Your agents run on eight layers you never have to build, buy or keep alive.

When your work needs one of them to behave differently, we change the layer. Below is each one, and what it saves you building yourself.

01

Data integration

Nobody runs an export again.

OAuth connectors

Drive import

CRM / ERP / BI

ETL sync

Otherwise you build

a connection into each system, then keep every one alive when a supplier renames a field.

02

Platform services

A busy morning does not lose a document.

Database

Queue

Object storage

Sandbox

Otherwise you build

the queue and the storage, and the job of knowing where every document went.

03

Retrieval

Every answer arrives with the page it came from.

Semantic search

Hybrid RAG

Vector / BM25

Page-level citations

Otherwise you build

search that works on a scan with no text layer, and keep it current when somebody reorganises a drive.

04

Model

You can see which model ran which step.

LLM router

Frontier models

Open source on-prem

Skills / prompts

Otherwise you build

the routing and somewhere to read it back, then rebuild both the week a better model arrives.

05

Orchestration

A job finishes, or it stops at the person who decides.

Agent runtime

Orchestrator

Retries

Artifact lifecycle

Otherwise you build

the running order and the retries, and check every morning for the runs that stopped in the night.

06

Governance

A person opens only what their job allows.

RBAC / auth

Tenant isolation

Activity logs

Guardrails

Otherwise you build

permissions twice, for your people and for the machines acting on their behalf, and answer for both.

07

Experience

One screen holds everything waiting on you.

Web / admin

Citations / artifacts

Streaming

Approval queue

Otherwise you build

that screen, and the record of who approved what and when.

08

Infrastructure

It runs in your account, and keeping it up is ours.

Private / hybrid / cloud

Docker / K8s

Secrets management

OTEL observability

Otherwise you build

the hosting and the release routine, and own the upgrade every time a piece underneath it moves.

What we build

We build the same four things into every install.

BG Image

ONE WEEK OF THIS JOB, OBSERVED

Request arrives in the shared inbox

4 min

Priced against the exceptions one person knows

1h 40m

Waits for one approval

3 days

Quoted, then chased on no schedule

25 min

Won or lost, logged or not at all

10 min

The waiting is the largest step and nobody had timed it. Four customers are priced as exceptions, and none of them are in the SOP.

BG Image

ONE WEEK OF THIS JOB, OBSERVED

Request arrives in the shared inbox

4 min

Priced against the exceptions one person knows

1h 40m

Waits for one approval

3 days

Quoted, then chased on no schedule

25 min

Won or lost, logged or not at all

10 min

The waiting is the largest step and nobody had timed it. Four customers are priced as exceptions, and none of them are in the SOP.

The data layer

One layer that reads every system the work touches and keeps it current. Everything else runs on top of it, and it is most of the engineering.

BG Image

Reading your systems

One layer, four sources

ERP · nobody has the password

Shared drive · 2,410 files

WhatsApp · where the price is agreed

Rate card.xlsx · merged cells

One layer, current every night

Reading

0%

0%

BG Image

Reading your systems

One layer, four sources

ERP · nobody has the password

Shared drive · 2,410 files

WhatsApp · where the price is agreed

Rate card.xlsx · merged cells

One layer, current every night

Reading

0%

0%

The workflow

Multi-step work run end to end: the sequencing, the retries, the handoffs between an agent step and a human one, and what happens when step three fails while nobody is watching. Every run is logged, so you can open any job and see what it did and on what basis.

BG Image

Run log

The last three runs

The investor report

Done

The counterparty file

Waiting on a person

The payment run

Running

BG Image

Run log

The last three runs

The investor report

Done

The counterparty file

Waiting on a person

The payment run

Running

The trigger

Every workflow starts on a trigger or a schedule. A document arrives, an invoice ages past its terms, Monday morning comes. The first you hear about it is when the work is done or when an exception needs you.

BG Image
Card Image
BG Image
Card Image

The delivery

The output has to land in the systems your team already opens. The record updated, the pack filed in the right folder, the report in the inbox before the meeting. We build against whatever you run today, including the system nobody has got round to replacing.

The install

We run every install in the same five stages.

Only someone who has done this knows to mention merged cells and a password nobody has. That is stage 01, and it is where most of the engineering goes.

Only someone who has done this knows to mention merged cells and a password nobody has. That is stage 01, and it is where most of the engineering goes.

01

Getting the data out of where it actually lives

Getting the data out of where it actually lives

An ERP nobody has the password to. A shared drive where nobody agrees which folder is current. A WhatsApp thread where the real price gets agreed. A rate card in Excel with merged cells, and scanned packs with no text layer. Getting clean, current data out of all of that is the first thing we build, and it is where most of the work goes.

An ERP nobody has the password to. A shared drive where nobody agrees which folder is current. A WhatsApp thread where the real price gets agreed. A rate card in Excel with merged cells, and scanned packs with no text layer. Getting clean, current data out of all of that is the first thing we build, and it is where most of the work goes.

02

Writing down how your people actually decide

Writing down how your people actually decide

Not the SOP. The SOP describes the clean case. We sit with the person who does the job and write down the judgment: which discount they approve without asking, which supplier gets chased on a schedule and which one gets a phone call, and the handful of customers who are an exception to all of it. You read it and correct it before we build anything.

03

Building it: what runs, on what trigger, into which system

Building it: what runs, on what trigger, into which system

Every workflow is written down as three plain answers before it exists. What runs. What sets it off, a schedule or an event. Where the output lands, which is always a system your team already opens. Then we build it and run it against real work from last month, so you can compare what it did with what your team did.

Every workflow is written down as three plain answers before it exists. What runs. What sets it off, a schedule or an event. Where the output lands, which is always a system your team already opens. Then we build it and run it against real work from last month, so you can compare what it did with what your team did.

04

The exceptions: what it does when it cannot be sure

The exceptions: what it does when it cannot be sure

Every workflow has a line it will not cross on its own. A confidence threshold, an amount, a document that does not match the rest of the pack, a customer already flagged as an exception. When it reaches that line it stops, completes the part it can, and tells a named person exactly what it needs. You set where each line sits.

Every workflow has a line it will not cross on its own. A confidence threshold, an amount, a document that does not match the rest of the pack, a customer already flagged as an exception. When it reaches that line it stops, completes the part it can, and tells a named person exactly what it needs. You set where each line sits.

05

Keeping it running: what happens when it breaks at 2am

Keeping it running: what happens when it breaks at 2am

Something will break. An integration changes, a supplier switches format, a volume spike hits a rate limit. Every workflow is monitored and the alert reaches us first. You get told what failed, what happened to the work that was in flight, and what changed so the same thing does not happen twice. After every change it is checked again against jobs you already finished.

Who builds it with you

An AI strategist, an engineering team, and the people who make the calls today.

The AI strategist

Maps the work, writes the runbook, and decides what a machine should never do alone

The engineering team

Builds the data layer first, then the workflows that sit on top of it

Your people

They make the calls today, and they are the only ones who know the exceptions

The person who signs

They set what the agents are allowed to do alone, and nothing is released until they open it and sign

An abstract field of characters, the texture of a process running

We meet you where you are, Dubai or Abu Dhabi

The boundary

What we will not do.

We do not sell software licences

No seat bill, no subscription, no login you pay for every month. You own what we build.

We do not hand you a document and leave

The analysis exists to be built. If you stop after it you take the spec elsewhere, and we say so before you buy it.

We do not run a pilot that never ships

Phase 02 ends in production, on a trigger or a schedule, inside the systems your team already opens.