Not RPA. Not a generic AI agent. See why →

The regulated back office,
running itself.

North Stack reads the email your team already gets, runs the routine jobs against Acturis the way a person does, and escalates the calls that matter — every decision cited to your own rules and written to an audit ledger you can hand to the FCA.

Live

A renewal lands, the agent classifies it, scores its confidence, and starts the run itself.

Runs inside the stack you already use

ActurisSystem of record
Microsoft 365Mail · Word · Excel
Insurer portalsQuote & bind
The MIDMotor database
FCA-readyAudit & Consumer Duty
Two people have to say yes

The team gets hours back. The principal gets the risk removed.

For the operations leader

Hours back, not another tool to run.

It clears the routine inbox work end-to-end: renewals compared, letters drafted, claims records found, risk re-keyed into Acturis. Your handlers stop doing the endless part and keep the judgement calls. In the demo that comes to 214 handler-hours a week, about 5.7 FTE of capacity handed back.

For the MD / principal

Auditable by design, not a black box.

Nothing improvises near client records. Every action runs a mapped process, stops at a human gate when it matters, cites the rule it used, and lands in a tamper-evident ledger. Under FCA and Consumer Duty, that is a control you’re adding rather than a risk you’re taking on.

Why it’s safe in front of regulated work

Three things a general-purpose AI agent can’t give you.

01

Predictable, not improvised

It runs a process you mapped, not a prompt it improvises. You know exactly what it will do before it ever runs, and every run is measured against the same map.

Mapped workflows · confidence-scored triage
02

Auditable, not a black box

Every step is a tamper-evident, hash-chained ledger entry, and every decision cites the exact rule and row it used. Proof you can hand to a regulator, not a chat transcript you have to trust.

Hash-chained ledger · cited decisions
03

No integration project

It drives Acturis, M365 and the portals through the same screens a person uses today. No API to wait for, no rip-and-replace, no year-long rollout.

Runs Acturis as-is · guardrails in the runner
How it works

Six steps, start to finish.

  1. 01
    Email comes in

    Everything that hits the shared inbox lands in one place.

  2. 02
    It gets read and sorted

    The agent works out what it is, which workflow it belongs to, and how sure it is.

  3. 03
    The routine runs itself

    Routine jobs run end-to-end through a process you mapped, and you can watch where each one is.

  4. 04
    It reads from your own sheets

    Pricing, rates, appetite and watchlists are versioned sheets it looks figures up in. It never invents them.

  5. 05
    The calls go to a person

    Anything sensitive or uncertain stops at a human checkpoint, with the proposed action laid out in full.

  6. 06
    Everything is on the record

    Every step, whether agent, human or system, is a tamper-evident ledger entry.

Reads the inbox like your team

It classifies, maps, and scores its own confidence

Every message is triaged for what it is, which workflow it belongs to, and how confident the agent is, with the key fields pulled out. Routine, high-confidence work runs; anything ambiguous is left for a person. It never improvises outside a mapped process.

Shared inbox triaging
Renewal pack — Holmfirth LogisticsAviva — Renewals reading
MTA — add vehicle · PN-4471Broker — J. Whitaker
Claim acknowledgementPennine Textiles
Re: Renewal pack — HolmfirthAviva — Renewals
Predictable by design

A process you can see, not a prompt you have to trust

Each job runs through a workflow you mapped: the exact steps, the systems each touches, where the human gates and exception lanes sit. That makes it knowable before it runs, which is the core difference from a general-purpose agent.

WF-E01 · one mapped run
TriggerRenewal pack received
ClassifyMap to WF-E01 · score confidence
Human gateBroker reviewif low conf → senior review
WriteProduce report · update Acturis
LedgerHash-chained entry written
Proof, not a transcript

Every decision cites the exact rule and version

When a run decides, it cites the sheet, row and version it used, alongside a field-by-field read-back against the system of record. You can trace a decision back to the row behind it instead of believing a conversation.

DecisionIssue YoY discrepancy report citing sources
Cited from your own sheets
Sheet Commission & panel ratesRow Holmfirth Logistics / FleetVersion v12 · 2026-06
FieldValueSystem of record
Vehicles24
Annual premium£48,210
Cover basisCommercial fleet
The honest comparison

Not RPA. Not a generic AI agent.

RPA is brittle and can’t read an inbox. A generic AI agent improvises and leaves you a transcript, not a defensible record. North Stack is the regulated middle: the flexibility of an agent with the control of a mapped, audited process.

CapabilityNorth StackGeneric AI agentRPA (UiPath / Blue Prism)
Handles a free-text shared inboxYesNo
Knowable before it runsNoYes
Tamper-evident audit trailTranscript onlyLogs actions
Cites the rule behind each decisionNoNo
Runs Acturis with no APIFragileBrittle replay
Human approval gates on writesNoLimited
The walkthrough

Ninety seconds, on the live product.

Request access

See it on your own processes.

Tell us a little about the work you’d want it to take off the desk, and we’ll set you up with the live North Stack demo environment and a founder-led walkthrough.

Already invited? Log in to the demo →

The request form is unavailable right now. Email i@choy.in and we'll get you in.