In many organizations, the AI conversation starts with the same question: which new tool should we buy? Yet a good part of the answer often already sits inside the software the organization uses. AI features are increasingly being added as updates to enterprise software such as ERP, CRM, human resources, office and collaboration applications. In this article, we look at why the AI features embedded in existing software are coming to the fore, and how to put them into an inventory before making a new investment decision.
Spending is growing, and the direction is shifting
According to a forecast Gartner published on September 16, 2026, worldwide AI spending will reach $2.7 trillion in 2026, which represents a 49.5% increase over the previous year. In the same table, spending on AI software rises from roughly $288 billion in 2025 to roughly $462 billion in 2026.
As notable as the size of the figure is where this spending is going. According to Gartner, software vendors of every kind are rapidly embedding agentic AI into their existing products to maintain their position in the market. Organizations, for their part, are turning to the simpler embedded features offered by the software providers they already work with, at a time when generative AI sits in the "Trough of Disillusionment" in 2026, the phase in which expectations have fallen. Gartner lists the purposes of these features as operational efficiency and workflow automation, improving customer engagement, and strengthening decision-making processes.
Gartner also finds that organizations look to external service providers for support less with comprehensive transformation projects than with smaller, indirect projects aimed at exploiting the AI features of their existing software.
The number of embedded features is growing fast
This trend is not new. In a prediction made in August 2025, Gartner stated that most enterprise applications would have embedded AI assistants by the end of 2025. According to the same prediction, 40% of enterprise applications will be integrated with task-specific AI agents by the end of 2026, up from less than 5% in 2025.
Gartner also draws attention to an important distinction here: assistants make tasks easier but depend on human input and do not operate independently. Agents, on the other hand, specialize in a particular task and can carry out complex work from end to end. Calling assistants "agents" is, in Gartner's words, a common misconception. This distinction is also necessary for classifying correctly what each feature does when you build an inventory.
The practical upshot is this: while an organization is debating whether to buy a new AI tool, a feature that could meet the same need may already be waiting to be switched on in software it uses, or may even have been switched on without anyone noticing.
Embedded features carry risks too
Although embedded features are easy to set up, they raise questions of their own. Gartner notes that the risks of vendor lock-in, data sovereignty and runaway costs are not deterring buyers from adopting these specialized capabilities. Overlooking a risk does not mean it has gone away. According to Gartner, organizations expect their providers to help them manage costs and to embed usage tracking into their workflows so that success can be evaluated.
This is exactly where an inventory proves useful: an organization that does not know which feature is switched on, what data it accesses, who uses it and how much it is used will find these risks hard to manage.
Why is an inventory a governance tool?
The U.S. National Institute of Standards and Technology (NIST) defines the inventory as one of the core elements of governance in its AI Risk Management Framework, published for voluntary use. According to the framework's implementation playbook, organizations should have mechanisms to maintain an inventory of AI systems, and these mechanisms should be resourced according to risk priorities.
The playbook says an inventory makes it possible to answer the following questions quickly: when was a given model last refreshed, how many models are currently in use across the organization, and how many users are affected by those models? According to the playbook, partial inventories may not deliver the value of a full inventory; the preferred approach is to record all systems or, at a minimum, to start with the high-risk ones. The same framework also recommends defining processes for the safe decommissioning of AI systems.
A practical inventory method
The steps below are a general framework derived from the sources above; each organization needs to adapt it to its own software landscape and risk priorities.
- Identify the owner. Appoint the person or team that will own the inventory from the start. Involving representatives from IT, business units, legal and information security keeps the inventory from being confined to a single viewpoint.
- Start with the software list. List every application the organization uses, working from license and subscription records. For each one, go through the provider's release notes and admin console and pull out its AI features one by one.
- Classify each feature. Record whether the feature is switched on or off, whether it is an assistant or an agent that carries out tasks on its own, which business process it is used in, and who has access to it.
- Document the data flow. Note which data the feature accesses, whether that data leaves the organization, and whether personal data is processed. This information lays the groundwork for contract and regulatory assessments; an inventory alone does not replace a legal assessment.
- Measure usage and cost. Where possible, record how often the feature is used and its cost impact, whether license-based or consumption-based. This step prevents setting aside new budget for a feature that nobody uses, or overlooking the cost of one that is in use.
- Match features to needs. When a request for a new tool comes in, check the inventory first: is there already a feature that could do the same job? If there is, try it first in a limited pilot; if not, move on to evaluating a new tool.
- Assess dependency. Write down how far you depend on a single provider's feature in critical processes, and what would happen if that feature were switched off or changed. Define the decommissioning steps at this stage as well.
- Update regularly. The inventory goes stale as providers add new features through updates. Set a review frequency and make one person responsible for tracking release notes.
Conclusion
Gartner's data shows that a significant share of enterprise AI use comes not from new platforms but from features added to existing software. This creates both an opportunity and a visibility problem: as features multiply quickly, it becomes harder to know which one runs where, on what data and at what cost. The inventory approach NIST recommends is a simple and practical way to gain that visibility.
Our view at Orient Global is this: before a budget request for a new AI tool reaches the table, putting together a short inventory of the features in the software you already use is often the most productive first step. An inventory both prevents unnecessary investment and clarifies the areas where a genuinely new tool is needed.
A note on logistics. Orient Global's consulting team includes a senior executive with more than 30 years in operations, logistics and supply chain. That executive has run logistics, warehousing, technical service and administrative operations end to end at a technology distribution company operating at a scale of roughly one billion dollars, leading teams of more than 100 people, and has designed and directed large-scale logistics and distribution networks, including nationwide warehouse and transport operations, across leading retail, logistics and distribution organizations. We bring the same discipline to small and mid-sized businesses with AI and automation, and it applies directly to the inventory described above: logistics chains run on many connected systems, such as ERP, warehouse, transport and order management, so start by listing what is in use, check which of those systems already include AI features, and only then decide where something new is really justified.
#interimmanagement #forwarddeployedengineering