Platform · Company Brain and Decision Assurance

The operating foundation for governed AI agents.

Every Annex agent runs on one shared layer. The Company Brain gives it your company’s live state, intent, authority and decision memory. Your company’s Annex Gate independently decides what executes, and Commit acts only through your own systems, keeping the reconciled outcome as evidence. Connect sends only what your Gate authorises to Annex Network, which verifies transactions between companies.

The operating stack

1 · Map1MapReviewed workflow map2 · Ledger2LedgerSource-linked facts3 · Gate3GateChecked authority4 · Commit4CommitCommitment states5 · Connect5ConnectEvidence for both sidesAgentworks on top of the stackAnother companyAnnex Networkverifies both sidesYour companypoint at an icon to look inside1 · Map12 · Ledger3 · Gate4 · Commit5 · ConnectAgentAnother companyAnnex NetworkYour companyselect an icon to look inside

Layer 1 of 5

Map

In every deploymentPart of Company Brain

Connect only the sources and actions an agent needs. Record which system is authoritative for each fact, and check the mapping against the source.

Inputs
The agent’s systems, e.g. ERP, inbox, spreadsheets
Output
Reviewed workflow map

The company keeps its applications. Nothing is replaced.

Company Brain and Decision Assurance

What is in the shared layer?

Map and Ledger make up the Company Brain; Gate and Commit make up Decision Assurance. All four are part of every deployment, scoped to the first agent. Connect is added when a second company is involved.

Annex components: whether each is part of every deployment, what it covers for a first agent, and which part of the platform it belongs to
ComponentEvery deploymentFor a first agentPart of
MapYesThe systems the first agent reads and writes, and the authoritative record for each fact.Company Brain
LedgerYesOnly the facts, intent and authority that agent needs, each with its source and freshness.Company Brain
GateYesThe rules, mandates and approval thresholds that apply to that agent’s actions.Decision Assurance
CommitYesThe commitments that agent can create, and the source-system path for each.Decision Assurance
ConnectNoNot needed for a first agent. Added when two companies agree to transact.Your link to Annex Network

Ledger · provenance and freshness

How does an agent know a fact is current?

How the Company Brain treats the facts an agent relies on. The exact limits are agreed per agent when the deployment is scoped.

  • Every fact has a source

    Each fact the agent uses points to the system and record it came from, so a person can check it.

  • Freshness is part of the fact

    Facts carry when they were last confirmed. Limits are set per kind of fact: stock may need a re-read before every commitment, a price list holds until its end date.

  • One record is authoritative

    When two sources disagree, the mapping says which one decides. The agent does not average or guess.

  • Stale means stop

    If a fact is too old or a source cannot be reached, the agent re-reads it or routes the work to a person. It does not act on an old copy.

Decision Assurance · authority

Business authority is not an app permission.

Access lets an agent read. App permission lets it use a tool. Authority decides whether it may commit your company. Execution happens only through your system of record, after a fresh check. Evidence shows that it actually happened.

Application permissions
  • What the agent’s credentials can technically do in each application.
  • Example: read the price list, create a draft order in the ERP.
  • Set by IT, system by system.

A permission says the agent can. It does not say the company agreed.

Business authority, checked by the Gate
  • What the agent may agree to on the company’s behalf, and until when.
  • Example: discounts up to 7% and quotes up to €5,000; above that, J. Weber approves.
  • Set by the business; the Gate checks it before each action.

Agents propose. The Gate decides what executes: a mandate has an owner, limits and an end date, and above a limit a person approves.

Inside the company, and between companies

Where your Gate ends and the Network begins.

The Annex Gate is local: it governs what agents may access, disclose, commit or execute for your company. When a transaction involves another company, the Network verifies what your Gate authorised, and the other company’s own Gate decides its side.

  1. 1

    Your agent proposes

    A request, an offer or a commitment that involves another company.

  2. 2

    Your Gate authorises

    Your permissions, mandate, limits and current state, and what may be disclosed.

  3. 3

    Annex Network verifies

    Identities, mandates, versions and message integrity, from the evidence each Gate authorised.

  4. 4

    Their Gate decides

    Anything that needs the other company’s authority, under its own rules.

For technical evaluators

What happens when something changes or fails?

Short answers about how the platform behaves. Each one is tested against your real systems before an agent acts.

What does Decision Assurance check before an action executes?

The Gate checks the action’s sources, your business rules, the agent’s mandate and any approval it needs, then re-reads the state at execution time. If every check passes, the action executes through your system; if one fails, it is blocked; if it needs a person, it is escalated. Agents hold no direct write access to systems of record.

What happens when a fact is stale?

The Ledger marks it. Before any commitment, the agent re-reads the authoritative source; if it cannot, it stops or routes the work to a person. Freshness limits are agreed per kind of fact when the deployment is scoped.

How are competing reservations handled?

Commit changes a reservation only through the system that owns the stock or budget, so that system decides who gets the units. If it cannot hold stock, Annex records an offer, not a reservation. Concurrency is tested against the real system before the agent acts, and Annex claims no more than that system enforces.

What if authority changes mid-transaction?

The Gate checks the mandate at each action, not only at the start. A mandate that has expired or been reduced stops the next step and sends it to a person.

What happens if an action fails?

The commitment stays in its last confirmed state. A failed write is recorded and retried or routed to a person, and a retry never creates a duplicate. The agent does not report a result it has not seen in the source system.

Where does the data stay?

In your own systems. Annex works with references to it: sources, states and evidence. Hosting region, model providers and retention are agreed per deployment.

Which systems does Annex work with?

The systems you already run. Nothing to replace: connectors are chosen from the systems the agent’s work already uses.

Which AI models does Annex use?

Annex is not tied to one model provider. It works with the models you choose, agreed per deployment. The Gate checks every action the same way, whichever model it came from.

Can agents Annex didn’t build take part?

Yes. The enterprise Company Brain gives governed context to the agents you already run, and Annex is designed to admit qualified third-party agents under the same Gate.

How is the Gate different from Annex Network?

The Gate is inside your company: it decides what your agents may access, disclose, commit or execute. Annex Network sits between companies and verifies identities, mandates and transaction integrity from what each Gate authorised. Neither can act for the other.

Does Annex Network see our Company Brain?

It receives the disclosures your Gate authorises for a transaction, with the evidence needed to verify them. It verifies them without requiring unrestricted access to your Company Brain, which stays with your company.

What does the next agent reuse?

A customer starts with one role. Every next role reuses the layer: each next agent starts with the context, authority and controls already mapped, and adds only what is new. Every decision and outcome is recorded, so value can be measured against your own baseline.

One shared layer

Why one layer instead of one integration per agent?

Every agent needs facts, actions and authority. The next agent reuses what the first one mapped and adds only what is new, which is designed to lower the effort of each next deployment.

Without a shared foundation
  • Annex SellIts own integrations
  • Annex BuyAnother integration project
  • Annex CollectDuplicate rules and context

Each new automation can become another isolated system.

With Annex
  • Annex Sell
  • Annex Buy
  • Annex Collect
Shared operating layer

Mappings, policies, authority and commitment checks, reused where they apply.

Make the first deployment count.

Start with one agent.

A deployment maps the first agent’s sources, authority and commitment checks, and tests them before the agent acts.