outcome. Reassign ownership to the function that benefits and revisit the screening tests for use cases.
- Disagreement about whether to proceed. This tends to mean the organisation has not agreed its readiness target. Repeat the readiness review with a wider group and settle the level to aim for.
- Policies that exist but are ignored. Governance written in isolation rarely survives contact with real projects. Simplify the approval route, attach it to decisions people already make and check that it gives answers quickly.
- Delivery capacity that cannot keep up. Compare the volume of requested work with the team described for the current stage. If demand is sustained, plan the step up to a larger engineering function. If it is patchy, consider ready-made products instead.
- Results nobody trusts. Look at the baseline and the measures agreed with the business owner. Disputed results usually trace back to measures chosen after the fact.
- Fading leadership attention. Short, regular reports on outcomes in the sponsor’s own terms keep the programme visible without demanding long meetings.
When several symptoms appear together, start with ownership. Most of the others become easier to fix once a named person is accountable for the programme and its results.
Frequently asked questions
How long does the free AIRI self-assessment take?
AI Singapore lists its online assessment as 20 minutes, 15 questions and free. That is short enough to complete in a leadership meeting, which makes it a practical way to start a discussion about readiness.
Is AI Aware enough to begin?
AIRI says organisations should minimally aspire to be AI Aware. Treat that as a floor for starting small, well-owned projects. An organisation planning to move beyond pilots should set its sights on the higher target discussed in the readiness phase.
Can smaller organisations pursue enterprise AI without building a team?
Yes. AI Singapore notes that smaller organisations could adopt ready-made AI solutions for their core or peripheral business activities. Bought products let them gain the benefit without carrying the cost of specialist engineers.
Do governance frameworks replace one another?
No. They serve different purposes. In Bradley’s overview of AI governance frameworks, the NIST AI RMF is voluntary guidance with no certification scheme, while ISO 42001 is designed for compliance certification. Many organisations use the guidance for daily risk practice and add certification only when outside assurance is needed.
