Agen8

Operations, rebuilt

We run the operations you should not be paying people to do.

Document handling. Invoice matching. Chasing what is missing. Answering the same status question forty times a day. We rebuild those functions to run on supervised AI workers — inside your systems, under your credentials, with a person approving anything that writes, sends or spends.

We start with two weeks of measurement. If the numbers do not justify going further, we say so and stop.

Where it runs
Your accounts, your region
First workflow live in
6–12 weeks
What you get monthly
An auditable savings statement

How it works

Four stages. Stop after any one.

Every stage has a fixed scope and a fixed price. Nothing about this requires you to commit to the whole thing on day one.

01

Measure the baseline

Two weeks, fixed fee. We automate nothing. We map the function with your people and put real numbers against it, then split the work three ways — and tell you honestly which parts are which.

  • Volume, minutes per unit, loaded cost, rework rate, cycle time
  • Rules and retrieval — automatable now, with confidence
  • Judgement — stays human, and we say so plainly
  • Broken process — should not exist; usually the cheapest win
02

Install the spine

Before a single workflow goes live, we install the layer that makes it safe to run and possible to measure. This is not optional and it does not come second.

  • Metering on every run, attributed to the workflow
  • An immutable audit log built to hand to an auditor
  • Approval gates on anything that writes, sends, spends or deletes
  • Access scoped to a role, never to a person's login
03

Rebuild the workflows

Four to ten weeks per bundle. We write the operating procedure — precise enough for a machine, readable enough for your team to argue with — and then prove it before it touches anything.

  • An exception path: when it is unsure, it stops and asks a named person
  • Shadow mode against live volume, while your team still does the work
  • Cut over one segment at a time, never all at once
  • You see the accuracy before it is allowed to act
04

Operate

The meter keeps running. Procedures get tuned as the business changes, and workflows get added when you want them.

  • The monthly savings statement, generated from the meter
  • Drift detection — quality graded against your own historical cases
  • Model swaps when a cheaper one is proven equivalent
  • Exit assistance, included, if you ever want to run it yourselves

The proof

Every month, one page.

Most AI projects cannot say what they achieved. That is why they die at renewal. Ours arrives as a statement you can audit against your own bill — because the run costs are on your account, not ours.

Document intake & customs prep

March 2026 Specimen
Agreed baseline — manual cost of this function
€52,000
Shipments handled by the workflow
3,847
— of which escalated to a person
just under 1 in 5
Remaining team cost
€20,800
Model and tool usage (billed to you at cost)
€1,940
Agen8 platform & support
€3,000
Net saving this month
€26,260

A specimen, not a client result — we are early and have no outcomes to report yet. The figures follow the illustrative model below. Every line here is generated from the meter rather than assembled by an account manager, and the usage line can be checked against your own provider invoice.

What gets installed

The spine.

The same control layer goes into every engagement. It is the difference between a consultancy and a body shop, and it is what makes the statement above possible at all.

Metering

Every run records what it cost and what it handled, attributed to the workflow. Not "we spent €4,000 on AI last month" but "document intake cost €312 across 1,840 shipments — 17 cents each, against a €2.10 manual baseline."

An audit log that survives an auditor

Every action: what asked, what ran, what it touched, what came back. Immutable and exportable. It exists before you need it, because the meeting where you need it is not the one to start building it.

Approval gates

Anything that writes, sends, spends or deletes is proposed rather than executed, until you promote it. Per workflow, per action, reversible, logged. This is the answer to "we tried AI and it made things up".

Drift detection

Output is graded continuously against held-out samples from your own history. When a model, a prompt or an upstream system changes, we hear it from the harness rather than from you.

Where it pays

High volume. Clear rules. A role you cannot fill.

We start with logistics because the arithmetic is most obvious there. The same workflows exist in every operations-heavy business, which is why the list below is not a limit.

Illustrative — a freight forwarder, 4,000 shipments a month

~10 people on document handling today, at roughly €52,000 a month
~60% extract, validate, classify, file, chase — rules and retrieval
4 people remain, on exceptions and judgement
≈€26k net saving per month, after run cost and platform

Assumes a German loaded cost of about €5,200 per person per month and roughly 25 minutes per document pack. Both get replaced with your figures in the diagnostic — we will not quote you a saving before we have measured one.

The same shape, other sectors

Manufacturing & distribution

Supplier quality documents, purchase-order matching, certificates of conformity, RFQ responses. The people who do this work are retiring faster than they are being replaced.

Professional services

Law, accounting, audit. Document review, client intake, obligation extraction, compliance monitoring. These firms understand billable-hour arithmetic instantly, which makes the conversation short.

Insurance & claims

First-notice-of-loss intake, claims triage, document validation, subrogation review. High volume, strict rules, and an audit-trail requirement the spine already satisfies.

Property & facilities

Lease abstraction, maintenance dispatch, tenant correspondence, contractor invoice checking.

Healthcare administration

Prior authorisation, claims preparation, coding support. Large and regulated — longer to sell, and the audit log does most of the work.

Your sector, probably

The pattern to look for: high volume, clear rules, documents or tickets as the unit of work, an exception rate someone complains about, and a role you have been trying to fill for months.

The workflows

What we rebuild.

Each of these is one function, with a written procedure, approval gates and a meter. Three are being built first — the three cost lines almost every operations-heavy business recognises immediately.

First to ship in build

Invoice matching

Finance

Three-way matching, exception handling, and chasing what does not reconcile. The most universally recognised cost line there is.

Ticket triage

Support

Classify, route, draft, and resolve the tier-1 volume that is pure retrieval — while everything that needs judgement reaches a person faster.

Lead research

Revenue

Enrichment, qualification and CRM hygiene, so the pipeline reflects reality without anyone spending a morning on it.

Next in build

  • Collections & dunningChasing what is owed, on a schedule, in your tone
  • Supplier onboardingIntake, validation and publication of new supplier data
  • Returns & refundsPolicy applied consistently, edge cases escalated
  • Quote & proposal draftingRates, surcharges and terms assembled for review
  • Month-end close prepReconciliation, accruals and what is still outstanding
  • Candidate screeningStructured screening against your criteria, and scheduling
  • Contract reviewObligation extraction and deviation from your standard terms
  • Recurring reportingThe reports rebuilt by hand every week, plus anomaly watch

Can you actually build

Three complete systems, open to inspect.

We have no client logos yet — we are early, and we would rather say so than imply otherwise. What we can show you is the engineering standard your spine gets built to. These are real, running systems you can log into, and they are the estate our workflows are built and demonstrated against.

Orrery

Sales system

Where the revenue workflows run.

A CRM built around the two questions the category usually answers badly: what is actually happening on this account, and which of these deals is quietly dying.

API tests
1,216
Endpoints
111
Stack
Next.js 16 · FastAPI

Bramble

Commerce system

Where the supplier and catalogue workflows run.

A storefront for independent makers. Guest checkout by default, stock held from the moment something enters a basket, and money as integer minor units end to end.

Tests
524 API · 1,040 web
Endpoints
91
Stack
Next.js 16 · FastAPI

Cairn

Work system

Where the delivery and reporting workflows run.

Offline-first across web, iOS and macOS. Every client holds a replica and reconciles through a change feed — two devices editing offline have to converge, and be provably right about it.

Sync ops
51 of 51 bound
Form factors
Web · iOS · macOS
Stack
Swift 6 · GRDB · FastAPI

Who you are hiring

We build the software. That is why we can change it.

Most firms in this space configure other people's tools and are limited to what those tools expose. We write the applications, which is a different job and a different set of options — and it is the reason the promises on this page are keepable.

When a workflow needs an approval screen your staff will actually use, we build that screen. When an agent has to reach into an ERP with fourteen years of history behind it, we write that integration rather than waiting for a connector. When you want cost per run broken out by department, that is a dashboard we build, not a CSV we export.

Most of what we install is not a model. It is the application around the model — the procedures, the permissions, the interfaces, the meter — and that is ordinary software engineering, done carefully.

Whole systems, not wrappers

Backends that carry real volume and real business logic, the APIs that let them talk to what you already run, and the database design underneath. The three systems above are ours end to end — schema, services, interface, deployment and observability.

Approval screens built for your desk

The human-in-the-loop step is where most automation is won or lost. Because we build the interface rather than inherit it, the person reviewing an agent's proposal sees exactly what they need to decide — not a generic queue with a JSON blob in it.

AI inside the system, not beside it

The model is embedded in the backend, where it can read, reason and act as part of the workflow — with the gates and the audit log around it. A chat window bolted to the side is a different product, and a much weaker one.

Your own knowledge, retrievable

The retrieval pipelines that let an agent search your contracts, procedures and history and cite what it used — built inside your infrastructure, on your data, so nothing proprietary leaves the estate to make it work.

Backend
Python · FastAPI · Postgres · Redis
Front end
Next.js · React · TypeScript
Native, where it earns it
Swift 6 · Kotlin

A short list on purpose. These are the tools the work above was actually built with — we would rather show you three systems in a stack we know deeply than claim eight we do not.

Pricing

Published, because you should not have to ask.

Two ways to buy the same thing. Take the fixed prices, or pay us out of what we save you — whichever splits the risk the way you want it split.

Stage What it covers Price
Diagnostic Two weeks. A costed map of the function, a savings model with the assumptions visible, and a ranked plan. Stands alone — take it and do nothing else if you like. €9,500
Solutioning & build The spine plus your first workflow, fixed scope. Each additional workflow in the same bundle, €18–25k. from €35,000
Support & operations Monitoring, tuning, procedure changes, and the monthly savings statement. Scales with how many workflows you run. €1,500–6,000 / mo
Model & tool usage On your own accounts, metered per workflow. We mark it up by nothing — which is why you can audit our statement against your invoice, and why our incentive is to make it cheaper. at cost

Or split the risk differently

Direct

Default

The prices above. You own the system outright from day one and there is nothing to unwind later.

Pay from savings

Build fee drops to €10,000, then 22% of verified net saving for 24 months, capped at €150,000. Offered where the baseline is unambiguous — document counts, ticket volumes. Never against something soft like "productivity", because neither of us could defend the number.

Hybrid

A reduced build fee plus a smaller share. Where most conversations land, and usually the right answer if you like the alignment but want a ceiling on it.

Terms of engagement

Four things we hold to.

The baseline is agreed before we build

In writing, signed by whoever owns the budget, at the end of the diagnostic. A saving computed after the fact is a number somebody can argue with — and someone whose team is shrinking always will.

It runs in your accounts, not ours

Your credentials, your region, your provider bill. If you want us gone, the system stays, the procedures are documents your team can read, and exit assistance is included. We would rather earn the renewal than trap you into it.

Nothing acts unattended until you allow it

Everything that writes, sends, spends or deletes is proposed first, and promoted per action, reversibly, by you. Every workflow runs in shadow mode against your live volume before it is allowed to touch anything.

We will tell you not to buy

Sometimes the diagnostic concludes that a process should be deleted rather than automated, or that the work is mostly judgement and there is little here for us. We say so and stop. A firm whose bad news is credible is one whose good news means something.

Start here

Two weeks and a number.

Tell us which function is costing you most and roughly how much of it goes through your team every month. If there is something here, we will show you the arithmetic. If there is not, that is a short conversation and we will both have saved a fortnight.

hello@agen8.ai

We reply to everything, usually within a day. If we are not the right people, we will say so and point you somewhere better.