Case studies

Every entry is a real, publicly documented AI implementation — a company deploying AI for a specific purpose, with reported outcomes, gathered by monitoring the open web. A use case's case study count is how many of these have been published about it in the last 24 months.

It is a proxy for real-world adoption — but a proxy, not a census: it reflects what companies and vendors have chosen to publish, not everything happening in the market. High volume with low proven impact indicates hype; high proven impact with low volume is anecdotal.

Growth (year over year)

Growth compares the case studies published in the last 12 months against the 12 months before that — a year-over-year read on whether a use case is gaining or losing momentum. It is always this 12-vs-prior-12 comparison, regardless of any date filter you apply elsewhere.

Treat it as a directional signal, not a precise figure: a positive number (shown in green) means more case studies were published than the year before; a negative number means fewer. When the year before had fewer than 5, we show the two counts instead of a percentage: a jump from 2 to 8 is “2 → 8”, not “+300%”.

Fastest growing ranks by momentum: how much faster a use case grew than all case studies in the same view, which grew too (partly because we track more each year), counting a small base for less. Within one industry, role or company size the counts are small, so read its order as a direction rather than a verdict.

No single vendor decides it. Most case studies are published by vendors about their own customers, and some vendors publish far more than others. So in each year, a vendor whose customers make up more than a third of a use case's case studies counts as a third, and growth reflects the use case rather than one vendor's publishing. The counts shown stay real counts.

Proven impact

Proven impact shows how strong the results are that companies have proven for each use case. It is a composite score where financial results (ROI, revenue and cost savings) are weighted stronger than time savings and operational metrics.

Only results a company measured and attributes to the AI count. Projections, rollout milestones (how long it took to build or deploy, how many staff use it), savings that went to a firm's customers, and results a case study doesn't attribute to the AI are left out.

It is a per-use-case average, not a per-company figure, and a use case needs at least 5 case studies to earn one. Each use case's average counts in 20 typical case studies alongside its own, so a handful of case studies can't top or sink a ranking: with 7 of its own, the typical ones mostly decide; with 100, its own do.

In an industry, role or company size view, a use case appears once it has at least 5 case studies there, and its score combines them with its record across every industry, adjusted to how case studies in that view compare overall. The more it has there, the more they count. We measured that a use case's results differ much less from one industry to another than from one use case to another, so a handful of case studies can't decide its place.

No single vendor decides a score either. Where one vendor's customers make up more than a third of a use case's case studies in a year, they count together as a third. In October 2026 that applied to a handful of use cases where one vendor's customers made up half or more, such as Heidi in clinical documentation and Harvey in legal research. Below the line nothing changes.

The word beside a score says how far it sits from the average use case, so the same number always reads the same:

Under the hood, each type of proven result carries a weight, and the rare, hard financial outcomes deliberately count for far more than softer ones:

Each type is worth double the one below, so one stronger result outweighs every weaker type together, and a single revenue or EBIT/ROI case punches far above its count. That is intentional: a hard financial outcome is the most concrete business result there is, and the rarest to find documented.

These are results companies chose to publish, which lean toward successes. The score compares use cases on the same footing, not what any one company should expect.

Build vs buy

How companies implement each use case: they bought a ready-made AI product and configured it, or built their own, writing code that shapes the AI, often on a vendor's model or cloud. Each case study is classified from its own story by one question: did the company, or a partner it hired, write that code? Configuring a vendor product, however extensively, counts as buying. The percentages cover classified case studies only (nearly all of them), and a vendor's showcases of its own AI are left out, as everywhere on the site.

Partner-built is the share of builds a third-party integrator wrote rather than the company's own team: about a fifth. In-house or with a partner is a separate question, for builds and buys alike: whether an outside firm helped, from advising to building. That is a consultancy or systems integrator named in the case study or the firm that published it, and any outside firm that wrote the code, a vendor's own engineers included.

We report it as the buy share — so “74% buy” means 74% bought or configured a vendor product and the remaining 26% built custom.

Above the bar, the answer is said in companies: most buy it ready-madewhen two thirds or more bought, most build their own when two thirds or more built, and split otherwise. Products they bought lists the vendors of the companies that bought: the model or cloud a build runs on is no product choice, so it isn't listed. A product is named in the answer only when more than half of the buyers chose it, and a platform only when more than half of the builders built on it. These are counts of case studies, never market share: vendors publish their own customers' stories at very different rates.

Agentic level

Agentic level captures how autonomous the deployed AI is, from L1 (a tool people operate) to L5 (it runs unattended, with no one in the loop). We place each case study at one level from what it describes the AI doing: whether it only produces an output, answers from real information, or acts in company systems; who stays in the loop; and what the results are credited to. Branding doesn't count: a deployment called an “agent” with no actions described stays at L2 at most. L5 is about autonomy, not about several agents working together.

A use case's level is the one half or more of its case studies sit at, shown with that share (“L4 Expert, 79% of its case studies”). When no level holds half, it reads Mixed and names its two most common levels: today 11 of 35 use cases. We never round an average into a level, because an average of 3.0 can come from case studies split between L2 and L4 with almost none at L3.

How sure it is. Placing a case study is a judgment, made from its text with a strict rubric by an AI model. When we scored 150 case studies a second time, 31% came back at a different level, mostly between L2, L3 and L4. So read one case study's level as approximate; a use case's level rests on many case studies and is the signal to use. A level that contradicts the rubric's own tests (an L4 that only reads data, say) is scored again.

No single vendor decides it. As with proven impact and growth, in each year a vendor whose customers make up more than a third of a use case's case studies counts as a third in its shares. The counts shown stay real counts.

The scale, level by level:

KPIs tracked

We cluster every metric reported across a use case's case studies into distinct KPIs — so a use case that tracks “meetings booked,” “conversion rate,” and “pipeline generated” counts three. The KPIs tracked figure is how many distinct KPIs peers report for that use case — a measure of how broadly its results are quantified.

Only KPIs reported by at least three companies are counted, so it reflects genuinely shared measures rather than one-off metrics.

When more than half of the companies behind a KPI's figure use one vendor, its breakdown says so (“All 14 firms use Harvey”): the figure then mostly reflects that vendor's customers. We say it rather than reweight, so every figure stays a plain median you can check row by row.

Industry, team and size

Every case study gets an industry, a team and a company size, which the views by industry, team and size filter on.

Industry is the industry its AI is used in: usually the company's own, from its profile (95% of case studies). When a company applies AI to another industry's work, such as a software firm building for hospitals or a consultancy working for a bank, it is that industry. There are 22 industries and 126 sub-industries. Technology covers IT services, hardware, cloud, semiconductors and cybersecurity; software companies are under Software & Platforms.

Team is the business function whose work the AI does, classified from each case study. Company size comes from the company's number of employees: SMB up to 250, mid-size up to 5,000, enterprise above.