AI Is Already Inside Your Organization

Picture an ordinary Monday in an organization much like yours. At six in the morning the email filter quarantines a phishing message before anyone is awake. At half past seven the planning system refreshes the demand forecast that will drive this week’s orders. At a quarter past eight the recruiting platform ranks the people who applied overnight. At twenty to nine a manager drafts a client proposal in a personal AI account, because it is quicker than the approved route. At nine the leadership team meets and agrees that it is time to start with AI.

The morning is invented, but nothing in it is unusual. The point is the order of events. In many organizations the AI program began years before the meeting that announced it. Nobody called it a program, because nobody had to approve it as one.

Three ways AI gets in

The argument of this chapter fits in one sentence. Your AI transformation may have started before your organization called it one, so the first leadership question is not what to buy but what is already running, and who answers for it.

AI reaches an organization through three routes, and it helps to name them, because each one hides in a different place.

Embedded AI arrived inside software the organization bought for another purpose: the planning system that forecasts, the security filter that scores messages, the routing engine that plans deliveries, the applicant-tracking system that ranks candidates. The team that bought the software knows it exists. Whether anyone owns the AI inside it is a separate question, and often the answer is no.

Official AI is what the organization licensed as AI: an enterprise assistant, a coding helper, a platform with a contract, an owner and a usage policy. This is the bucket leadership knows best, because it passed through procurement and usually through a board paper.

Informal AI is what employees brought themselves: personal accounts, browser extensions, a subscription paid on a departmental card. As The Speed of AI Adoption showed, this bucket is large. Among employees who use AI at work, about 70 percent rely on free public tools, against 42 percent who use tools their employer provides¹.

Executive conversations tend to be about the official bucket. The other two are where surprises live.

Much of the AI at work never announces itself

Much of the AI inside an organization does not look like a conversation. It looks like a result. A shopper sees “you may also like” and never sees the model behind it. A cardholder sees a fraud alert, not the score that triggered it. A driver sees a route, not the prediction of traffic that produced it. A recruiter sees a ranked list of applicants and does not always see why one name is above another.

This quiet layer is not new. Ranking, scoring, forecasting and routing have been built into enterprise software for years, usually without the label. Mature technology disappears into the products that use it, which is a blind spot. A leadership team whose picture of AI contains only assistants is looking at the newest and most visible part, and missing much of the AI that already makes decisions.

AI now arrives by software update

The embedded bucket is growing for a simple reason: software vendors are adding AI to the products their customers already use. In 2024 Gartner told finance leaders it expected about 80 percent of independent software vendors to have embedded generative AI in their enterprise applications by 2026, up from less than 1 percent in 2023². That is a forecast for this year, not a count, but it shows where vendors said they were heading.

When AI arrives this way there is no purchase order, no business case and often no announcement beyond release notes. The organization does not decide to adopt it; it finds that it has.

Customers of the workplace messaging service Slack found this out in May 2024. A widely shared post pointed out that Slack’s privacy principles said its systems analyzed customer data, including messages, content and files, to develop machine-learning models for features such as channel and emoji recommendations and search results. The use was on by default, and a workspace owner who wanted out had to email Slack to ask³. Slack responded that these were not generative models, that customer data was not used to train large language models, and it later explained that the models rely on derived signals rather than reading message content⁴.

The lesson of the episode is not that the vendor behaved badly. It is that many organizations had been running AI features in a tool they used every day without knowing it, and learned the terms from social media rather than from their own records. The question for any buyer of software is whether it finds out from its own review or from someone else’s post.

Even a mandated count keeps growing

If any organization should know where its AI is, it is one that is legally required to list it. Since an executive order of December 2020, US federal agencies have had to publish inventories of their AI use cases.

In December 2023 the Government Accountability Office reviewed the inventories agencies had published. Twenty of 23 large agencies reported about 1,200 current and planned uses of AI, and three reported none. Only five of the twenty provided complete information; the other fifteen had incomplete or inaccurate entries⁵. The 2024 consolidated inventory listed 2,133 use cases from 41 agency submissions⁶. The 2025 inventory, published in 2026, listed 3,611 individually reported use cases at all stages of development. OMB counts submissions, not agencies: 41 submissions reported individual use cases, out of 56 in all, the rest filing only common commercial tools or confirming no AI use. It also added a separate, consolidated category for common tasks that rely on commercial off-the-shelf AI products, the embedded bucket by another name⁷.

Some of that growth is genuinely new use. Some is better counting under clearer rules, and the rules changed between rounds, so the years are not strictly comparable. Both readings carry the same message for a private organization with no legal duty to look. When people start counting seriously, the number goes up, and the first list is rarely the full one. The US National Institute of Standards and Technology makes the same point in its voluntary AI Risk Management Framework: one of its basic governance expectations is that mechanisms exist to inventory AI systems⁸. How to build that inventory is the work of AI Inventory and Risk Classification, in Module 07.

Unseen AI creates three problems at once

AI that leadership cannot see does not stop working. It keeps forecasting, ranking and drafting, and it leaves three problems behind.

The first is no map. Different functions use different tools, and nobody holds the full list. When a board member asks where the organization uses AI, the honest answer is that nobody knows, and the federal experience suggests that even a careful first answer will be low.

The second is duplicated spend. Two departments buy two tools to summarize the same kind of document. Nobody compares their quality, and neither team learns from the other. The waste is small in any one case, large in total and invisible until someone looks across functions.

The third is uneven rules. Privacy, security, accuracy and accountability are handled differently in every corner. The official assistant has a policy; the personal account has none. The old fraud model has an owner and a review cycle; the ranking feature switched on in last quarter’s software update may have neither.

None of these is solved by buying another platform. A fourth tool does not fix an unseen third, and a strategy that ignores the systems already in production is a wish list.

Bought is not the same as owned

The most important consequence of embedded AI concerns accountability. Buying a system from a vendor does not move the responsibility for what it decides to the vendor. A forecast that sets staffing levels is a people-and-money decision. A model that ranks applicants is a fairness-and-law decision. Neither has to call itself AI to matter.

The clearest current illustration is a lawsuit, Mobley v. Workday, in federal court in California. A job applicant alleged that the AI screening tools in a widely used hiring platform rejected him for more than 100 jobs because of his age, race and disability. The vendor argued it was not the employer and made no hiring decisions. In July 2024 the judge allowed the case to proceed on the theory that the employers had delegated to the vendor’s tools their traditional function of rejecting candidates or advancing them to interview⁹. In May 2025 the court allowed applicants aged 40 and over to join as a nationwide collective¹⁰.

In June 2026 the court let most of the claims in an amended complaint go forward, and as of October 2026 the case was still in pretrial proceedings, with no finding of liability¹¹. The claims are unproven, but the lesson does not depend on the outcome: the employers bought a hiring system, and AI inside it was shaping who got an interview. How many could have named the owner of that screening step?

A practical test is to ask, of any system that shapes customers, money or people: if it made a bad call on Friday, whose name would come up in Monday’s meeting? If the honest answer is nobody, the vendor still does not own the outcome. You do. Who that owner should be, and how accountability is divided, is the subject of AI Roles, Ownership and Accountability in Module 07.

The law increasingly says the same thing, and it addresses the organization that uses the system, not only the one that built it.

Seeing takes curiosity, not blame

Whether leaders get an honest picture depends on how they ask. There are two ways to get it wrong.

One is to look away and assume that, because nothing was approved, nothing is running. The other is to announce a crackdown and go looking for culprits. As The Speed of AI Adoption showed, a ban stops the organization’s learning without stopping the use; it mainly moves the use out of sight. People who fear confiscation name only the approved systems, and leadership governs the official stack while the real one keeps running.

The stance that works sits between the two. Assume some informal use exists. Ask what problem people were trying to solve, because that is real demand the organization has not yet met. Ask the owners of every major business system what their software now decides on its own, because they may not have asked their vendors. Treat what comes back as information rather than confession. Leaders who handle the first disclosures calmly tend to get a second round that is more complete.

Story: a sports league that thought it was starting

The following is an illustrative composite, not a real organization. Its parts are common ones.

A national professional sports league runs the competition for its member clubs: the fixture schedule, central ticketing for its biggest matches, the broadcast and sponsorship deals, and the integrity of the game. Its board asks for an AI strategy, and the leadership team does what many teams do. It books vendor demonstrations, drafts a shortlist and writes a steering paper with the word transformation on the cover.

Meanwhile, AI is already at work. The scheduling software has used its vendor’s optimization engine to build the fixture list for years. The ticketing platform’s demand-pricing feature, switched on in a vendor update, now sets the price of every section of the stadium. An integrity service scores betting patterns for signs of match fixing. Office staff draft documents with a licensed assistant. The commercial team pastes draft sponsorship terms into personal AI accounts to write proposals faster. Two departments pay separately for two meeting-summary tools.

Then, before a long holiday weekend, the demand-pricing feature pushes the price of family seats for the big holiday match far above what the league had promised, and fans complain loudly and in public. Nobody can say who owns the pricing rule or who is allowed to change it. A month later the league learns that confidential sponsorship terms have been sitting in personal accounts that no one had listed. The league thought it was starting. It was already in the middle.

A few months later the chief operating officer pauses the vendor tour and changes the first move. Leadership tells every function, in writing, that nobody will be penalized for tools they disclose, and asks one plain question: what already ranks, scores, forecasts, routes, prices or drafts in your work?

The answers fall into the three buckets. Embedded: the scheduling engine, the demand-pricing feature and the integrity scores. Official: the licensed office assistant. Informal: the personal accounts in the commercial team and the two summary tools. The pricing feature gets a named owner with authority to set limits before the next big match. The commercial team gets an approved assistant and a clear rule that confidential deal terms never go into personal accounts. The league keeps the better summary tool and stops paying for the other.

Only then does it return to the vendors, and the conversation is much shorter, because it knows which problem is still unsolved. Nothing about the league’s ambition changed. What changed was the order of work: see, then choose.

What this means for leaders

The practical consequence is a change in the first question. Before asking which AI to buy, ask where AI already works and who answers for it. That question is cheaper and more revealing than a vendor tour, and it improves the eventual purchase, because it shows which problems are already solved, which are solved twice and which are not solved at all.

It also changes what leaders should expect of the people who own business systems. The owner of the planning system, the hiring platform or the customer database now owns whatever AI those systems contain, whether or not they chose it. Many have not been told. Telling them is a leadership act, not an IT task.

The first round of looking will be incomplete. How leadership reacts to it decides whether there is a second.

Check yourself

1. If nobody in the organization uses a chat assistant, the organization is not using AI.MythFact
2. New AI features often reach organizations through updates to software they already pay for.MythFact
3. When an organization is required to list its AI, the first list is usually complete.MythFact
4. If a vendor supplies an AI hiring tool, the duties around its use fall only on the vendor.MythFact
5. Banning unofficial AI tools gives leadership an accurate picture of AI use.MythFact
6. AI inside software the organization bought still needs a named internal owner.MythFact

Reflection: find the quiet systems

What comes next

AI may already be everywhere in your organization, including in systems that never say its name. That raises a fair question. If AI has been working quietly for years, why does this moment feel so different? Something changed about the part people can see. The next chapter, Generative AI Changes the Game, explains what changed and what did not.