Systems we have designed and built, every one specified in full first.
Financial services, regulated industry, media, logistics and professional services. Each is a small set of narrow, testable agents with defined handoffs. Client engagements are described without identifying detail.
Investment management firm
An operating estate under one supervisor, built to remove the daily chase across systems and inboxes.
Intake & triage
Reads every inbound request, classifies it, and routes it to the right desk before a person opens the inbox.
Reconciliation runner
Matches trades and cash movements across the custodian and the internal ledger, then builds the exception list.
Correspondence drafter
Prepares client replies from approved language and holds them at the approval gate.
Desk supervisor
Coordinates the other agents, tracks case state, and escalates anything that stalls past its window.
Regulated industrial sector
Multi-jurisdiction filing with human review, and an inspection-ready evidence trail assembled as the work happens.
Filing assembler
Gathers the required fields per jurisdiction, validates them, and prepares each submission for sign-off.
Deadline watchdog
Holds every statutory date and escalates until the filing is confirmed closed, never just sent.
Evidence collector
Captures the confirmation, the timestamp, and the underlying record into an audit trail as it goes.
Regulation monitor
Watches the published rules and raises a change request when a requirement moves.
Specialty finance
A full policy lifecycle that holds case state for years and replaces the spreadsheet everyone depends on and nobody trusts.
Application intake
Extracts structured data from messy inbound documents and opens a clean case record.
Underwriting support
Assembles the file against the playbook and ranks the deviations for a human to judge.
Servicing runner
Drives each policy through its scheduled events and holds state across the full term.
Collections ladder
Runs the ageing sequence with the right tone at each step, and stops the moment a human takes over.
Royalty recovery platform
Built for a client in music rights. Reconciles catalogues across rights databases that disagree by design, and surfaces registrations artists were already owed.
Catalogue matcher
Resolves the same work across databases that spell, code, and credit it differently.
Registration checker
Compares what is registered against what should be, and flags the gaps worth chasing.
Claim assembler
Packages the evidence for each recoverable registration so a person can file with confidence.
Provenance tracker
Tags every number as measured, stated, or estimated, so nothing is ever invented in a claim.
Freight brokerage operations
Built for a mid-sized broker drowning in load coordination. Keeps every shipment moving without a person holding the thread.
Load intake
Reads inbound tenders, checks them against lane rules, and books or flags each one.
Carrier coordinator
Tracks every external carrier, what they owe, and where each truck actually is.
Exception desk
Catches the delayed, damaged, and disputed loads early and routes them to a human with context attached.
Billing runner
Assembles what is billable per shipment, raises it, and reconciles the payment against it.
Accounting firm back office
Built for a growing firm losing partner time to admin. Takes the repetitive filing and chasing off senior people.
Document filer
Turns client uploads into structured, correctly filed records without a person renaming files.
Deadline watchdog
Holds every client filing date and chases the missing inputs before the date, not after.
Correspondence chaser
Follows up on outstanding items in the firm's own voice and stops when the client replies.
Approval router
Puts every action that needs a partner into one clean queue instead of a scattered inbox.
Every agent above runs the same five stages. None of them is the prompt.
This is the shape of every system on this page. The model does a small amount of work in stage two. The other four stages are what make it survivable in production.
Ingestion
The trigger and the source of truth. An inbox, a webhook, a form, a nightly query. Fields are validated and normalised before anything reasons over them.
Decision logic
The playbook, written down before the model sees it. Rules first, judgement second, and a recorded reason for every branch taken.
Human gate
Anything above the risk line stops here as a proposed action with its evidence attached, waiting for one approval click.
Execution
The write happens in your system, once, idempotently, with the record id and timestamp captured on the way through.
Feedback loop
Every override, correction and exception is logged and reviewed. The playbook changes; the agent does not drift on its own.
No agent we build gets the final word on anything that matters.
Speed without a gate is just a faster way to be wrong at scale. Each agent pattern we design carries an explicit human gate, decided with you and documented in the blueprint.
The risk line is set before build
For every action, we agree in writing whether the agent may do it alone, must propose it, or must never touch it. Money out, client-facing words, and anything a regulator reads sit above the line by default.
A gate is a queue, not a pause
Held actions arrive as one clean list with the draft, the source records and the reason it was flagged. Approve, edit, or reject. No hunting through an inbox to work out what the agent wanted.
Every decision leaves a record
Who approved what, when, and on which version of the playbook. That trail is the difference between an agent you can defend in an audit and one you quietly stop using.
What happens to the ten percent that goes wrong.
Anyone promising an agent that never fails is selling a demo. Before we build, we write the failure list with you and design the response to each one. Here is the standard set.
The source system is down or slow
Work queues instead of disappearing. Retries back off, the run is marked incomplete rather than complete, and someone is told before the next scheduled window closes.
The agent is not confident enough to act
Low confidence is a routing decision, not a guess. The case goes to the human queue with what it did understand, and the reason it stopped.
The model returns something malformed or invented
Output is validated against a schema and checked against the source record before any write. Anything that fails validation never reaches your system.
The same job runs twice
Every write is keyed to the record it belongs to, so a repeat run updates instead of duplicating. No second invoice, no second email.
A rule changes and nobody tells the agent
Playbooks are versioned and reviewed on a cadence. Overrides are counted, and a rising override rate is treated as a defect to fix, not noise to ignore.
It needs to stop, right now
Every agent has an off switch that a non-technical manager can reach, and the process it took over is documented well enough to run manually the same day.
None of these is one clever agent.
Each is a small set of narrow agents with defined handoffs.
Your industry has a system like these. It just has not been drawn yet.
The best first agent is usually the one nobody nominates. Tell Max what your business does and we will find it together, at no cost, in a blueprint you own.