Skip to contentABR Labs.

ABR / engineering

AI engineering.

Anyone can demo an agent; production is the test that matters. We build the agent that owns the job, wire it into the systems your team already uses, and stay on it while it runs.

Built in your AWS tenancyTested before it actsRun month to month

the checks · oursMore on this layer
the agent
the working memory
your systems

every build stacks the same four layers

The jobs an agent can own.

The month-end, the pipeline, the paperwork: the work that fills finance, accounting and HR weeks. We build agents that own these jobs end to end, and the irreversible steps wait for your sign-off, under thresholds you set.

month-end closereconciliation draftingtransaction codinginvoice matchingapplicant screeningthe hiring pipelinevariation processingpayment runs
  1. Month-end closes

    The period ends and the work lands.

  2. The agent reconciles and drafts

    It reconciles billing, ERP and bank feeds, then drafts the journal entries with the supporting docs attached.

  3. Checked against your policy

    checked against policy

    Every entry is checked against your close policy and approval limits before it can post.

  4. The controller signs off

    your call

    A person approves before anything touches the ledger.

  5. The close stops sprawling

    The reconciliation and drafting that filled the week is done when the controller sits down.

It runs one hundred times before it acts once.

A new agent runs one hundred times on your live work before it touches anything. Ninety-nine have to match what your team would have done, judged by your own reviewer.

one hundred runs · the bar sits at ninety-nine

A working memory of how your business runs.

The agent is only as good as what it knows about your business. The working memory holds your data, your rules and the decisions behind them, so the agent acts on what your company already knows.

The more of your business it holds, the more the agent can take on, and every action still passes through the checks.

An organised lattice of many small nodes, some linked and some standalone, a few picked out in burgundy: a structured field of what the business knows.
approval limitssource: delegation policy · visible to: finance, ops
supplier listsource: your ERP · visible to: finance
hiring rulessource: HR policy · visible to: HR
variation budgetssource: contracts register · visible to: ops

Every answer carries where it came from and who is allowed to see it.

After discovery, the work has two priced stages.

Build is one project with a finish line, and Run is the monthly subscription that keeps it working after it ships.

Every engagement starts with a fixed-fee discovery. See how the three stages are priced.

Still scoping the problem before you commit to a build? Start with AI consulting.

your ERPemail + documentsoperational systemsapprovals channelthe agentthe working memorythe checksthe recordeach actionsources on recordsign-offs
one archetype build · no client data · hover, tap or tab through the parts · drag sideways for the rest

fydis.

Our product

Guardrails and audit for AI that acts in regulated industries, where you have to prove what happened. It ships inside every ABR build, and it stands on its own.

13:02:11payment run 4471 · approval limitsverified
13:02:14new payee added · approved payee listheld
13:02:19export request · client data policyblocked
Visit fydis.io (opens in a new tab)

Start with one job.

We will walk you through the build that would own it, and quote it before any work begins.