Named agents that
own a function.
Not a chatbot on a website. An agent is given a job, a set of systems, a declared autonomy level and a set of hard limits — then it does that job every day, and every decision it makes is logged with the inputs that produced it.
A job description,
written in software.
Every agent we deploy has the same six attributes defined before it goes live. If any one of them cannot be answered, the agent is not ready.
| Attribute | What it means | Where it is fixed |
|---|---|---|
| Name and function | The single job it owns — not a general assistant | Schedule A |
| Autonomy level | Supervised (acts, then audited) or draft-only (proposes, human sends) | Schedule A + enforced in code |
| System access | The exact systems and permission scope it holds | Access register |
| Hard limits | Actions it can never take, whatever the model outputs | Platform guards |
| Accuracy benchmark | The measured standard it must hold, sampled weekly | Acceptance criteria |
| Escalation path | Who receives an exception, and how fast | Escalation ladder |
Twelve that already
have a template.
Not a fixed catalogue — a starting point. A template removes two to three weeks from the build phase because the process map, guards and test set already exist.
Inventory & stock agent
Syncs stock across channels, flags mismatches, prevents oversell, raises reorder alerts against your thresholds.
Pricing agent
Adjusts prices within a hard margin floor, monitors competitor movement, and refuses any change that breaches the floor.
Listing & catalogue agent
Drafts listing content, fills attributes, flags policy risks. Draft-only by default because listing errors reach customers.
Refund & returns agent
Assesses returns against policy, prepares the decision, and escalates anything above the approval threshold to a named human.
Support triage agent
Classifies incoming tickets, drafts replies, routes by intent and priority, and hands over to a live agent on defined triggers.
Scheduling agent
Books, reschedules and confirms appointments across calendars, handles no-show follow-up and slot optimisation.
Collections agent
Runs the reminder ladder, records promises to pay, escalates ageing accounts. Never mentions legal action without a human.
Reconciliation agent
Matches payments to invoices, isolates exceptions into a review queue, prepares the month-cycle summary.
Invoice processing agent
Extracts, validates and posts supplier invoices with confidence scoring and a review queue for anything doubtful.
Lead qualification agent
Scores and enriches inbound leads, asks qualifying questions, and books the ones that meet your criteria.
Review & reputation agent
Monitors reviews across platforms, drafts responses, and escalates anything involving a health, safety or legal claim.
Reporting agent
Assembles the daily and weekly operating pack from your systems, with variance commentary rather than a raw data dump.
You can see every
agent, all the time.
A private workspace showing what each agent is doing, what it has done, what it cost and whether it is holding its accuracy benchmark. Pause any agent with one control.
See it on a call →- Status per agent — running, paused, degraded, awaiting approval
- Task volume today, this week, this month
- Accuracy against the benchmark, from the weekly QC sample
- Exception queue depth and oldest item age
- Consumption spend per agent, with burn rate per hour
- Last run, next scheduled run, and the last ten decisions with inputs
- Pause / run control and a hard-stop on consumption
- Autonomy level shown on the agent card, never hidden in settings
What agents may never do, enforced in code
- Publish medical or health claims — blocked at draft and routed to human review
- Price below the agreed margin floor — hard refusal plus an alert to the pricing owner
- Issue a refund above the approval threshold without a named human approver
- Send any reply that mentions legal action — escalated to the team lead
- Act on an unverified customer or account identity
- Move money, change bank details or alter payout instructions
- Delete records — archive and flag only
- Operate outside the log: every decision is stored with its inputs for 24 months
These are platform-level guards, not prompt instructions. A model cannot talk its way past them, and neither can a badly worded ticket. Full governance detail →
Send us one process. We’ll tell you what it costs to automate it.
The discovery call is free and carries no obligation. If AI is the wrong answer for your process, we would rather say so in the first hour than in month three.