Enterprise Adoption: How Many Companies Pay, Who Gets Paid, How Deep It Goes

Ramp is monthly, the share of companies with a paid AI transaction, from a sample of 70,000. BTOS is fortnightly, the share of companies that say they use AI. The gap is a difference in definition: a survey misses free tools and employees paying out of pocket. After BTOS changed its question in November 2025 the figure jumped from about 10% to nearly 20%; the dashed stretch is read from press releases.

Left: every industry rises monotonically; none has turned back. From the second half of 2025 non-tech industries accelerated, manufacturing up 14 points in a year and finance from 60% to 74%, so diffusion moved from early adopters into the mainstream; even the slowest, hospitality, is at 30%, far from any ceiling. Right: mid-sized companies lead throughout; large companies lagged in 2023 and 2024 and caught up fast from the second half of 2025, meaning enterprise procurement cycles have completed, a sign of going mainstream.

The US Census Bureau's fortnightly survey, the share answering yes to "used AI in any business function in the past two weeks", from November 2025. Three groups by level, each on its own axis. Information is up 8 points in eight months and professional services up 6, leading the field; the low group, hospitality, transport and mining, has barely moved. Diffusion is the already-deep industries going deeper, not a broad roll-out, the same direction as the frontier pulling away in OpenAI's signals.

Left: the same moment and the same companies, only the question differs. The old question asked about AI in producing goods or services, which most companies felt did not apply, giving 3% to 10%; the new question asks about any business function, and hospitality went from 2.5% to 8.4%, transport from 3.1% to 7.2%, so back-office adoption is more common than it looked and the official figure understated for years. Right: current use against expected use in six months by industry, grey for now, green where expectation exceeds today, with the gap labelled, the thickness of each industry's adoption pipeline; expectation has run three to five points above current use throughout, so the pipeline has stayed full.

Left: the new-question period, the official fortnightly series from November 2025; current use is about 20% and expectation runs three to four points above it throughout. Right: the old-question period, a different wording so the level is only about 10%, but expectation likewise runs three to five points above current use. What the two periods share is a pipeline that has always been full: not one reading had expectation below current use.

The share of AI job postings is evidence that companies treat AI as a long-term capability, a stronger commitment than a subscription. Canada and Ireland went vertical from the second half of 2025, more than doubling in six months, reflecting the concentration of AI data centers and multinational tech expansion there; the US rose steadily from its 2023 low to about 5.8%. The large continental European economies clearly lag, and that gap is market depth for US AI vendors. Ramp's research finds hiring growth is higher at high-adoption companies; in the data so far, AI adoption and layoffs have not shown up together.

Left: the share of companies with a paid transaction at each vendor in the month; a company can use several. OpenAI held a steady 30% to 40% lead for over two years, then stalled from September 2025; Anthropic accelerated from single digits in 2024 and overtook it in May 2026. Google has been stuck at 5% to 6% for three years, xAI at about 3%, and DeepSeek has never passed 0.3%, so the US enterprise market never switched to it on price. Right: OpenAI still has slightly more of category spend, but 70% of new customers pick Anthropic and 60% of switchers move to Anthropic. The installed base favors OpenAI; the new business favors Anthropic.

Three tiers: the top 1%, the top 10% and the typical company. Dashed lines are estimated from Ramp's published chart; dots are figures Ramp disclosed in its articles. The typical company pays about one seat; the top 1% are building their own AI workflows, nearly 600 times more. A few companies betting heavily, not broad adoption. In August the top 1% fell 9.7% month on month for the first time, which Ramp attributes to seasonality, price cuts and switching to cheaper models; watch whether September recovers.

Three charts, three questions. Cumulative growth: the three tiers start from the same base and all rose three to four times in a little over a year. Year-on-year growth: the top 10% at about +245% is still the fastest, while the top 1% eased from +265% to +200% after its first monthly decline in August. Multiples: the top 1% divided by the top 10% has stayed between 10 and 13 times for three years, so the very top is not pulling further away; the top 10% divided by the median widened from about 30 to about 54 times, so the shift is in the second tier, heavy use spreading from a top-1% exception into a top-10% norm. All derived from estimated series; read the trend and the order of magnitude.

OpenAI's economic research team publishes usage intensity inside enterprise customers: the growth multiple of weekly active Codex users relative to February 1, 2026, on a log axis. February 2026 is the GPT-5.3 Codex launch month, so it lines up with Ramp's company-count share for the same period. Left: non-engineering functions such as legal, sales and recruiting show far higher multiples than engineering; right: the other job titles show the same shape. Depth surging while breadth stays flat is existing customers going deeper, and non-engineering staff using agents are building their own automation, not chatting. The multiples are indexed to a low base; OpenAI does not disclose absolute numbers.

Two lines at the same company part ways: the enterprise side above is deepening, the consumer side here is diluting. Left: message purpose, with work under 20% and other personal uses at two thirds. Middle: "doing" means asking ChatGPT to act rather than to inform; 45% of work messages are doing, only 22% of non-work messages. Right: the work share by plan rises with the paid tier, Pro 58% against free 26%.

Six snapshot releases published by Anthropic, split by platform rather than company, asking three things: how much people hand off (automation versus augmentation), how concentrated the tasks are, and whether use is for work or personal. It carries no inside-the-enterprise intensity, so it is not the same thing as OpenAI's enterprise signals; what can be compared is direction: tasks are spreading, the API side is hands-off (nearly 90% automation), the chat side is collaborative (about half). The June 2026 release changed its method, so later values are not fully comparable with earlier ones.

Left: paying companies as a share of the whole sample. Foundation models at 81.7% stand alone; the center of gravity of enterprise AI spending is still the model layer. The largest tool category, content creation, is a third of that, and agent-style applications (AI customer service 3.3%, AI sales 0.3%) are still very early. Right: year-on-year change. The model layer is all up; three of eight tool categories are down (AI coding, software design, AI customer service), the first quantified evidence of money concentrating in the model layer. Note that Claude Code and Codex subscriptions are booked under the model vendors, so part of the tool-layer decline is native model features absorbing standalone tools, not individual companies failing.

Ramp publishes only a monthly cross-section per category, but with one-month, three-month and one-year changes, so four official points can be backed out and indexed to a year ago = 100. The five declining categories are all steady slides, not single-month drops: AI coding 79, AI customer service 81, software design 83, AI sales 84, vector databases 92, the list of what the model layer absorbed. The rising group is the model layer and new entry points: open-source hosting 173, AEO 162, foundation models 120, GPU cloud 118. The flat group is content creation and video generation.