Skip to content

AI Agents & Operations Systems

Agents that take the admin off your team, and ask before anything important.

Single agents and full AI operations systems that read, draft, chase and update inside the tools you already use. Every action is logged, and anything involving money, publishing or advice waits for a person.

In plain terms

What this is, and when you need it

An AI agent is software that can carry out a task, not just answer a question. It reads an email, looks up the customer, drafts the reply, updates the record and asks you before it sends anything that matters.

A business needs one when the same manual work repeats every day across inboxes and systems: chasing documents, triaging requests, preparing reports, keeping records in step. Start with one agent on one job. When it earns its keep, more specialists join the same router, tools and approval rules, and it becomes an operations system.

What's included

What you can have built

  • Single agents

    One agent on one job, such as inbox triage or document chasing, done properly and measured.

  • Multi-agent systems

    A router or chief of staff with specialist agents for finance, operations, sales or support.

  • Approval gates

    Agents prepare, people decide. Money, publishing and advice always wait for a sign-off.

  • Tool connectors

    Controlled access to email, calendars, CRM, accounts and files, each with the least access it needs.

  • Business memory

    Clients, suppliers, preferences and past decisions kept where agents can look them up.

  • Monitoring and audit log

    Every step, tool call and approval recorded, with alerts when something needs a person.

How it's built

Design principles

The rules every build follows, whatever the business.

  1. Approvals before autonomy

    Each action has a level: done automatically, done and reported, or held for approval. Levels change only by a decision someone makes.

    How this works, in the Lab
  2. Memory that belongs to you

    Short, curated memory of how your business works, stored in your systems and readable by your team.

    How this works, in the Lab
  3. Tools with the least access

    Agents reach your software through one controlled tool layer. Read access where possible, write access only where needed.

    How this works, in the Lab
  4. Monitored, not left alone

    Traces of every run, tests against real tasks, and alerts when an agent is unsure or a tool fails.

    How this works, in the Lab
  5. Clear data boundaries

    What each agent can see is decided up front, written down and enforced in code, not in a prompt.

    How this works, in the Lab
  6. Cost under control

    The right model for each step, caching, and a monthly limit. Bring your own model keys or use a managed allowance.

    How this works, in the Lab
  • Approvals. Money, publishing and advice wait for a person.

  • Data control. Runs in accounts you control, never used for training.

  • UK GDPR. What is processed, where and by whom, written down.

  • Logging. Every action recorded: who did what, and when.

Where it fits

What it looks like in practice

  • Property

    A chief-of-staff agent that briefs the directors each morning and chases every open item.

  • Letting agents

    Tenant messages answered day and night, repairs triaged and contractors booked, with spend approved by you.

  • Accountancy

    Missing records chased, receipts read into the books and quarterly updates prepared for sign-off.

  • Online retail

    'Where is my order?' answered from live data, refunds drafted and held for approval.

  • Professional services

    Enquiries qualified, meetings booked and proposals drafted from your own templates.

The engagement

How it runs, start to finish

The full process
  1. Blueprint

    Pick the first job, map the tools it touches and set the approval levels. You get the agent design and a fixed price.

  2. Build

    The agent works on your real data from week one, behind approvals, with a demo every week.

  3. Launch

    Your team uses it with everything held for approval at first, then levels relax where it has earned trust.

  4. Run

    Monitoring, model updates and new agents added to the same system as the business asks for them.

Stack

What it's built with

Chosen per project. These are the tools in production today.

Models
  • Claude
  • GPT
  • Gemini
  • OpenRouter
  • LiteLLM
Frameworks
  • LangGraph
  • Claude Agent SDK
  • MCP
  • FastAPI
Data
  • Postgres
  • pgvector
  • SurrealDB
  • Redis
Infra
  • Azure Container Apps
  • Vercel
  • Fly.io
  • Docker
Channels
  • Email
  • Microsoft Teams
  • Slack
  • WhatsApp
  • Web chat

Related work

An AI chief of staff for a UK chief executive

The situationThe leadership of a growing UK property group was running the business from their inboxes. Email, calendars and business records lived in different places, and every decision meant someone gathering the facts by hand first. They wanted the admin handled for them, without handing an AI the keys to the business.

A working illustration: the same kind of system, with an invented company

Northgate Commandcommand.northgate.example
Done

A chief of staff AI for a managing director. A request is typed into the command bar: sort the overnight emails, reply to a supplier about a delay, and ask the finance director agent about Q4 cashflow. A live trace shows the router sending the work to specialist agents. Eleven emails are triaged and a call is added to today's actions. A reply to the supplier is drafted and waits for approval. The board pack is summarised. The finance director agent reports £412k projected cash at year end, 9 percent ahead of plan, and flags two invoices over 45 days. The executive approves the reply and it is sent.

Done

  • 11 emails triaged
  • 1 reply approved and sent
  • Board pack summarised
  • Finance assessment ready

Demos are illustrative. Names, companies and figures are invented.

What was built

  • A chief of staff that routes every request. It reads each request and hands it to the right director agent.

  • Six specialist director agents. Each one responsible for its own area of the business.

  • Microsoft 365 through Microsoft Graph. Email, calendars and files, plus the business records in Dataverse.

  • Approval tiers with signed approvals. Every action is checked against what the agent may do, and anything that matters needs a signed approval.

  • A morning brief and a live activity trace. The day laid out before it starts, and every step the agents take written to an audit log.

The result

The leadership get a chief of staff that prepares, drafts and chases across the business, while every decision that matters still waits for a person. It runs entirely inside their own Microsoft cloud, so their data never leaves infrastructure they own.

Read the AI chief of staff case study

FAQ

AI agents: common questions

How much does an AI agent cost?

A focused agent on one job is usually a small four-figure build. A multi-agent operations system with several specialists, memory and many connected tools runs from the mid four figures into five. The Blueprint gives you a fixed price before you commit.

How long does it take?

A single agent is typically live in two to four weeks. A multi-agent system takes four to eight weeks or more, delivered in stages so the first agent is working before the last is built.

Will it send anything without us checking?

Only what you choose. Every action has an approval level, and anything involving money, publishing or advice waits for a person. You can see every pending action and every decision in the log.

What happens when the agent gets something wrong?

It is designed to fail safely. Agents say when they are unsure, risky actions are held for approval, every run is traced, and alerts reach a person. Mistakes become test cases so the same error does not come back.

Will it work with our software?

If your software has an API, sends email, or exports data, an agent can usually work with it. Microsoft 365, Google Workspace, most CRMs, Xero, QuickBooks and the main property and e-commerce platforms are routine. The Blueprint checks your exact setup.

Who owns the agents and the data?

You do. The code, prompts and configuration are yours once the build is paid for, and the system can run in your own cloud account with your own model keys.

How is this different from ChatGPT or Copilot?

Those are general assistants you ask questions inside their app. An agent system is built around your business: connected to your tools, working on a schedule, following jobs through to the end and asking before anything important. It can use the same models underneath.

Not sure where to start? Book a Blueprint

A fixed-fee architecture engagement: what to build, how it fits your tools, and a fixed price to build it. £160, credited against the build.