LOThe question
Where AI actually lands in a company, and who holds it
Adoption is narrower than the coverage suggests and lands in the same three places. The interesting question is not which function buys the tool — it is which role inherits the work the tool creates, because in the verified record that role is almost never the one that chose it.
Three functions, and rarely more than three at once
Among US firms that use AI, 57% use it in three business functions or fewer. The most common are sales and marketing (52%), strategy and business development (45%), and IT (41%). Overall 18% of firms use AI in any business function, 32% weighted by employment, expected to reach 22% within six months. For very large firms in information, professional services and finance the rate reaches 50 to 60%.
If your company has AI in more than three functions, you are at the front of this distribution, not behind it. That is the opposite of the impression most coverage gives, and it is worth knowing before you buy anything on the grounds of falling behind.
A second, independent reading points the same way. The Gen AI Jobs Transformation Map that Singapore's central bank commissioned with the sector's skills bodies expects roughly 80% of the country's financial-sector value to come from four business functions — sales and marketing, customer operations, risk management, and engineering and technology. Different country, different method, different decade of regulation, and sales and marketing is at the top of both lists.
What gets replaced first is other software
In the same survey, 16% of AI-using firms replaced existing software or equipment with an AI-integrated version — roughly twice the share reporting any AI-related employment change in either direction. Ninety-five per cent report no headcount change at all.
That lands on a specific person. The process you configured, the rules you wrote and the boundary you set around what the system may do without a human all move to a platform that ships its own versions of them. The verified record does not say anybody lost a job; it says the object of somebody's job was swapped underneath them.
Some of it is already inside without a decision
The same survey finds 23% of firms report workers using AI in work-related tasks, against 18% adopting it at firm level. Diffusion runs both ways: worker use sometimes happens without formal adoption, and formal adoption sometimes happens without worker use.
Vendor telemetry sees the bottom-up half directly, because only the party operating the tool can. Anthropic's usage data shows the median agentic session producing a blog post contains a single human prompt, against thirteen exchanges in a chat conversation — and the gap appears across 26 of 31 output types and survives holding the model constant. Whoever is producing that output inside your company did not raise a purchase order for it.
The duties arrive assigned to nobody
Since August 2026 the EU AI Act requires deployers of high-risk systems to assign human oversight to natural persons with competence, training and authority. A survey of 200 SRE and DevOps leaders reported that 43% of AI-generated code changes still need human review. IBM reported its AskHR agent containing 94% of common HR questions across more than 80 applications — which moves what is left to people, and what is left is what was already hard.
The recurring shape across these records: capability arrives, a duty arrives with it, and the duty is inherited by whoever already understands both the system and the business consequence. That person usually has the knowledge and not the authority — and the record shows what happens next, because the review step is the first thing cut when it produces no artefact.
- Since 2 August 2026 the EU AI Act requires deployers of high-risk AI systems to assign human oversight to natural persons with the necessary competence, training and authority (Article 26)
- A survey of 200 SRE and DevOps leaders reported 43% of AI-generated code changes still need manual debugging in production after passing QA and staging
- IBM reported its AskHR agent containing 94% of common HR questions across more than 80 automated HR tasks, with support tickets down 75% since 2016 and HR operating costs down 40% over four years
- Deciding what it may do unsupervised
The part that argues against us
We sell an assessment of what AI changes for the roles a company employs. The strongest single finding in this note argues that for most companies the answer so far is: not the headcount. Ninety-five per cent of AI-using firms report no employment change. If you came here expecting a number for how many people you will not need, the verified record does not contain one, and we would rather say so than sell you the number somebody else will.
What the record does contain is narrower and more actionable: which functions adopt first, which duties arrive with the adoption, which roles inherit them without authority, and which of your roles the public evidence cannot see at all. That last one is the honest limit of any assessment, ours included, and it is why the empty cell is left visible.
- Census BTOS supplement: 16% of AI-using US firms replaced existing software or equipment with AI-integrated solutions
- C.H. Robinson's Q2 2026 10-Q shows average headcount down from 12,858 to 11,471 while revenue rose 19%, naming automation and AI to cut manual processes as the restructuring initiative
What this does not establish
The adoption figures are US firms only, for a reference period of November 2025 to January 2026, and count firms rather than workplaces or roles. A firm answering yes has AI somewhere in a business function; nobody asked which roles that function contains or whose tasks changed. The duty records are rules in force, not enforcement. And the per-role readings here come from occupation pages whose judgements are labelled task by task — most of them are inferred, and the page says so rather than this one.
Everything cited here
Each one opens the full record: its source and source tier, the dates, the scope it applies to, who verified it and what it does not establish.
- Census BTOS supplement: 18% of US firms used AI in a business function, and 57% of those in three functions or fewer — U.S. Census Bureau, Center for Economic Studies (CES-WP-26-25)
- Singapore's financial regulator and the sector's skills bodies forecast that Gen AI will augment most financial-sector roles over five years rather than displace them — Generative AI Jobs Transformation Map for Singapore's financial sector (MAS, IBF and Workforce Singapore)
- Census BTOS supplement: 16% of AI-using US firms replaced existing software or equipment with AI-integrated solutions — U.S. Census Bureau, Center for Economic Studies (CES-WP-26-25)
- Deciding what the system may decide alone — task on Business systems owner
- Anthropic's usage telemetry: the median agentic session that produces a blog post contains one human prompt, against 13 turns in chat — Anthropic Economic Index report: Cadences
- Since 2 August 2026 the EU AI Act requires deployers of high-risk AI systems to assign human oversight to natural persons with the necessary competence, training and authority (Article 26) — Regulation (EU) 2024/1689 (Artificial Intelligence Act), Article 26 — verbatim mirror of the Official Journal text
- A survey of 200 SRE and DevOps leaders reported 43% of AI-generated code changes still need manual debugging in production after passing QA and staging — VentureBeat
- IBM reported its AskHR agent containing 94% of common HR questions across more than 80 automated HR tasks, with support tickets down 75% since 2016 and HR operating costs down 40% over four years — IBM — AskHR case study (first-party)
- Deciding what it may do unsupervised — task on AI implementation lead
- C.H. Robinson's Q2 2026 10-Q shows average headcount down from 12,858 to 11,471 while revenue rose 19%, naming automation and AI to cut manual processes as the restructuring initiative — C.H. Robinson Worldwide, Inc. — Form 10-Q for the quarter ended 2026-06-30 (SEC EDGAR)