Use cases

What we build, described as patterns. The first three are systems we have designed and run ourselves. The others are example builds showing how the same parts fit common needs for businesses and agencies.

Systems we have built

Designed, built and in daily use by our own team.

A multi-agent analytics assistant over your data warehouse

Ask a business question in plain English and get an answer grounded in your warehouse, with the query that produced it, and a dashboard or report when the question deserves one.

What we build

  • A coordinating agent that breaks a question into steps and delegates to specialised sub-agents for SQL, analysis and visualisation.
  • A semantic layer describing your tables, metrics and business definitions, so the agent uses the same numbers your team does.
  • Generated dashboards and written reports that cite the queries behind every figure.
  • Read-only warehouse access, query cost limits and a full log of what was asked and run.

How we built it

We have built a system of this shape for day-to-day business intelligence: a lead agent receives the question, a data agent writes and checks the SQL against documented models, an analyst agent interprets the result, and a reporting agent turns it into a chart or a short write-up. Analysts review anything that goes to leadership; routine questions no longer wait for a free slot in their week.

MCP connectors for analytics and search-performance data

Let AI assistants query web analytics and search-performance data directly, so marketers ask "which landing pages lost organic clicks last month?" instead of building the report by hand.

What we build

  • Model Context Protocol servers that expose analytics and search-performance reporting as well-described tools.
  • Sensible defaults for date ranges, dimensions and metrics, so plain-English questions map to correct queries.
  • Read-only, per-account credentials and limits that keep usage within API quotas.
  • Documentation that lets your team add new reports without touching the core.

How we built it

We have built MCP connectors of this kind for web analytics and search performance platforms. Once connected, an assistant can compare periods, break traffic down by channel or page, and combine the two sources in one answer, while every request stays read-only and traceable to the person who asked it.

An agent platform with long-term memory and sub-agents

A working environment where agents remember context between sessions, delegate to specialists and use your tools, built for daily analytics and engineering work rather than for demos.

What we build

  • Long-term memory that stores decisions, preferences and project state, and retrieves only what is relevant to the task at hand.
  • Specialised sub-agents for areas such as data, documentation, communication drafts or code, each with its own tools and permissions.
  • Tool orchestration across command-line tools, APIs and internal systems, with approval steps for anything irreversible.
  • Session logs and hand-over notes, so work can be resumed by a person or another agent.

How we built it

We use a platform of this kind ourselves for analytics and engineering work. A request is routed to the right specialist, which pulls in relevant memory, calls the tools it is allowed to use and reports back; drafts that would reach other people wait for explicit approval. The same architecture can be shaped around a client team's systems and rules.

Example builds for businesses

How these patterns typically apply to a company’s own operations.

A support agent for an online shop or SaaS product

An agent that answers customers on chat and email from your help content and live order or account data, and hands over to your team with full context.

What we build

  • Answers drawn from your help centre, policies and product documentation, with sources.
  • Read access to order, subscription or account status, so the agent can answer "where is my order?" properly.
  • Escalation rules for refunds, complaints, sensitive topics and low-confidence answers.
  • A weekly review of unresolved conversations that feeds back into the knowledge base.

Example build

A typical build for a small online business: the agent handles order status, delivery, returns policy and account questions; anything involving money, a complaint or an unclear request goes to a person with a summary and the data the agent already looked up. The team sees every conversation and can take over at any point.

Document and email intake for an operations team

Incoming invoices, orders and requests are read, classified, extracted and routed automatically, with a review queue for anything uncertain.

What we build

  • Classification of a shared inbox into request types, with priority and the right owner.
  • Field extraction from PDFs and attachments into your ERP or accounting system.
  • Validation rules, for example matching supplier and amounts against purchase orders.
  • A review screen where a person confirms flagged items in seconds rather than retyping them.

Example build

A typical build for a back-office team: every message to the shared address is classified within a minute, invoices have their fields extracted and checked against open orders, and only mismatches or low-confidence documents reach a person. Everything else is filed and logged.

Example builds for agencies

How these patterns typically apply to an agency serving many clients.

Client reporting automation for a marketing agency

Monthly and weekly client reports assembled automatically from ad platforms and analytics, with a written commentary drafted for the account manager to review.

What we build

  • Data pulls from each client's ad platforms and analytics into one consistent model.
  • Report templates per client tier, with agreed definitions for spend, conversions and return.
  • A drafted commentary explaining what changed and why, for the account manager to edit and approve.
  • Alerts when a client's tracking breaks or spend moves outside expected ranges.

Example build

A typical build for an agency: on the first working day of the month every client report is ready in draft, numbers are consistent across accounts, and account managers spend their time on the commentary and the conversation with the client rather than on exporting spreadsheets.

Tracking health checks across client sites

An agent that checks conversion tracking and consent behaviour across many client sites and reports problems before they show up as a gap in the data.

What we build

  • Scheduled checks that load key pages and confirm conversions, events and consent signals fire as specified.
  • Comparison of tracked conversions against platform and back-office figures to catch silent drift.
  • A plain-English summary per client of what broke, since when, and the likely cause.
  • A shared log so the agency can show clients exactly when issues were found and fixed.

Example build

A typical build for an agency managing dozens of sites: instead of discovering a broken purchase event at the end of the month, the team gets a short message the morning after a site release, with the affected pages and the change that most likely caused it.

Example builds are illustrative. They show the kind of system we would design for that situation, not work delivered to a named client.

Tell us what is taking up your team’s time.

A short message is enough. We reply with an honest view of what AI can and cannot do for that problem.

Start a conversation