Embedded AI in the software you already own

Gartner's September 16, 2026 forecast contains two numbers that together tell one story. Worldwide AI spending will reach $2.7 trillion in 2026, up 49.5 percent year over year, yet despite that scale generative AI has landed in Gartner's Trough of Disillusionment. The point behind this apparent contradiction is where value flows. Companies are ordering fewer large, custom transformation projects and increasingly using simple AI features that vendors have embedded into the software they already run. For leadership, the signal is that advantage today comes not from another purchased system, but from making good use of the AI in tools the company already pays for, while keeping cost, data, and output quality under control.

What happened

On September 16, 2026, Gartner published its forecast for worldwide AI spending. It projects that spending will reach $2.7 trillion in 2026, a 49.5 percent increase over the prior year. Behind that number sits a shift in what companies spend on.

Gartner places generative AI in the Trough of Disillusionment, the stage in its technology maturity cycle where, after a period of inflated expectations, some deployments did not deliver the promised results and the market brings its approach back to reality. At the same time, software vendors are rapidly embedding agentic AI features into their existing products to defend their market position against new, cross-functional agents that operate above individual applications.

The result is that companies reach for simpler, ready-made AI features from vendors whose software they already use, and they do so to raise operational efficiency, automate workflows, improve customer engagement, and support decisions. Gartner states plainly that buyers are shifting away from ordering large business transformation and toward smaller, indirect projects whose purpose is to exploit AI features in the software they already have.

Notably, the risks tied to this approach, namely vendor lock-in, data sovereignty, and run-away costs, are not deterring buyers from adopting these proprietary features. Gartner adds that companies turn to providers to help manage cost and to embed usage tracking into workflows, so they can judge whether a given feature pays off.

This picture lines up with other 2026 research. Deloitte's State of AI in the Enterprise report, based on a survey of 3,235 leaders across 24 countries, shows that worker access to AI rose by about half in a year and that twice as many leaders as a year earlier report transformative impact, yet only about 34 percent of companies are genuinely reimagining how the business works. McKinsey's annual State of AI survey finds that adoption is widespread, but most organizations have not yet started to scale AI across the company. In other words, easy gains come from ready-made features, while the harder, deeper value still waits on process redesign.

Why it matters for business

For leadership, the most important takeaway concerns purchasing priorities. If the market is shifting from large projects toward ready-made AI features in existing software, the first question is no longer "which AI system should we buy," but "which AI features have we already paid for, and are we using them." For many companies, the cheapest and fastest value sits in tools already in place: the office suite, the sales system, the accounting software, or the ticketing tool.

That also changes how you calculate returns. A ready-made AI feature wired into a tool employees already use has a shorter path to impact than a separate system, because it needs no new integration, no new login, and no training from scratch. A shorter time to first result is an advantage, especially for a smaller company without the budget and team for a multi-month project.

There is a second side, though. Gartner's Trough of Disillusionment signal is a reminder that a ready-made feature alone does not move the result if the process around it does not change. Minutes saved on a single task stay with the employee as convenience, not as lower company cost, until someone redesigns the workflow and removes work from it. So choosing a ready-made feature makes the start easier, but does not remove the decisions about who owns the process and how the result is measured.

Then there is the bill. Demand for AI infrastructure, including AI-optimized IaaS, is growing 96 percent in 2026 by Gartner's estimate and stays relatively insensitive to price, which shows up in usage-based billing models. A feature that looks cheap at the start can, at high volume, generate a bill that grows faster than the benefit. Cost control stops being a technical topic and becomes part of the business decision.

Business use cases

Office suite, email, and documents

Business problem: employees lose time writing repetitive emails, summarizing long documents, and preparing meeting notes, and each of these tasks looks minor on its own but consumes hours a week across a team.

AI solution: AI features embedded in the office suite and email, meaning an assistant that drafts emails, summarizes documents, and produces notes and task lists from meetings, with the employee reviewing and approving.

Data and process to put in place: organized access to documents and calendar, clear rules on which content may be passed to the assistant, and a short guide on when to use AI output and when to write it yourself.

Potential effect: less time on administrative work and faster flow of information, with the same team.

Risk and limitation: the assistant can produce content that sounds credible but is wrong, so client-facing writing and decisions need human review, and sensitive data requires clear rules on what may and may not be shared.

Sales and CRM

Business problem: sales reps spend a lot of time updating the sales system, writing emails to clients, and preparing quotes, which leaves less time for conversations and closing deals.

AI solution: AI features embedded in the sales system that suggest message text, prepare call summaries, organize customer data, and flag which opportunities need attention, while the rep makes the decisions and approves the quote.

Data and process to put in place: current data in the sales system, a consistent price list and discount rules, customer contact history, and a rule that a person approves the final quote and terms.

Potential effect: more opportunities handled in the same time and a shorter sales cycle, which translates into revenue rather than just comfort.

Risk and limitation: a suggestion based on outdated data leads to a wrong quote, and handing the model discount decisions that should stay with a person is a risk to watch.

Customer service and tickets

Business problem: the support team is drowning in repetitive queries, response time is rising, and adding more people raises cost without improving quality.

AI solution: AI features embedded in the ticketing tool that propose an answer based on the knowledge base and customer history, pre-categorize tickets, and for simple cases prepare a ready reply for approval, routing hard or risky cases to a person.

Data and process to put in place: an organized and current knowledge base, ticket categorization, access to customer history, and clear rules on when a reply can go out automatically and when it needs approval.

Potential effect: shorter response time and more tickets handled without growing the team.

Risk and limitation: with an incomplete knowledge base the model can give a wrong answer, so a confidence threshold and quality control are needed, and the rate of tickets handled without a person must be distinguished from genuinely resolved cases, because closing a conversation is not the same as solving the problem.

Accounting, invoice flow, and reporting

Business problem: accounting and administration spend a lot of time manually re-keying invoices, matching payments, and preparing repetitive reports.

AI solution: AI features embedded in accounting software or an ERP system that read documents, pre-classify them, match them to orders and payments, and prepare draft reports for a person to check and approve.

Data and process to put in place: a structured document flow, integration with the accounting system, coding and approval rules, and control over who approves items above set thresholds.

Potential effect: a shorter month-end close and less manual work, while a person remains accountable for the books.

Risk and limitation: errors in financial data are costly, so verification and a full audit trail of every operation performed by the model are required, and financial data needs special attention to the question of where it goes.

What companies can do now

The first step is to take stock of what the company already has. Before deciding on a new purchase, check which AI features are already available in the tools the company pays for: the office suite, the sales system, the accounting software, the ticketing tool. It often turns out that some capabilities are already paid for but unused.

The second step is to pick one process that genuinely costs time or money and is repetitive. Rather than turning on every feature at once, treat one area as the first full project, for example handling queries, quoting, or invoice flow.

The third step is the owner and the data. Someone in the company has to be responsible for seeing the deployment through and have the mandate to change how work is done. In parallel, clean up the data the AI will work on, because even a good feature returns results you cannot trust if the knowledge base, price list, or documents are out of date.

The fourth step is security and data sovereignty. Before turning on a vendor's AI feature, establish which data may be passed to it, where that data goes and how it is processed, and write down short rules for employees. That limits both legal risk and shadow AI, meaning the use of tools outside the company's control.

The fifth step is cost and measurement. Turn on usage tracking from the start, set limits and alerts, attach cost to a process and an owner, and pick one hard metric, response time, cost per case, or month-end close time, and measure it before and after. Without that number you cannot tell real value from the impression of being modern, and a usage-based bill easily slips out of control.

Risks

The first risk is vendor lock-in. The deeper a company builds a process on one vendor's proprietary feature, the harder and more expensive it becomes to change it later. Gartner notes that this risk is not deterring buyers, which does not mean it disappears, so it is worth knowing from the start what an exit would look like.

The second risk is data sovereignty and security. Turning on an AI feature means company data is processed by the vendor's model, so you need to know what data goes there, where it is stored, and under what terms. Without that it is easy to breach requirements or lose control of sensitive information.

The third risk is run-away costs. Usage-based billing means the bill grows with volume, and at large scale it can grow faster than the benefit. Without limits, usage tracking, and cost attached to a process, the company pays for something it does not measure.

The fourth risk is shadow AI and the absence of a process owner. When employees turn on AI features themselves and use them their own way, without anyone responsible and without rules, value scatters and risk rises. A deployment without an owner ends up as a tool nobody steers.

The fifth risk is the absence of output quality control. The model can produce an answer that sounds credible but is wrong. Without a confidence threshold, human review, and an audit trail, such errors enter the process and undermine trust in the whole deployment. In customer service you also have to separate cases closed without a person from those genuinely resolved.

Key takeaways

Gartner's September 16, 2026 forecast shows a market that spends more on AI, $2.7 trillion in 2026, while bringing expectations of generative AI back to reality. Value is shifting from large transformation projects toward simple AI features embedded in software companies already use. For a small or mid-size company, that is a practical route: start with the AI in tools you already pay for, pick one process, assign an owner, clean up the data, set security rules, and measure the result. This approach offers a faster start and a lower cost of entry, but it does not remove the two decisions that still separate companies making money on AI from the rest: who owns the process, and how you keep control of cost and output quality.