What we do · AI agents
Agents that do the work, inside limits you set
Agents that read the email, check the record, draft the answer and raise the order, then wait for a person where it matters. Built with limits, evaluations and a log of every action.
- Price
- From $2,500 one-time; most builds $5,000 to $15,000
- Time
- 3 to 8 weeks
- For
- Teams with a multi-step task that needs reading or judgement, a clear goal and a person willing to approve the outcome.
- You own
- The agent, its prompts and its code
In short
An AI agent reads an input, decides what to do within written limits and acts across your tools: drafting the reply, updating the record, raising the order. We build agents with a person approving anything that matters, guardrails on what they may touch, evaluations before go-live and monitoring after it. Builds start at $2,500, one-time; most land between $5,000 and $15,000.
Is this you?
If three of these sound like your week, ai agents is the place to start.
- A task needs reading or judgement, so it was never automated
- Someone triages an inbox of requests, enquiries or documents all day
- You tried an AI pilot and nobody trusted the output
- The same multi-step lookup happens across three or four tools
- You want AI to act, but only within limits you can audit
What this is not
So the scope is clear before the call.
- Agents that act on customers, money or records with no limit and no person behind them.
- Tasks that run the same way every time. A plain workflow is cheaper and easier to trust.
- Phone and voice work, which has its own page: see voice agents.
Most often for
What we have built
AI agents in practice.
- Real estate
An agent fleet for property operations
Twelve agents in build with a control plane, decision queue and policy engine, on top of six live automations.
Read the story - Distribution
Emailed RFQs read by AI
Enquiries read, parts matched against the SAP item master, and quotations prepared for a person to send.
Read the story - Sales
Leads qualified with a person in the loop
Capture, enrichment and AI qualification, with the first reply approved before it goes out.
Read the story - Tool
How much should an agent be allowed to do?
Our check for scoring an agent’s risk before it goes live.
Read the story
AI agents
Limits first, prompt second.
Bring the task you would hand to a new hire. We will show you what an agent could do alone, what it should ask about, and what it should never touch.
Score your agent idea
How it runs
AI agents, step by step.
- 1Week 1
The task is written down step by step: inputs, decisions, actions, and the limits on each. We agree what the agent may do alone and what needs approval.
- 2Weeks 2 to 3
The agent is built against real examples from your history. An evaluation set measures how often it gets each step right before anyone relies on it.
- 3Weeks 4 to 5
Shadow mode. The agent proposes, a person decides, and every disagreement is logged and reviewed.
- 4Go-live
Approval is relaxed only where the numbers support it. Monitoring, cost tracking and alerts run from day one of hypercare.
What you get
- The agent, running in your accounts on the model and tools agreed in the blueprint
- A written policy: what it may do alone, what needs approval, what it may never do
- An evaluation set and the score it reached before go-live
- An approval queue and a full log of every decision and action
- Monitoring for errors, cost and drift, with alerts to a named person
How we know it is done
- The agent meets the agreed evaluation score on held-back real cases
- Every action it can take is listed in the policy and nothing outside it is possible
- A person can pause it, review its log and reverse its actions
What you own
- The agent, its prompts and its code
- The policy and the evaluation set
- The decision log
- The model and platform accounts, in your name
of corporate generative-AI pilots show no measurable P&L impact
Source: MIT NANDA, 2025Built on
The tools you already run, where we can.
Proof
Customer stories.
“An agent is only as safe as the limits written around it. We write the limits first and the prompt second.”
How it starts
Every project starts with an audit.
One to two weeks with the people who do the work. A costed blueprint in order of payback. The $500 is credited in full to the build.
- Audit1 to 2 weeks. $500, credited to your build.
- BlueprintIncluded in the audit’s second week. Included with the Audit; quoted separately for multi-phase programmes.
- Build2 to 8 weeks. From $2,500 one-time; most builds $5,000 to $15,000.
- Run60 days of hypercare, then monthly. Hypercare included; care plans from $50/month.
Questions
What people ask about ai agents.
When should we use an agent rather than a workflow?
Only when the work has several steps, needs reading or judgement, and has a clear goal. If a step runs the same way every time, a fixed workflow is cheaper and easier to trust. Most good agents sit on top of plain automations.
How do you stop an agent doing something it should not?
It can only call the actions we give it, with the permissions we give it. Anything above an agreed limit, such as a value, a customer-facing message or a deletion, goes to a person for approval. Every action is logged.
How do we know it is working?
Before go-live it is scored against real past cases. After go-live, monitoring tracks errors, approval rates, overrides and cost per task, and a sample of decisions is reviewed each week during hypercare.
Which models do you use?
Claude and OpenAI through their APIs, chosen per task. Model costs pass through at cost and are itemised. The agent is written so the model can be changed later.
Who owns the code, the workflows and the accounts?
You do, from the first sprint. Every tool, automation and database lives in accounts in your name. Code sits in your repository. Documentation is part of the deliverable. If you stop working with us, everything keeps running.
Reading
- Systems
How much should an AI agent be allowed to do?
An agent with an unscoped token dropped a production database and its backups in nine seconds. The permission model we use for every agent: read widely, write narrowly, spend with a cap, and escalate when unsure.
- Systems
Six months of AI agents in production: what broke
Not the model. Field renames, a confidence score that drifted, a supplier who changed invoice format, and a Slack channel nobody read. What failed across the agent systems we run for clients, and what we changed.
- Systems
Invoice intake with an agent: the first thirty days
What the first month of an invoice intake agent looks like from the inside: how the queue behaves, which invoices escalate, what the finance team changes, and the three settings that decide whether it survives month two.
Also what we do
- What we do
Automate
Teams re-typing data between tools, chasing status by hand, or waiting a day for a report.
- What we do
Modernise
Companies running a $5M to $100M operation on spreadsheets, a legacy ERP or a tool nobody dares change.
- What we do
Advise
Founders and COOs making technology decisions without anyone senior to check them.
- What we do
Extend
Teams that already run a platform but are out of capacity, or need a skill they do not have in-house.
- What we do
Harden
Founders who shipped something real with an AI builder and need it to survive real customers, investors or an audit.
- What we do
Integrations
Teams whose tools each hold part of the truth and who re-key data between them every day.
- What we do
Internal tools and portals
Teams running the business from spreadsheets and inboxes, or sending customers and suppliers status by email.
- What we do
Dashboards and reporting
Teams that export, clean and paste reports every week, or argue about whose number is right.
Related: Automate, Integrations, Audit
AI agents
Talk to the people who would build it.
A discovery call. Bring the work your team still does by hand. You leave with a yes or no on fit and the first thing we would build.
Book a discovery call