Quick answer
About 1.3% of new San Francisco business registrations started in 2025 (101 of 7,953) have an AI marker in the registered name, up from 0.2% each year through 2022. That counts names only: no public file measures AI agents, and a name cannot show whether a business uses or sells AI.
Key takeaways
- The share of new San Francisco registrations with an AI marker in the name was flat at 0.2% through 2022, then rose each year: 0.6%, 0.9% and 1.3% in 2023, 2024 and 2025.
- The number of new registrations per year was about flat from 2022 to 2025, so the rise is in the share, not in more registrations overall.
- We found no evidence that these registrations are more concentrated in the five neighborhoods with the most registrations.
- The data cannot tell a one-time step from a smooth climb, and a name is a proxy: it misses AI firms with no AI in the name and can catch names that only contain the letters.
- 2026 is a partial year with a changed registration form, so it is reported apart and is not part of the trend.
How many new San Francisco registrations have AI in the name?
About 1.3% of new San Francisco business registrations started in 2025 have an AI marker in the registered name, up from 0.2% each year through 2022. What we measured is new business registrations in San Francisco with an AI marker in the registered name. No public dataset measures AI agents or AI companies as such, so this counts names only. The share was flat through 2022 and then rose each year to 2025.
The table counts one registration per business account, dated by its earliest San Francisco location, from the city's public register. An "AI marker" is a standalone AI, .ai, A.I. or the phrase "artificial intelligence" in the name. The last column is a 95% range: if the same count were repeated under the same conditions, a range built this way would contain the true share about 95 times in 100.
| Start year | With an AI marker in the name | All new registrations | Share | 95% range |
|---|---|---|---|---|
| 2017 | 18 | 10,185 | 0.2% | 0.1% to 0.3% |
| 2018 | 17 | 10,053 | 0.2% | 0.1% to 0.3% |
| 2019 | 17 | 9,112 | 0.2% | 0.1% to 0.3% |
| 2020 | 9 | 7,194 | 0.1% | 0.1% to 0.2% |
| 2021 | 12 | 7,218 | 0.2% | 0.1% to 0.3% |
| 2022 | 14 | 7,687 | 0.2% | 0.1% to 0.3% |
| 2023 | 49 | 7,628 | 0.6% | 0.5% to 0.9% |
| 2024 | 66 | 7,773 | 0.9% | 0.7% to 1.1% |
| 2025 | 101 | 7,953 | 1.3% | 1.1% to 1.5% |
Across 2017 to 2025 there were 74,803 registrations in total, and 303 registrations had an AI marker in the name. The number of new registrations per year was about flat from 2022 to 2025. So the rise is in the share, not in more registrations overall. The technical paragraph gives the same comparison restricted to registrations with a business activity code on file.
What did we measure, and how?
We counted registrations whose registered name contains an AI marker, as a share of all new registrations, by the year the business started in San Francisco. The source is the Registered Business Locations dataset from the Office of the Treasurer & Tax Collector, published on DataSF under a public-domain licence (PDDL). We downloaded it on 2026-10-06 and analysed start years 2017 to 2025. The name is the one on file now, and the register stores names with normalised capitals, so a person's given name Ai looks the same as AI the technology. Sole proprietors can register under their own name, so such names can match.
This series set out to look at AI agents. A name cannot identify agent businesses: in this register the word "agent" in a business name mostly belongs to insurance and real estate businesses, so we did not count it. We also started from an assumption that AI-related registrations would be a tech-only story; the section on activity codes below says what the data showed. The direction of the rise was visible from earlier counts before we froze the plan. The yearly shares and totals were not.
The headline is the share of all new registrations. Many carry no activity code: 77% of 2017 registrations have none, against 39% of 2025 registrations. The model we fitted treats "no code" as its own category. We read nothing into that change.
Technical detail (skippable): the 2025-versus-2022 odds ratio, a ratio of odds and not "times as likely", is odds ratio 6.43 (95% CI 3.97 to 10.41, p < 0.001) with business-code status adjusted. Without that adjustment it is odds ratio 7.05 (95% CI 4.48 to 11.10), and among registrations that report a code it is odds ratio 4.98 (95% CI 2.38 to 10.42). A test of whether the yearly rise gets steeper gives odds ratio 1.05 on the squared-year term (95% CI 1.02 to 1.07, p < 0.001); a flat period followed by a rise also produces that, so it cannot separate a step from smooth acceleration and is not proof of it. A third test, for neighborhoods, gave odds ratio 1.74 (95% CI 0.80 to 3.80, p = 0.163). All three were corrected together; the first two hold after correction and the third does not. Neighborhoods were treated as clusters, effective n = 8,548 (41 neighborhoods) for registrations with a location, and the intervals on the odds ratios account for that. The yearly share intervals in the table are plain Wilson intervals and do not.
Is the rise a step or a smooth climb?
The data cannot say which. Through 2022 the share sits at 0.2% with 0.1% in 2020, and then it is higher in each later year. With nine yearly points, a single jump and a gradually steepening climb cannot be told apart, so we give no date for when the rise began and no explanation for it.
Are these registrations concentrated in a few neighborhoods?
We found no evidence that they are more concentrated in the five neighborhoods with the most registrations. Descriptively, 53% of registrations from 2022 to 2025 with an AI marker (120 of 228) sit in those five areas, against 36% of all registrations (10,790 of 30,131). The test in the technical paragraph above does not separate that gap from no difference once neighborhoods are treated as clusters. In words, most of the registrations with an AI marker in those five areas are in one neighborhood. We publish no neighborhood table: small counts are suppressed.
Are they all tech businesses?
The data neither confirm nor contradict that. We expected the registrations to be mostly tech. Among registrations from 2022 to 2025 with a business activity code on file, 29% of those with an AI marker (95% CI 22% to 37%; 37 of 129) have a code on file outside the information and computer-systems-design groups (NAICS 51 and 5415). A further 44% have no code on file at all, so that 29% is not all of them. Our pre-set rule gives "indeterminate".
This describes the activity code on file, which the dataset documentation calls self-reported and not verified by the city's tax office. It says nothing about where AI is used. For the Census survey of AI use by industry, see Is AI Only Used by Tech Companies?.
How reliable is an AI marker in a name?
For the names that matched, one reading check found 93 of 100 (93%) clearly refer to AI as a technology. This is a reading check, not a validation. One rater read 100 randomly drawn matched names, and that rater was an AI model that could see the AI marker and is not independent of the rule. Seven names were unclear and were counted as not AI.
The check says nothing about the other direction. Recall cannot be measured here. Well-known AI firms are missed whenever their names carry no AI marker.
What was 2026 like?
2026 is a partial year and is reported apart, not as part of the trend. Through 2026-10-06 the share was 0.9% (57 of 6,226). The city changed its registration and tax process on 2026-07-06 (Proposition M, per the publisher's dataset note), and the share with no activity code fell from 78% to 1.3% at that date. So 2026 and earlier years are not like for like.
What can this data not show?
It cannot show how many AI companies exist, whether any operate, or where AI is used. The limits, in full:
- A registration is not an operating business or an employer.
- Businesses that never registered, are exempt or are outside San Francisco are absent.
- AI businesses with no AI marker in the name are absent; names that only contain the marker are included.
- Closed businesses stay in the register, and a start date can come before the registration.
- Registrations of one owner count separately.
- About 1,800 rows with a misspelled city were left out under the protocol.
- The address is the registered one.
- The reading check had one rater.
- Everything here is an association between a start year, a name and a registration; none of it can say why a change happened.
How can you check it?
Open the public register and re-count. The data are on the DataSF dataset page, and the dataset's documentation and metadata give the licence and the 2026 note. The CSV endpoint returns the rows. Anyone can re-count with the matching rule described above (a standalone AI or A.I., a .ai ending, or the phrase "artificial intelligence"). We publish no name list, no certificate numbers and no data file. For the neighboring question of how many businesses of different sizes use AI, see AI use by business size and industry.
What can you do with this?
Use three reading rules the data support. First, when someone says AI businesses are multiplying in San Francisco, ask what was counted: here it was names in a register. Second, compare shares, not raw counts: total registrations per year were about flat from 2022 to 2025, and the share is what moved. Third, treat any single year with care, and treat a partial year or a year with a changed form, like 2026 here, as separate.
A registration with AI in the name shows what someone chose to file. It does not show a working product, and its absence does not show the opposite.
Analysis by Deivy Hernandez, from public data files.
FAQ
Will the share keep rising in 2026 and after?
The data cannot say. They end with full-year 2025, and the 2026 figure is for part of a year and was collected under a changed registration form, so it is not a like-for-like next point. A projection would need a model this study did not fit.
Can I use this to find AI competitors near me?
No. The study counts registrations by name and publishes no names, and a registration does not show an operating company, a product or a location where customers can find it. A name search in the city's own register shows only what a business chose to file.
Is this the same as the share of businesses that use AI?
No. This counts what registrants put in a name. The Census Bureau's survey asks employer businesses whether they used AI; our look at AI use by metro area covers that survey.
Will a re-count today give the same numbers?
Not exactly. The city says the dataset is updated daily, and this study used the extract of 2026-10-06. Start dates and names on file can change, so a later count of the same years can differ slightly.