- Which task, specifically, and how often
- What it reads and what it writes
- Where a wrong answer costs something
- What 'done' looks like to whoever owns it today
Custom AI Agents
When the job is specific to you
Some work does not fit a product. An agent scoped to one workflow — reading your systems, following your rules, escalating on your terms — is often a fortnight of build and a permanent removal of somebody's worst recurring task.
Scope my agentNo commitment — a scoping call and an honest read on whether this fits.
The short answer
A custom AI agent is software scoped to one workflow in your business: it reads from the systems you already run, acts within limits you set, escalates to a person on your terms, and logs everything it did. It is not a chatbot and it is not a platform subscription — it is a build, and you own it.
What's included
What we actually build
Every group below is a real deliverable, not a capability list.
- The scope, written down and agreed
- System connections, read and write separately permissioned
- The decision rules, in your language
- The escalation path
- The audit log
- CRMs and helpdesks
- Spreadsheets, databases and internal APIs
- Email, calendars and messaging
- Accounting and ERP, read-only unless you say otherwise
- Anything with an API
- A written scope it cannot act outside
- Write access granted per system, not blanket
- A human approval step on anything irreversible
- Rate limits, so a bad loop is contained
- Every action taken, with its input and its reasoning
- Every escalation, and why
- A weekly summary of what it did and what it refused
- The source, in your repository
- Documentation and runbooks
- A rebuild guide, so you are not locked to us
- Thirty days of support
Measurement
What it is expected to move
Figures are typical ranges from comparable builds, and the ranges are wide where they honestly should be.
Typical build2–5 weeksscope dependent
Most useful agents are small. The long ones are long because of system access, not because of the model.
Task removal5–15hper week, per agent
The measure that matters is somebody's recurring hours, not a percentage of a process nobody timed.
Auditability100%of actions logged
Every action, its input and its reasoning. If you cannot inspect what it did, you cannot responsibly let it act.
OwnershipFullcode and prompts
Delivered into your repository with documentation and a rebuild guide. There is nothing to buy back.
Engagement options
Three ways to start
Indicative starting points. Final scope depends on your systems and how much of the groundwork already exists.
Start
Single agent
FromFrom $7,500One workflow, scoped and built, with the audit trail and the handover.
- Discovery and written scope
- Up to three system integrations
- Decision rules and escalation path
- Full audit logging
- Source and documentation delivered
- 30 days of support
Most chosen
Agent set
FromFrom $18,000Two or three agents that share a system layer and a set of rules.
- Everything in the single agent
- Shared integration and permission layer
- Consistent escalation across agents
- A single audit view
- Team training session
- Quarterly review
Ongoing
Managed
FromFrom $2,800/moWe run them, watch the logs, and change them as the work changes.
- Everything above
- Monitoring and incident response
- Rule and prompt maintenance
- New integrations as systems change
- Monthly readout
- A named engineer
Before you ask
Custom AI Agents, answered
The questions that come up on every scoping call, including the uncomfortable ones.
Scope my agentHow is this different from a chatbot?
A chatbot answers. An agent acts — it reads your systems, makes a decision within limits you set and writes something back. That difference is why the scope, the permissions and the audit log matter far more here than the model does.
What should we NOT automate this way?
Anything where a wrong action is expensive and hard to reverse, and anything where the rules are genuinely a matter of judgement rather than policy. We will say so during discovery. The best candidates are high-frequency, low-ambiguity tasks somebody already does the same way every time.
Do we own it?
Completely. The source goes into your repository, with documentation, runbooks and a rebuild guide so you are not locked to us. There is no licence and nothing to buy back.
What access does it need?
As little as possible, granted per system and read-only by default. Write access is requested explicitly, per system, and anything irreversible sits behind a human approval step. If a scope needs more access than you are comfortable granting, that is a signal about the scope.
What if it does something wrong?
It logs every action with its input and its reasoning, so 'what happened' is always answerable — which is the first requirement of letting software act on your behalf. Beyond that: written scope limits, per-system permissions, approval on anything irreversible, and rate limits so a bad loop is contained rather than compounding.
How long does it take?
Two to five weeks for a single agent. The variable is almost always system access — getting API credentials and agreeing permissions takes longer than the build in most organisations.
Tell us the task somebody hates
The best candidates are usually the recurring jobs nobody wants: the same report, the same reconciliation, the same copy-paste between two systems. Describe one and we will tell you honestly whether an agent is the right answer.
Digital OttersWhat you get
2–5 weeks for a single agentYou own the codeEvery action logged and inspectableWe will say when NOT to automate