Report

# How accurate are Ireland’s NACE codes?

Irish company records as of 25 June 2026 · 347,902 companies · Classification audit, September 2026 · Last updated 16 November 2026

Around three in five of the 347,902 Irish companies in the data have a NACE code that is missing, wrong or impossible to check. Take any 100 companies in the data: 39 have no NACE code at all, 6 are holding companies with no trading activity to check against, 14 have a code that is wrong, and 41 have one that holds up. The 39 and the 6 are counted across all the records. The split of the rest into 14 wrong and 41 right is estimated from a model’s judgements of a sample, nearly a third of which it could not judge, and from one person’s checks of 100 of the disagreements.

NACE is the European Union’s classification of economic activities. A company registering in Ireland today gives one NACE code for its main activity. We took a commercial dataset of Irish company records as it stood on 25 June 2026: 347,902 companies, each with whatever NACE and SIC codes are recorded for it. Below, “the register” means these records, not the CRO’s own register. We counted what those records say. Then we sampled 2,400 companies and withheld each one’s registered classification codes from the evidence. A language model worked out what each business does, where the evidence allowed. Three people classified 419 of the same companies, so we could check the model against them.

## Key findings

Counted figures cover all 347,902 companies in the records. Estimated figures come from the sample of 2,400. How each estimate was measured is under [“How we measured it”](https://www.sunstone.ie/resources/guides/irish-nace-code-accuracy/#method).

* **Around three in five** of the 347,902 Irish companies in the records have a NACE code that is missing, wrong or impossible to check. An estimated 41% (95% confidence interval 38% to 43%) carry one that is present and correct. That estimate rests on a model’s judgements of a sample and one person’s checks of 100 of the disagreements, and sits outside the pre-registered gates. (Estimated)
* **39.1%** of the 347,902 companies (136,101) have no NACE code. (Counted)
* **9.0%** (31,412) have no code at all, in either NACE or SIC. (Counted)
* **164 of 93,068** companies formed before 2010 (0.2%) carry a NACE code. (Counted)
* **52%** of companies with both a NACE and a SIC code (31,223 of 60,026) have codes that map to different sections; 39.5% still do once SIC 74990 (non-trading, not an activity) is removed and holding companies (NACE 6420) are set aside as a sensitivity check. (Counted)
* **25.2%** of companies with a NACE code, leaving out holding companies, have a NACE code that is wrong (95% interval 20.8% to 29.6%). This is a post-hoc estimate resting on one adjudicator’s 63 settled cases. (Estimated)

**Chart: Every 100 companies in the data** (Counted over all 347,902 records, except the wrong and right split, which is estimated)

* No NACE code: **39.1% (136,101)**
* Holding company: **6.4% (22,414)**
* Wrong code: **13.7% (about 47,787, ±2.4 points)**
* Right code: **40.7% (about 141,600)**

“No NACE code” and “Holding company” are counted exactly over all 347,902 records. Only the split between wrong and right code is estimated, from a model’s judgements and one adjudicator’s 63 settled cases (see [“Adjudication”](https://www.sunstone.ie/resources/guides/irish-nace-code-accuracy/#adjudication)).

## Do the records agree with themselves?

No, and this check needs no model and no sample. 60,026 companies in the records carry both a NACE code and a SIC code, an older classification. On 31,223 of them, 52%, the two codes map to different sections. That is the records disagreeing with themselves, counted across every company that holds two codes.

Two codes drive part of it, so here is the same figure with each set aside in turn.

**Chart: Contradiction rate, setting two codes aside in turn** (Companies carrying both a NACE and a SIC code whose codes map to different sections)

* All pairs: 52.0% n=60,026
* Less SIC 74990: 45.6% n=50,458
* Less NACE 6420 too: 39.5% n=42,640

SIC 74990 is not an activity at all: on Companies House’s list it means “non-trading company”. On its own it accounts for a quarter of every contradiction, and those cases set a sector against a company’s status rather than against another sector. The second deduction is the holding-company boundary. NACE 6420 is a pure holding company (section K, finance), while SIC 70100 is the activities of head offices (section M, professional activities). Both schemes separate a passive holding company from an active head office, so a company carrying both codes is a real disagreement, but one that only evidence about that company can settle. We set aside every company with NACE 6420 to show how much they drive the figure, not because those codes are always right.

Remove both and 39.5% remains: two in five of the 42,640 companies left still hold two activity codes that point at different sectors. That is the most conservative version: 45.6% if holding companies stay in.

## Which companies have a code at all?

NACE alone covers three fifths of the records. The older SIC scheme fills most of the rest. That is why “two in five companies have no activity code” is true of NACE but overstates the problem.

**Chart: Every company in the records, by which scheme classifies it** (347,902 companies · records as of 25 June 2026 · counted)

* NACE only: **43.6% (151,775)**
* SIC only: **30.1% (104,689)**
* Both: **17.3% (60,026)**
* Neither: **9.0% (31,412)**

UK SIC 2007 was derived from NACE Rev. 2 and shares its division numbering, so one lookup maps both to the same sections. Taking either scheme, 91.0% of companies carry some classification, and 9.0% carry none at all. That 91.0% counts SIC 74990, which means “non-trading company” rather than an activity, so the share with a real activity code is somewhat lower.

Most missing NACE codes belong to companies formed before 2010. Of the 93,068 companies formed before 2010, 164 carry a NACE code, and 99.92% of companies with a NACE code were formed in 2010 or later. The pattern suggests the records started capturing NACE codes around 2010, and the older companies were never given one. Most are not stale records: 86% of companies without a NACE code had their record updated in the two years before 23 September 2026. That does not show they are trading.

## Why can’t holding companies be checked?

22,414 companies, 10.6% of the companies with a NACE code, carry the one code that says a company does nothing observable. A holding company’s website is usually its group’s website, and the group trades in something. A construction group’s Irish holding company will have a construction website while being, correctly, a holding company. So the web evidence describes the group, the code describes the company, and reading the website cannot separate them. That is why holding companies are reported as their own group throughout, rather than folded into an accuracy figure, and why that choice moves the share of companies whose NACE code is missing, wrong or cannot be checked:

* **52.9%** if holding companies count as fine;
* **56.5%** if holding companies are left out entirely;
* **59.3%** if holding companies count as bad.

## Why might the codes look like this?

Most likely because each company chooses its own code. The people forming an Irish company give one NACE code for its main activity on Form A1, the form that registers the company with the CRO. The code is the company’s own description of itself, chosen at the start, often before the business has traded, and the CRO register records the code the company gives. The NACE codes in the records we checked most likely trace back to those choices.

After that, the business can change while the code stays the same. A software firm registers as a consultancy, then sells its own product. A manufacturer comes to mainly distribute. And because the code names one main activity, a business that does several things can only be found under one of them. When nothing on the list fits, a catch-all “other” code may be the easy choice.

The classification also splits the same trade from the inside. “Pharmaceutical” spans three sections depending on the verb: 2120, manufacture, is section C; 4773, a dispensing chemist, is section G, retail rather than healthcare; and 8621, a medical practice, is section Q. A pharmacy is shopkeeping. That is not a trick question. It is the everyday difficulty of applying this scheme.

## What should you use instead?

Classify businesses by evidence of what they do today, not by the code they chose at registration. Much of that evidence is public: official records such as company registers and filed accounts, a business’s own website, its online business listings, and the professional profiles of the people who work there. Read together, these show what a business actually sells, and to whom.

That is how SaleSight, Sunstone’s sales intelligence platform, classifies businesses in Ireland. It uses MAPS (Market Activity Profiling System), Sunstone’s own classification of what a business actually does today, built from evidence of how it operates rather than its registration. Separately from this audit, our own analysis finds that a market searched that way shows on average around 2.5 times as many relevant companies. The figure varies by sector.

## See what your code search is missing

Book a discovery call. In the demo, we’ll show you SaleSight working on your own market in Ireland, including how many more relevant companies it finds than a search by industry code.

[Book a discovery call](https://www.sunstone.ie/#contact) 

## How we measured it

We wrote the protocol down and committed it to version control before any result existed, so thresholds could not move once the numbers were visible.

1. **Freeze the frame.** Pin the records to one snapshot and draw a stratified sample of 2,400 companies across 22 groups: the 21 NACE sections, using the SIC code where NACE is missing, plus companies with no code, with a floor of 40 (or the whole group where it is smaller) so that small groups stay measurable.
2. **Build blinded dossiers.** Assemble each company’s evidence (registered name, address, website text, directory listings) and withhold its registered classification fields. A deny-by-default check refuses any field not on an allowlist, so no registered code goes into the dossier, though a website or listing could still mention one in its own text. Before the run with web search, a 100-case pilot checked the titles and web addresses of the search results (the page text comes back encrypted, so it could not be checked) for the words “NACE” and “SIC”, and for a four- or five-digit number within 40 characters of either, and found none in any of the 79 cases that searched.
3. **Judge twice.** Once reading only the dossier, and once with a web search tool. The same 2,400 companies both times. We chose the search run for the headline after both had come back.
4. **Anchor with people.** Three labellers classified 419 of the sampled companies (536 labels) from the same blinded dossiers, and could also look a company up online. A block of 100 was set aside for all three to label, but not all of it came back, so human-to-human agreement rests on 151 paired comparisons. Twenty known-answer control cases were mixed in, though their format made them easy to spot.
5. **Gate what may be said.** Six pre-registered thresholds decide what the pre-registered analysis may claim. Code reads them, not whoever writes the report. The adjudicated error rate was added later and sits outside them.

**Disagreement is not error.** When two sources differ, that alone does not establish which is wrong. A pizza company filed under “other manufacturing” is a register error; a games subsidiary correctly filed as a holding company looks like a disagreement but isn’t. Only adjudication, where a person settles each disagreement, separates the two, which is why counted and estimated figures are kept apart throughout.

### Two runs, similar answers

The same 2,400 companies were judged twice under different conditions. The two runs give similar rates, so the figure does not depend much on whether the model can search. Both measure disagreement, not error.

**The two runs side by side**

|                                                            | Dossier only           | With search            |
| ---------------------------------------------------------- | ---------------------- | ---------------------- |
| Model disagrees with register, excluding holding companies | 31.1% (25.8% to 36.4%) | 33.8% (30.5% to 37.2%) |
| Model disagrees with register, including holding companies | 33.5%                  | 34.7%                  |
| Judged cases, excluding holding companies                  | 380                    | 863                    |
| Judged cases, including holding companies                  | 394                    | 921                    |
| Model declined to classify                                 | 64.9%                  | 30.2%                  |
| Judged “not a trading business”                            | 13.1%                  | 5.4%                   |

Giving the model a search tool more than doubled the usable sample, from 380 judged cases to 863, and narrowed the interval from about 5.3 points either side to about 3.3. The point estimates moved 2.7 points and the intervals overlap heavily.

### What if you score whatever code a company holds?

The study was designed to score NACE codes. But a company filed only under SIC has still been classified by the register, just in the older vocabulary. Scoring whichever code is present answers a broader question, and reaches nine companies in ten instead of six, though that includes SIC 74990, “non-trading company”, which is not an activity.

**How often the model disagrees with the register, under each definition of “classified” (stratified, weighted, holding companies excluded)**

| Definition     | Records covered | Model disagrees        | Judged cases |
| -------------- | --------------- | ---------------------- | ------------ |
| NACE code only | 60.9%           | 33.8% (30.5% to 37.2%) | 863          |
| Any code       | 91.0%           | 34.6%                  | 1,477        |

Widening the question adds 614 judged cases, and lifts coverage from 60.9% to 91.0% of the records. The rate itself moves by 0.8 of a point.

This is a post-hoc analysis, and it would open a publication gate that the pre-registered one fails (see [“Publication gates”](https://www.sunstone.ie/resources/guides/irish-nace-code-accuracy/#gates)). That is the most suspect shape a result can take, so we state it plainly. The pre-registration argued that no rate could be called national, because coverage for any code-based rate was “capped at roughly 60.9%”. That premise was wrong: it treated “a code” as “a NACE code”. Because UK SIC 2007 shares NACE’s division numbering, a code-based rate reaches 91.0%, above the 90% the national-scope gate requires. That 91.0% counts SIC 74990, “non-trading company”, as a code. It is a status, not an activity, so with it set aside coverage may fall below 90% and the gate may not open.

The rate barely moved, from 33.8% to 34.6%, well within the margin of error, so the wider definition does not change the answer. We report both. The NACE-only rate remains the pre-registered primary result.

### The human benchmark

Three people classified companies from the same blinded dossiers, independently, and could look a company up. People often disagree on this task: pairs of labellers agreed between 57% and 81% of the time.

* **66.9%** human agrees with human, over 151 paired comparisons
* **67.1%** model agrees with human, over 419 paired comparisons
* **20/20** model on known answers, under both prompts

Human-to-human agreement (66.9%) and model-to-human agreement (67.1%) are the same to within rounding: the model agrees with the labellers about as often as they agree with each other. On twenty control cases with answers fixed before anyone saw them, the model scored 20 out of 20 under both prompts. These twenty were written to be unambiguous, so they show the model is not broken on easy cases, not how accurate it is on hard ones.

**Chart: Where model and human land on the same case** (412 companies labelled by both, in each run · human abstention was 51.9% in both runs)

Dossier only

* Both committed: **127**
* Both said insufficient evidence: **169**
* Human abstained, model answered: **45**
* Model abstained, human answered: **71**

With search

* Both committed: **174**
* Both said insufficient evidence: **91**
* Human abstained, model answered: **123**
* Model abstained, human answered: **24**

The “human abstained, model answered” segment is the interesting one. With search, the model answered 123 cases the human did not, and the agreement measure counts every one as a disagreement, whether the model’s answer was right or not.

Headline agreement fell from 67.0% to 58.0% when the model got a search tool. This does not show the model got worse: human abstention was 51.9% in both runs, and the model with search answered more of the cases the humans left open, each of which counts against it. Where both sides commit to a section, agreement is 84.3% without search and 85.1% with it. Whether its extra answers are right was not measured.

### Robustness: change the model, the prompt or the evidence

Each check changes one condition, keeps the same companies, and compares the verdicts case by case. A stable rate built from unstable verdicts would be a weak result, so the table reports both.

**Robustness checks**

| Check                                                               | Rate before | Rate after | Kappa | Agree, both committed |
| ------------------------------------------------------------------- | ----------- | ---------- | ----- | --------------------- |
| A different model (400 cases)                                       | 33.6%       | 39.8%      | 0.798 | 89.5% (n=276)         |
| A search engine’s business category withheld (344 cases, no search) | 42.7%       | 43.9%      | 0.829 | 96.4% (n=281)         |
| Prompt fully reworded (400 cases)                                   | 33.6%       | 34.2%      | 0.875 | 96.7% (n=270)         |

Kappa is agreement corrected for chance: 1.0 is perfect, and the study’s bar for “reliable” is 0.6. All three checks clear it. The second model disagrees with the register more often, but it also declines less, which may mean it is answering harder cases. Withholding the search engine’s category made the judge decline more often, and changed few of its conclusions when it did answer (96.4% agreement).

### Adjudication: when the register and the model disagree, which is wrong?

In the cases one person could settle, usually the register. A disagreement only says that two sources differ. To find out which is at fault, one adjudicator took 100 disagreements, with full information and free to search, and chose between two unlabelled answers, without being told which came from the register.

The headline’s split between wrong and right codes multiplies two measurements: how often the model, reading the dossier and its own web searches, disagrees with the register (33.8%), and, when they disagree, how often a person settling the case found the register at fault (74.6%). That puts the register wrong on 25.2% of companies with a NACE code that are not holding companies, with a 95% confidence interval of 20.8% to 29.6%. The second measurement rests on one adjudicator and 63 resolved cases. It assumes the third of disagreements nobody could settle split the same way as the rest.

**Chart: 100 adjudicated disagreements**

* Register wrong: **47**
* Both wrong: **2**
* Register right: **13**
* Both defensible: **4**
* Could not tell: **34**

Of the 66 cases the adjudicator could settle, the register was wrong in 49: 74.2%, with a 95% interval of 63.7% to 84.8%. Excluding holding companies, it was 74.6% of 63 (63.9% to 85.4%).

* **33.8%** model disagrees with the register
* **74.6%** of the disagreements one person could settle, the register was at fault
* **25.2%** register wrong (20.8% to 29.6%)

The correction runs in the register’s favour. Before adjudication, every disagreement was implicitly counted as a register error, which put the “present and correct” share at 36.0%. In about a quarter of the cases the adjudicator could settle, the register’s code was right or just as defensible. Correcting for that lifts the share to 40.7%, against the case we set out to make.

* **How the denominator was ruled.** “Could not tell” is left out: it is a limit of the evidence, not a fault of either source. “Both defensible” stays in, because dropping it would remove exactly the cases most favourable to the register.
* **Controls.** Twenty hidden cases with known answers were mixed in. The adjudicator chose correctly on all 12 they committed to and declined the other 8. Declining is not scored as an error.
* **What to hold this to.** One adjudicator; a third of cases could not be settled; the sample was not weighted to the register; and the analysis was designed after the judged results came back, so it is not part of the pre-registration.

### Publication gates: what the pre-registered analysis may claim

Six thresholds were fixed in advance and are enforced in code. They apply to the run with search; the run without search was not tested against the model-agreement gate.

**The publication gates**

| Claim                                                                       | Gate                                                                    | Why                                                                                                                      |
| --------------------------------------------------------------------------- | ----------------------------------------------------------------------- | ------------------------------------------------------------------------------------------------------------------------ |
| Publish a headline error rate                                               | Blocked                                                                 | Agreement with people 0.58; needs 0.85                                                                                   |
| Publish reliability from model agreement                                    | Open                                                                    | Cross-model kappa 0.798                                                                                                  |
| Call any rate “national”                                                    | Blocked for NACE only; any code reaches 0.91, but that counts SIC 74990 | NACE-only coverage 0.61; any-code coverage 0.91, which counts SIC 74990 (non-trading) and may fall below 0.90 without it |
| Publish rates for each section                                              | Blocked                                                                 | 0 of 21 sections reach n=150                                                                                             |
| Coverage caveat on every estimate (the abstention and unjudged-share gates) | Fired                                                                   | 30.2% of sampled companies unjudged, above the 25% threshold                                                             |
| Publish counted disagreement (not a pre-registered gate)                    | Open                                                                    | Counted, not sampled                                                                                                     |

The 52% contradiction figure is counted, not sampled, over every company holding both codes (60,026), so none of the sampling gates touch it. The estimated rates are the ones that carry caveats. Under the fired coverage gate: the model gave no answer for 30.2% of the sample, and each rate rests only on the judged cases eligible for it: for example, 863 for the NACE-only rate behind the headline and 1,477 for the any-code rate. The adjudicated 25.2% error rate was designed after the results came back, so it sits outside these gates altogether and is reported as an additional analysis.

The error-rate gate did not open on this evidence. It requires 85% agreement with the human labellers, and even the best-matched pair of labellers agreed with each other only 80.7% of the time, on 57 cases. It was pre-registered and the results have been seen, so it stays: moving a threshold after seeing the data it blocks is exactly the failure pre-registration exists to prevent.

### What would change these numbers

* The records are a snapshot from June 2026, and it is the only one the study had. Everything here describes that snapshot.
* The 52% is a contradiction rate, not an error rate. It establishes that the records disagree with themselves; it does not say which code is wrong. The sampled study is what measures error, and adjudication estimates that 25.2% of coded, non-holding NACE entries are wrong (20.8% to 29.6%).
* Four in five companies in the records have no website on record. The typical piece of evidence is a company name and an address. That is the biggest single constraint here, and it is a fact about the records rather than a flaw in the method.
* The human anchor, not the model, is now the binding constraint. Three labellers abstained on about half their cases, which suggests about half cannot be decided from the evidence on file. Doing more needs more human labelling, not more computing.
* Adjudication rests on one person. A second reading was planned and did not arrive, so there is no measure of how far two adjudicators would agree. A third of the disagreements could not be settled even with full information, and the 74% describes only the two thirds that could. The adjudication sample was not weighted to the register.
* One robustness check was not run: asking the same model the same questions a second time. The three that were run all cleared the agreement bar; the second model’s rate was six points higher.

**Technical note.** Data: a commercial dataset of Irish company records as of 25 June 2026, 347,902 companies, each with whatever NACE and SIC codes are recorded for it. Contradiction and coverage figures are counted over all 347,902 records. Stratified sample of 2,400 companies across 22 groups (21 NACE sections, using SIC where NACE is missing, plus companies with no code), with a floor of 40 for each group, or the whole group where it is smaller. Judge: a large language model at low reasoning effort, run twice, once on the dossier only and once with a web search tool. Human anchor: 3 labellers, 419 distinct companies, 536 labels and 20 control cases. Adjudication: one reader, 100 disagreements plus 20 hidden controls. Robustness: a second model, a reworded prompt, and a search engine’s business category withheld. Counted facts and sampled estimates are stored separately throughout, because the usual way figures like these get misquoted is a sampled number reported as though it had been counted.

## Sources

* Companies Registration Office, [Information Leaflet No. 1: Company Incorporation](https://cro.ie/information-leaflets/company-incorporation-leaflet-1/) (the NACE code on Form A1).
* Regulation (EC) No 1893/2006, [establishing NACE Rev. 2](https://eur-lex.europa.eu/eli/reg/2006/1893/oj).
* Eurostat, [NACE overview](https://ec.europa.eu/eurostat/web/nace).
* Companies House, [Choosing a Standard Industrial Classification (SIC) code for your company](https://companieshouse.blog.gov.uk/2021/10/12/choosing-a-standard-industrial-classification-sic-code-for-your-company/), and its [list of SIC codes](https://resources.companieshouse.gov.uk/sic/) (74990, 70100).

---

This is a Markdown copy of https://www.sunstone.ie/resources/guides/irish-nace-code-accuracy/, published by Sunstone (Sunstone Technologies Ltd).
