Skilled people spend their day on preparation instead of judgement

An agent is only useful when it is given a real job, real access and clear limits. We build agents that way.

The system object

ROLEOrder-desk copilot
KNOWSprice list · stock · terms
CANdraft · check · prepare
CANNOTsend · discount · commit stock

a job description before it is software

  1. 01

    Role

    the job the agent supports

  2. 02

    Context

    the case in front of it

  3. 03

    Knowledge

    sources it may read

  4. 04

    Authority

    read · draft · propose · execute

  5. 05

    Proposed action

    with sources cited

  6. 06 · Human authority

    Approved action

    the person confirms

  7. 07 · Exception

    Exception

    routed to the role owner

  8. 08

    Audit

    prompt · sources · approver

Where it fits

If preparation eats the day, the role is a candidate.

A sales team researches every prospect from scratch. An operations lead answers the same internal questions by hunting through documents. A finance analyst assembles the same pack each month before any analysis begins. The knowledge exists somewhere in the business; reaching it and shaping it into a usable form is the slow part.

General-purpose chat tools help individuals, but they do not know your systems, your rules or your history, and their output cannot be trusted with anything consequential. What is needed is an assistant built for a specific role, with controlled access to the right sources and a defined boundary on what it may do alone.

  • Experienced staff doing repetitive research, drafting or summarising before real work starts01
  • Internal questions that are answered by asking a colleague rather than a system02
  • Institutional knowledge locked in documents, tickets, email threads and past proposals03
  • Teams experimenting with public chat tools using company information04
  • A backlog of work that is well understood but too voluminous for the people available05

The approach

We design the job first, then the agent

Every agent we build starts with a written description of the role it supports, the tasks it takes on, the sources it may read, the actions it may take and the point at which it must hand back to a person. Only then do we choose models, retrieval methods and integrations.

Most roles are best served by a single well-scoped agent with strong connections to data. Multi-agent designs are used where the work genuinely divides into distinct specialisms; we do not add complexity to look sophisticated.

  1. 1. Role and task definition

    We sit with the people doing the job and separate the parts that are preparation, retrieval and drafting from the parts that require their judgement.

  2. 2. Knowledge connection

    The agent is connected to the documents, systems and records it needs, with permissions that mirror what the role is already allowed to see.

  3. 3. Action boundaries

    Actions are classed as read, draft, propose or execute. Anything that changes a record or reaches a customer is proposed for approval unless you decide otherwise.

  4. 4. Evaluation and iteration

    Agents are tested against real historical cases before launch, and reviewed against a sample of live outputs afterwards, so quality is measured rather than assumed.

What we build

Agents and copilots we build

Illustrative of what a system in this area can include. Every build is designed from your workflow, not from a catalogue.

  • 01Sales research and preparation agentsAssemble account context, recent activity and relevant history into a briefing before a call, and draft follow-ups for review.
  • 02Operations assistantsAnswer procedural questions from your own documentation, check order or job status across systems, and prepare exception summaries.
  • 03Finance and reporting copilotsGather figures from finance systems, reconcile against expectations and draft commentary that an analyst then finalises.
  • 04Knowledge-connected internal assistantsA single place for staff to ask questions answered from policies, contracts, manuals and past work, with sources cited.
  • 05Human-in-the-loop workflow agentsAgents that carry a multi-step task, pause at defined approval points and continue once a person confirms.
  • 06Multi-agent workflows where justifiedSpecialised agents that hand work between them, for example research, drafting and compliance checking, with one accountable owner.

The agent proposes; the person decides

Human authority · consequential actions are approved by the role owner

An agent that acts without limits is a liability. Ours are built so that the person in the role remains the accountable decision-maker for anything consequential, and so that every proposal and action can be traced back to its sources and its approver.

  • Approval points defined per action type, visible in the interface people already use
  • Every output shows where its information came from
  • Full logs of prompts, sources, actions and approvals for later review
  • A simple way for staff to correct the agent and have the correction stick

An agent is only as useful as what it can reach

The value of an agent comes from connecting it to the systems where your work already lives. We integrate through supported APIs and existing permissions, never by giving an agent blanket access. Where a system cannot be integrated safely, the agent works alongside it and a person completes the step.

  • CRM platforms (Salesforce, HubSpot, Zoho CRM, Pipedrive)
  • ERP and finance systems (SAP, Microsoft Dynamics, Oracle NetSuite, Odoo)
  • Email, calendar and chat (Microsoft 365, Google Workspace, Slack, Teams)
  • Document libraries and knowledge bases (SharePoint, Confluence, Notion, Google Drive)
  • Support and ticketing systems (Zendesk, Freshdesk, Jira Service Management)
  • Internal databases, data warehouses and reporting tools

How an engagement runs

From role definition to a working agent

  1. 1. AI Leverage Call

    We discuss the roles under pressure and whether an agent, a workflow automation or a simpler tool is the right response.

  2. 2. AI Opportunity Blueprint

    Task analysis for the chosen roles, source and permission mapping, action boundaries and the agent architecture.

  3. 3. AI Agent Sprint

    One agent for one role, connected to real systems, evaluated against real cases with the people who will use it.

  4. 4. Production intelligent system

    The agent hardened for daily use, with monitoring, logging, access controls and a feedback loop.

  5. 5. Ongoing optimisation

    Review of outputs and corrections, expansion to adjacent tasks or roles, and model updates handled without disruption.

Scope grows only as value is shown. The next workflow is chosen, not assumed.

Pick the role where preparation eats the most time

That is usually where an agent pays for itself first. Tell us what that role does in a typical week and what it needs to know. Discuss an AI opportunity with us and we will map what an agent could take on and what should stay with the person.

Questions about this service

How is this different from giving staff a chat tool subscription?

A subscription gives individuals a general assistant with no knowledge of your systems and no controls. A built agent is connected to your data, limited to defined actions, tested against your cases and logged. The first helps a person write faster; the second changes how a role works.

Can the agent take actions in our systems, or only answer questions?

Both are possible. We class actions as read, draft, propose or execute and agree which the agent may do alone. Most businesses start with read and propose, then extend to execute for low-risk actions once trust is established.

What happens when the agent is wrong?

It will be, sometimes. That is why consequential actions require approval, outputs cite their sources, and every interaction is logged. Corrections from staff feed back into the agent's instructions and evaluation set so the same mistake is less likely next time.

Which AI models do you use?

The model is chosen for the task, considering quality, cost, speed and where the data may be processed. We design so the model can be changed without rebuilding the agent.

Do we need a multi-agent system?

Usually not at first. Most roles are well served by one focused agent with good data access. We introduce multiple agents only when the work divides into clearly separate specialisms and coordination adds real value.

Where is capability hiding in your business?

Bring one workflow that costs your team more time than it should. We will tell you honestly whether AI, automation or a better system belongs there.