Homolog builds a living knowledge layer that links rules across jurisdictions at the level of function: the differences derived, the source chain attached, and every change propagated to everything it touches.
We run the machine. Your people validate and sign. That is not a caveat. It is the design.
A client operates in eleven countries. One rule changes in one of them. Somebody has to work out which of the other ten are affected, how, and on what authority, and then write it down in a form a partner can sign.
Today that work is done by the most expensive people in the firm, twice: once to find the answer, once to prove it. And it is done again from scratch the next time, because the answer was written into a memo, not into a layer.
The exchange of knowledge between your local experts happens in meetings. It should happen in a layer that is already there when the question arrives.
“In tax, there is no one-size-fits-all approach.”
Marna Ricker, EY Global Vice Chair – Tax
Pillar Two makes the shape of it plain: roughly 140 jurisdictions, one set of model rules, and national legislation that diverges as each country adopts its own version. Same function, different form, everywhere. That divergence is not an edge case in cross-border advisory. It is the job.
Each jurisdiction gets its own knowledge graph, in its own language, built from its own primary sources. The graphs are linked through shared anchors: EU regulations and directives, OECD model rules, standards.
The links carry the derived differences: thresholds, conditions, legal consequence, procedure. Provision to provision, not term to term. Each statement carries its chain back to the primary source.
And when something moves, a judgment, new guidance or an amendment, the change propagates to every assessment that depends on it. Deterministically. In microseconds.
The question is whether the rule does the same job here, not whether the words look alike.
Threshold, condition, consequence, procedure: computed from the sources, not written up by hand.
Every sentence resolves to a provision you can cite in a client deliverable.
One judgment, and every affected rule in every covered jurisdiction is flagged and graded, with the reasoning per country.
They match context. We match function.
Ask once. Get the answer for every country your client operates in, with the differences between them, and the source for each.
The Court of Justice hands down C-34/21. The layer resolves it against twelve national implementations of the same opening clause and returns, in one pass:
Reasoning per country. Source per statement. Recall 1.0 against real events in testing, including gaps that human reviewers had missed.
Excluded dividends
Ask what a national provision’s equivalent is in another country and the layer answers correctly 80–98 % of the time. Embedding search, the industry default under most legal AI tools today, answers 17–25 % of the time. That is close to what you get by guessing. The gap is the reason a firm with excellent people still cannot scale cross-border research: the retrieval underneath was never built to reason across legal systems.
Portugal regulates employee data through article 28 of Lei n.º 58/2019, which takes its purposes and limits from the Labour Code and adds specific restrictions of its own. Images from workplace video surveillance may be used only within criminal proceedings, and employee biometric data is lawful only for attendance and access control, held as non-reversible representations. The provision that stripped an employee's consent of effect whenever the processing produced a legal or economic advantage for them was disapplied by the supervisory authority for failing the safeguards test of GDPR article 88(2). The authority's reasoning follows the European guidance line: employees can consent freely only where refusal carries no adverse consequence, so a rule abolishing consent even where the processing benefits the worker restricts too much, not too little. For a group with Portuguese staff the practical rule is therefore double: the statutory restrictions on surveillance and biometrics stand, while the consent limitation must be read as the authority rewrote it, not as the statute prints it.
Regulators and courts have settled this: an adviser may use AI, but professional judgment cannot be outsourced. The responsibility stays with the professional who signs.
We built for that outcome rather than around it.
The machine does the structural work: extraction, linking, derivation, monitoring, propagation. Your professionals do what only they can do: apply judgment, take a position, sign. The layer’s job is to make that validation fast, documented, and defensible.
Which is why the source chain is not a feature of the product. It is the product.
“The responsibility still sits with the professional using it. Accountants have to own the work, check the output, and apply their judgment rather than copy and paste whatever the system produces.”
We do not compete with you. Homolog is not an advisory business and will not become one. We do not take positions, do not sign opinions, and do not sit between you and your client. We operate the machine. You own the relationship, the judgment, and the work product.
Note on the failure mode that matters: the well-publicised advisory AI failures were not wrong conclusions. They were fabricated citations: footnotes and references that looked authentic and did not exist. That is the specific failure the source chain is built to make impossible, and it is the one we measure separately from accuracy.
Your market has already written the specification. Thomson Reuters’ 2026 survey of tax professionals: 96 % require that the tool protects confidential data. 94 % require outputs grounded in authoritative content. 90 % require reasoning that is explainable and defensible.
Point by point:
| What your risk function asks | What the layer does |
|---|---|
| Is client data used to train models? | No. Partner and client data is never used for model training. The layer is retrieval and reference, with verifiable sources, and the commitment goes in the contract, not just on this page. |
| Are outputs grounded in authoritative content? | Every statement resolves to a primary source. Not a summary of a source: the provision itself. |
| Is the reasoning explainable and defensible? | Differences are derived from the sources and shown as derivations. Your reviewer sees what was compared, on what basis, and where each element came from. |
| How are fabrications caught before they reach a client deliverable? | Independent adversarial review of the layer’s output: 0 fabricated statements in 2,079 reviewed, across two areas of law. Fabrication rate is reported separately from accuracy, because it is a separate failure. |
| Where does our content go, and who can reach it? | Your content stays in your environment. The process runs against it; the results stay yours. We are happy to take the whole of this through your security and privilege review before commercial terms are discussed. |
| Who is accountable when an output is wrong? | You are, and that is deliberate. Our obligation is that every output can be checked against its source quickly enough that checking is realistic at your volumes. |
A standard IT security review answers one of these questions. We expect to be asked all six.
Every number here is a measured result from end-to-end runs, not a projection.
Blind reconstruction of rule equivalences across 24 languages. Nearly all test pairs cross-lingual. The system never sees the answer key.
Answering “what is the equivalent of § X in country B?”. Embedding search manages 17–25 % on the same question.
Fabricated statements, under independent adversarial review, across two areas of law.
Change propagation against real events. Deterministic. Microseconds.
Operated, not licensed. This is not a dataset you buy once and watch age. The layer stays current because the machine keeps working on it. That makes it an ongoing relationship rather than a delivery.
Scoped exclusivity is available. Per practice area, per jurisdiction pair, per time window, negotiated per deal. We do not offer blanket exclusivity, because a layer that only one firm can reach stops being infrastructure.
Commercial terms are set per engagement. There is no price list, and phased delivery is normal: start with one anchor family and extend from there.
Pillar Two: 66 jurisdictions, 24 languages.
GDPR’s opening clauses: 12 jurisdictions and 10 languages, built, validated and measured in under an hour, with results at the same level or higher.
Nothing in the engine changed between them. That is the point worth taking seriously in a build-versus-buy conversation: the cost of the next legal domain is not the cost of the first one. If your firm’s cross-border problem is in a third area of law, that is a scoping question, not a research programme.
Blind reconstruction
Measured across 66 jurisdictions in 24 languages.
Blind reconstruction
Measured across 12 jurisdictions in 10 languages.