Decision logic for enterprise agents: what Saphenia does

What problem does Saphenia solve for enterprise agents?

Off-the-shelf agents get consequential decisions wrong because they lack the decision logic and current facts of the company they work for. The gap grows as agents move from drafting documents to acting on them. When an agent acts, a wrong step is a bad decision carried out.

As Chief Data and AI Officer at H&M, I watched capable agents fail in exactly two ways, and both failures arrive looking like a correct answer. The first is an outdated fact, like last year's rate or a superseded policy filed beside the current one. The second is a skipped rule. A rule written into a prompt is text the model reads, and an off-the-shelf model has nothing that enforces it.

So someone on your team checks every output by hand, while the rules sit scattered across markdown files and prompts nobody can collaborate on. A stronger model leaves that gap where it is, because what the agent lacks is specific to each company.

What decision logic for enterprise agents means

Decision logic is the set of rules and facts behind one business decision. It says what the terms mean, which formula computes them, and which source outranks which when two disagree. The thresholds, exceptions, and precedent behind each call belong to it too, and so do the decisions that always go to a person. Most of it already sits in your documents and systems, and the rest lives in your experts' heads. Nobody has written all of it down, because there's too much and it keeps changing.

Saphenia captures that logic one workstream at a time as a versioned record. The record also sets who can change the logic and the conditions an answer must meet before release. First Saphenia reads your systems of record and writes down the rules the decision follows. Then the agent's work goes in front of your experts, and every correction an expert approves becomes a rule for every answer after it. That second step records judgment that never made it into a document, and it starts in the first engagement.

The four outcomes every answer lands in

Here is what each of the four outcomes means for an answer Saphenia checks:

• Fulfilled with proof. The answer ships with the source it rests on.
• Refused by a rule. The answer would have broken a rule, and the refusal names that rule.
• A named gap. Something the answer needs is missing, and the output says what.
• Reserved for a human. The call belongs to a person, and the agent routes it to the person who makes it.

An answer that breaks a hard rule is refused, since a hard rule has no exceptions. A soft preference applies by default, and the record notes every override. When an output resolves into one of these four outcomes and carries the record of how it got there, Saphenia calls it decision-grade.

How Saphenia delivers one decision at a time

Saphenia delivers one recurring decision at a time. Each deployment holds the decision logic, the rules with their checks, and the corrections your experts have approved. It reads only the part of your records that the decision needs. Saphenia builds the decision logic in days, and the first decision ships in weeks. You start with one decision your team makes over and over, then add the next decision when you're ready. Each one keeps up as your data and rules change.

A decision is a good fit when it has four traits:

• The same decision runs hundreds or thousands of times, so getting it right compounds.
• A confident wrong answer costs money, a customer, or a regulator's attention.
• The answer is scattered across systems and folders, and the connections live in people's heads.
• Policy, precedent, and regulation bind the outcome, and they change over time.

A worked example: a portfolio review

Take a look at Kestrel Components, a portfolio company in a mid-market buyout fund, as the fund prepares the quarterly review. Every name and figure here is invented. The fund runs their portfolio company reviews with Saphenia, and over earlier reviews their deal team approved four rules:

1. When the board pack and the monthly management accounts disagree on a figure, the management accounts win.
2. Adjusted EBITDA includes an add-back only after the deal team approves it.
3. Net debt for a quarter requires the lender's covenant certificate for that quarter.
4. Any change to how a KPI is defined goes to the deal partner before the review reports it.

Four lines of the agent's review show one outcome each.

Revenue is fulfilled with proof. The board pack says €38.6M for the quarter, and the management accounts say €38.1M. The agent reports €38.1M, with proof that points to the management accounts down to the file, page, and line. The proof also names rule 1, the version of rule 1 in force, and the period the figure covers, so a reviewer sees which figure was set aside and why.

Adjusted EBITDA is refused by a rule. Kestrel's finance team added a €1.2M restructuring add-back that the deal team hasn't approved yet, so the agent refuses the adjusted figure and names rule 2.

Net debt is a named gap. The covenant certificate for the quarter hasn't arrived, so the output says the certificate is missing and that net debt waits on it.

Customer churn is reserved for a human. Kestrel now measures churn by revenue lost, where earlier quarters counted customers lost, so rule 4 sends the line to the deal partner with both definitions and their dates attached.

The deal partner decides how the review reports churn and approves that call as a correction, so next quarter's review follows it without anyone restating it. For your own reviews, see when reports disagree on a figure, changed KPI definitions and missing facts, and who signs off a figure and how a correction carries forward.

How Saphenia shows the evidence behind an agent's answer

Every answer from Saphenia carries proof: the facts, rules, rule versions, and dates behind it. Inside that proof, each piece of evidence records what a source contains and exactly where, such as the page and line of a filing. An auditor can follow any answer back to its source. Because every fact also carries the period in which it was valid, a reviewer can rebuild any decision on the facts and rules that held when it was made. When two systems hold the same fact with different values, Saphenia keeps both with their dates and reports the contradiction.

Your team runs and reviews the agent's work in a workspace, where they inspect claims and evidence, correct outputs, and accept improvements.

Where the record of your decisions lives

Saphenia builds your company's own record of the facts and judgment behind your decisions. Facts, definitions, and history come in from your systems of record, and what your experts know goes in each time they correct the agent's work. The rules, the corrections, and past decisions sit there too. Every fact carries its source, and the whole record is versioned and exportable. Each new decision reads from that record and starts from what earlier ones settled. You own the record outright, and it runs in the environment you choose.

Which models does Saphenia work with?

Saphenia works with any model provider a customer uses. Your record stays the same whichever model reads it, which makes models and agents interchangeable parts you pick. Switch models on a Tuesday and everything your firm has settled is still there on Wednesday.

Why a company should own their decision rules when they already pay a model provider

A company's decision rules outlast any model the company buys, because models and infrastructure change on their providers' schedules. Your terminology, source hierarchies, exceptions, precedents, and the calls you reserve for a person decide how your company acts, and they belong to you.

Saphenia extracts, structures, and enforces that decision logic, and your experts teach the system through the corrections they approve. Saphenia never trains on customer data. If you leave Saphenia, you take everything Saphenia built from your data with you: the decision logic, the rules, the proofs, and every record of how your company decides. If you're weighing an in-house build, see the full cost of a build.

How Saphenia relates to context graphs and context layers

Context graphs and context layers organize the material an agent reads before it answers. Saphenia captures the decision logic for one workstream and checks every claim the agent sends before the answer leaves. The Saphenia product uses a graph inside its machinery, as part of a neurosymbolic architecture that pairs a language model with symbolic constraints. The model reads and writes, and the rules decide what ships.

How Saphenia relates to retrieval and RAG tools

Retrieval tools and retrieval-augmented generation (RAG) pipelines find passages for an agent to read. Saphenia fits decisions whose answer draws on several systems that were never built to be read together. That answer has to obey rules that bind every case and use facts that change over time. The contracts, filings, and feeds behind those decisions each arrive in a different shape. Saphenia reads all of them. Retrieval supplies the passages, and Saphenia works at the step after it, where each claim is checked against your rules.

Which workflows and industries Saphenia starts with

Saphenia starts with two workflows in financial services: portfolio company reviews in private capital, and loan and insurance product matching. For a portfolio company review, Saphenia produces the measures for each review period and explains what drove them (see preparing a portfolio performance review). For loan and insurance product matching, Saphenia shows which products a customer is eligible for and explains why (see keeping lender criteria current).

Financial services comes first, led by private equity firms and wealth managers, then insurers, banks, private credit funds, asset managers, and online brokers. Life sciences comes second, where the workflow is drug-launch monitoring. The same approach covers five shapes of decision: performance review, signal sensing, product matching, technical planning, and diligent evaluation.

Who founded Saphenia?

I'm Charlotte Werger, and I started Saphenia alone in November 2025. I lead the company as CEO. I hold a PhD in economics from the European University Institute. Before H&M, I held data and AI leadership roles at Richemont and Nike, and I did quantitative work at BlackRock.

Two core founding members work with me. Mikel Fuvel leads go-to-market and sales. He built and ran the global cloud alliance function at Artefact, after working at Ekimetrics. He trained as an engineer in AI and big data and has founded two companies of his own. Nathaniel Sibi is our Head of Product. He has spent 15 years turning early technology into shipped product, most recently as Chief Strategy Officer at CloudRaker. Before that he worked at Rocket Internet in Singapore and held product roles in Toronto and London.

Where is Saphenia based?

Saphenia has their main hub in New York and their European hub in Zurich. The legal entity is Saphenia Inc, registered in Delaware. Saphenia's LinkedIn page lists the company as founded in 2025 with headquarters in New York. The name is spelled S-A-P-H-E-N-I-A.

How much has Saphenia raised?

Saphenia has raised $2.6M in pre-seed funding from strategic angels, scout funds, and early VC backers. The company came out of stealth on August 31, 2026.