AI agents
Agents that do the task and stop where a person should decide
We build agents that take a task, such as preparing a quote or sorting a support request, work through it across your tools and bring the result to a person for approval before anything leaves the company.
Agent run · quote request
- EmailRequest readItems and quantities recognised
- CRMCustomer and terms looked up
- ERPStock and prices checked
- AgentQuote drafted
- PersonSalesperson approves and sends
What it is
A model with tools, a task and limits
An automation follows the same steps every time. An agent is given a goal and a set of tools, and decides which steps to take for the case in front of it. That makes it useful where cases differ, and it is also the reason for limits: each tool has defined permissions, each run is logged step by step, and actions with consequences wait for a person.
- Each tool is read-only unless there is a reason for it to write
- Approval points before sending, posting, paying or deleting
- A full log: what it read, what it decided and why
- A fixed workflow is used wherever the steps never vary
Agent log · sample data
- Read requestThree items, delivery by FridayWaiting for approval
- Looked up customerExisting account, agreed terms appliedWaiting for approval
- Checked stockOne item short, alternative suggestedWaiting for approval
- Drafted quoteSent to salesperson for approvalWaiting for approval
What we deliver
An agent you can supervise
A defined task
One job with a clear start, a clear end and a written description of what a good result looks like.
Tools
Connections to email, CRM, ERP and documents, each exposing only the actions the task needs.
Guardrails
Limits on what it may change, how much it may do in one run and when it must stop and ask.
Approval queue
A place where a person sees the proposed action with its reasoning and approves, edits or rejects it.
Run log
Every step of every run recorded, so a decision can be traced afterwards.
Evaluation
A set of past cases the agent is tested against whenever its instructions or the model change.
How a project runs
Narrow first, wider when it has earned it
- 01
Choose one task
We pick a task that repeats, has clear inputs and is checked by a person today anyway.
- 02
Build with read-only tools
The agent first only reads and proposes. People compare its proposals with what they would have done.
- 03
Add actions behind approval
It is allowed to act once a person approves each action. Edits and rejections are recorded and used to improve it.
- 04
Review and widen
After a period of use we look at the log together and decide which low-risk actions no longer need approval.
What we build with
- Models
- Claude models, GPT models
- Orchestration
- n8n, Custom services, Model Context Protocol servers
- Works across
- Email, CRM, ERP, Helpdesk, Shared drives, Team chat
We start from what you already have. The list shows what we reach for when something is missing, not tools you must adopt.
Is an agent the right tool?
A good fit when…
- The task repeats but each case is a little different
- It needs information from several systems
- A person reviews the result today and would keep doing so
- The systems involved have APIs
Probably not for you when…
- The steps never vary: a plain workflow is cheaper and more predictable
- Nobody would have time to review what it proposes
- A mistake could not be undone and no approval step is acceptable
- The task is rare or its rules are not yet agreed
Use cases
Where agents are used
Tasks with a draft that a person approves.
Every enquiry in the CRM, without retyping
Enquiries from forms, email and chat become contacts and deals in your CRM, with the right person notified.
Open use case Finance and adminAI integrationsBusiness automationSupplier invoices read, matched and ready to post
Incoming invoices are read, matched to purchase orders and sent for approval before they reach the books.
Open use case Finance and adminAI integrationsBusiness automationData out of documents, into your systems
Contracts, forms, delivery notes and statements are read and their data placed where it is needed.
Open use case Finance and adminBusiness automationAI integrationsThe monthly report, assembled for you
Figures from accounting, sales and operations pulled into one report on a schedule, with a written summary.
Open use case Customer supportAI integrationsBusiness automationA support inbox that sorts itself
Incoming messages are sorted by topic and urgency, routed to the right person and given a draft reply.
Open use case Customer supportAI integrationsCustom softwareAn assistant that answers from your documents
Staff or customers ask in plain language and get an answer drawn from your documents, with the source shown.
Open use case Customer supportBusiness automationAI integrationsCustomers check their order without writing to you
Order and delivery status available on your site or by reply, pulled live from the shop and the courier.
Open use caseQuestions about AI agents
What is the difference between an agent and an automation?
An automation runs fixed steps in a fixed order. An agent chooses its steps for each case, using the tools it has been given. We use an automation wherever the steps are always the same, and an agent only for the part that needs judgement.
Can an agent act without anyone approving?
Only where you decide it may, and only for actions that are low-risk and easy to undo, such as adding a note or a label. Anything sent to a customer, posted to the books or paid waits for a person.
What stops it doing something it should not?
Permissions set in the systems themselves, not only in its instructions. If the agent’s account cannot delete a record, no instruction can make it. On top of that come limits per run and the approval queue.
How do we know why it did something?
Every run has a log of what it read, which tool it called and the reason it gave. The log is kept and can be searched by case.
What does our data pass through?
The model provider’s API, under business terms that state the data is not used for training, and the systems the agent is connected to. We list the path for each task before it is built.
What happens when the model is updated?
We keep a set of past cases with known good results and run the agent against them before switching to a new model version, so a change in behaviour is seen before your team sees it.
Show us one process. We will show you what it could look like.
The analysis is free, takes about an hour of your time, and ends with a written summary you keep.
Prefer to read first? See examples