The enterprise field guide to AI

Rolling out AI looks like a software purchase. Buy seats, send an announcement, done. In practice, the companies that get every team using AI well make choices at every layer. Which models? Who can use what? Where does the data go? Where does the work live afterwards? Does the AI get better at your work, and who owns that improvement?

The picture below is the map for the rest of this guide. The lower layers are shared across the industry and getting cheaper fast. The upper layers are where your organization makes its own decisions, and where control is won or lost.

A large language model reads text, broken into small pieces called tokens, and predicts what comes next. Trained on enough text and code, that one skill turns into many: drafting an email, summarizing a call, reading a contract, writing a SQL query. Newer models also read images, audio and documents.

A handful of labs build the most capable models. Industry produced over 90% of notable AI models in 2025, according to Stanford’s AI Index. Some are closed: you reach them through the vendor’s service and never see the model itself. Others are open-weight: anyone can download and run them (Layer 4).

Two numbers describe a model in practice. The context window is how much it can read at once, from a few pages to several books. The price is what each token costs to process (Layer 3). The leading model changes every few months.

Pre-training is the expensive part. A model reads a vast amount of text and code and learns to predict the next token. That takes months on large clusters of chips and produces a model that knows a lot but does not follow instructions well.

Post-training turns that raw model into a useful assistant. Three steps do most of the work:

• Supervised fine-tuning. The model studies examples of good answers written by people.
• Preference tuning. People compare two answers and pick the better one. The model learns from those choices. This is often called reinforcement learning from human feedback.
• Reinforcement learning with checkable rewards. The model attempts tasks with a clear right answer, such as code that must pass tests or math that can be verified, and is rewarded when it succeeds. This is what produced today’s reasoning models, which work through a problem step by step before answering.

Post-training is moving fast. On OSWorld, a test of whether an AI agent can complete real tasks on a computer, the best score reached 66.3% in 2025, against a human baseline of 72.35%, Stanford reports.

Inference means running a trained model to produce an answer. Training happens once. Inference happens every time anyone uses the model, so it is the cost that grows with adoption. Providers charge per token, usually quoted per million tokens.

The price of a fixed level of capability has collapsed. Stanford’s AI Index found that GPT-3.5-level performance fell from $20.00 to $0.07 per million tokens between November 2022 and October 2024, a more than 280-fold drop. Epoch AI measured declines between 9× and 900× per year depending on the task, with a median of 50× per year. A newer Epoch study, which measures cost per completed task rather than per token, finds the cost of a given level of performance falling about 13× per year since 2023.

Two things push the other way. Reasoning models spend extra tokens thinking before they answer. That buys better results on hard problems at a higher cost per answer. And usage grows as people find more work for AI. A falling unit price does not mean a falling bill.

An open-weight model publishes its trained parameters. You can run it on your own servers or in your own cloud account, so prompts and data never leave infrastructure you control. You can also fine-tune it. In exchange you take on hosting, scaling, security patches and evaluation.

Open models are close to the frontier, but the gap moves. Stanford’s AI Index measured the top closed model’s lead over the top open model at 8.04% in early January 2024 and 1.70% in February 2025. Its 2026 edition reports the gap reopened: 3.3% in March 2026, up from 0.5% in August 2024.

Open models are widely used. On OpenRouter, a service that routes requests across hundreds of models, open-weight models reached about one-third of tokens by late 2025. Chinese open models averaged about 13% of weekly tokens over the year, peaking near 30% in some weeks.

Different models are good at different work, and the mix changes quickly. On OpenRouter, programming went from roughly 11% of tokens in early 2025 to more than half by late 2025. Reasoning models went from a negligible share to more than half over the same year. OpenRouter alone lists more than 300 active models from over 70 providers.

A model router sends each request to the model best suited to it, by task, cost, speed or data sensitivity. Put a router behind your company’s identity and policy controls and it becomes an AI gateway: one place where IT decides which models each team may use, what data may reach them, and what gets logged.

Shadow AI is employees using AI tools the company has not approved, usually personal accounts. It is common. In 2024, 78% of AI users said they brought their own AI tools to work, according to Microsoft and LinkedIn’s Work Trend Index. A 2025 Gartner survey of 302 cybersecurity leaders found 69% of organizations suspect or have evidence that employees use prohibited public generative AI.

It shows up in breaches. IBM’s 2026 Cost of a Data Breach study found that security incidents involving shadow AI more than doubled, to 43% from 20%. Among breached organizations with an AI-related breach, 92% lacked proper AI access controls.

Bans rarely hold. Samsung restricted generative AI on company devices in May 2023 after staff reportedly pasted source code and meeting notes into a public chatbot. What works better is giving people a sanctioned tool that is as easy as the personal one. Security vendor Netskope, measuring its own customers, saw the share of generative AI users on personal AI apps fall from 78% to 47% as company-managed accounts rose from 25% to 62%.

The controls are familiar from other software:

• Single sign-on (SSO) so people log in with their work identity, and SCIM so access is removed the day someone leaves.
• Role-based access (RBAC) so the People team’s apps can read HR systems and Sales apps cannot.
• Data rules such as PII redaction and limits on which data classes may reach which models.
• Named approvers for anything that leaves the company: an email to a customer, a message to a candidate, a change to the CRM.

Data residency is where data is stored and processed: which country, which cloud, whose servers. Data sovereignty goes further. It asks whose laws apply and who controls the data. For AI, it also covers what happens after a prompt is sent:

• Retention. How long the provider keeps prompts and outputs. Zero data retention means nothing is stored after the response.
• Training use. Whether your data may be used to train the provider’s models. Get the answer in the contract.
• Subprocessors. Which other companies touch the data on the way.
• Deployment. Whether you can run the platform in your own cloud account for regulated work.

Demand is growing quickly. Gartner forecasts worldwide sovereign cloud infrastructure spending of $80.4 billion in 2026, up from $59.3 billion in 2025, and expects it to shift 20% of current workloads from global to local cloud providers. Gartner also predicts that by 2027 more than 40% of AI-related data breaches will be caused by improper use of generative AI across borders.

Most AI work today disappears. A rep drafts a renewal email in a personal chat. A support lead builds a perfect prompt for refund cases and keeps it in a notes app. A recruiter summarizes forty candidates in a tab that closes. The prompt, the output, the edits and the decision are scattered across personal accounts. When that person leaves, the work leaves too.

The alternative is to treat AI work as a company record, like email or the CRM. A good record of AI work keeps:

• Prompts and context: what was asked, and which documents or systems the AI read.
• Outputs and versions: what came back, and how people edited it.
• Decisions and approvals: who accepted it, who approved sending it, and when.
• Provenance: which model and which policy were in force.

That record is searchable, so the next person can pick up the thread. It supports retention rules and legal hold. It can be exported, because it belongs to you.

Every day, people make deliberate judgments about AI output. A support lead accepts a reply as written. A sales manager rejects a follow-up email because the tone is wrong. A recruiter corrects a screening summary that missed a certification. Each of those is a labeled example of what good looks like at your company.

Kept in the record, those decisions can improve your AI over time, privately:

• Evaluations. Accepted and corrected outputs become a test set for your own work, so you can compare models on what matters to you.
• Better defaults. Prompts, templates and routing rules get tuned toward what your reviewers approve.
• Tuning, when it pays. Preference data from accept, reject and correct decisions is the same kind of signal labs use in post-training (Layer 2).

Two conditions matter. Learning should come from decisions people choose to make, not from watching how they work. And it must stay inside your organization. Your accepted work should improve your AI, never a model that serves anyone else.

Adoption is rising steadily. The Census Bureau’s biweekly business survey shows the share of U.S. firms using AI in any business function climbing from 17.3% in late November 2025 to 23.8% in early October 2026. Larger firms are well ahead of the national rate.

Using AI somewhere is not the same as using it well everywhere. In McKinsey’s 2026 survey, nearly nine in ten respondents report regular use of AI in at least one business function, but only 44% say AI is scaling across their enterprise, and 37% attribute any EBIT impact to it. In Microsoft’s 2025 Work Trend Index, 24% of leaders said their companies had deployed AI organization-wide, while 12% were still in pilot mode. Its 2026 edition found that only 26% of AI users say their leadership is clearly and consistently aligned on AI.

Many pilots never become company-wide practice. Gartner predicted in 2024 that at least 30% of generative AI projects would be abandoned after proof of concept by the end of 2025, and in 2025 that over 40% of agentic AI projects will be canceled by the end of 2027. These are forecasts, not measurements. The pattern behind them is familiar: a tool with no clear job, no owner, and no way to show results.

The companies that move from pilot to everyone tend to do four things:

• Start with specific work, not a blank chat box: the follow-up email after a sales call, a first reply to a support ticket, a screening summary for a recruiter.
• Start with one department that has a clear owner and a measurable output, then expand.
• Give leaders a report on work done and outcomes kept, not logins or keystrokes.
• Make the safe path the easy path, with approvals built in rather than added later.

The EU AI Act sorts AI uses into four tiers: unacceptable (banned), high risk, transparency obligations, and minimal risk. Amendments known as the Digital Omnibus entered into force on July 27, 2026 and moved the high-risk deadlines later. The dates below are those published by the European Commission after that change.

Every company using AI already has duties. Since February 2025, providers and deployers must take measures to support the AI literacy of their staff. The amended text says this does not require guaranteeing any specific level for each person. Fines for prohibited practices reach €35 million or 7% of worldwide annual turnover, whichever is higher.

In the United States, rules come from states and cities, and many focus on hiring and other consequential decisions.

Auditors and regulators ask the same questions: who used AI, for what decision, with which model and data, and who signed off. Hiring is a common focus. Illinois requires employers to notify employees when AI is used in employment decisions. New York City requires bias audits for automated hiring tools.

Spending is rising fast. Menlo Ventures estimates enterprises spent $37 billion on generative AI in 2025, up from $11.5 billion in 2024 (its earlier figures exclude inference spending). Zylo’s 2026 SaaS Management Index found spending on AI-native applications up 108% year over year. Without a plan, that spend spreads across personal subscriptions and team-by-team tools.

There are three broad ways to deploy AI across a company. None is right for everyone.

Seat structure matters for companies of 25 to 500 people. Here is what the major plans state on their own pages.

We built tabbr for the upper half of the stack, where your organization decides. We start with adoption, because AI nobody uses protects nothing. Each department gets a library of ready apps its teams can run on day one. We keep three kinds of control with you from day one: