AI agent platforms
Bounded AI agent platforms that work over your tools: each agent has permitted actions, asks approval on material decisions and keeps a record of every step.
What we solve
An AI agent is not an autonomous employee: it is a delimited workflow that can look up information, prepare actions and execute only those it was authorised for. Everything else escalates to a person.
We design limits before capabilities: what it may touch, with which permissions, what needs human approval and how errors are detected and corrected.
Common problems
- Answering repetitive enquiries across several internal sources
- Preparing tasks (registrations, bookings, notices) for someone to validate
- Giving a first answer outside working hours with real context
- Gathering scattered information before a person decides
An illustrative example
Fictional case: a maintenance business receives “Is my Thursday visit confirmed?”. The request says “Thursday requested” and the schedule says “technician not yet assigned”. These are different states; there is no confirmed visit yet.
Proposed first scope: consult those two authorised sources and draft “The Thursday visit has been requested; the technician assignment still needs confirmation”, with a reference for each state. A person checks the sources and decides what to communicate. This scope does not book the visit, change the schedule or send the message.
An illustrative scenario explaining the service; not the outcome of a real customer nor a promise of results.
What the engagement includes
- An agent bounded to your authorised sources and actions
- Human approval and escalation rules
- A record of each enquiry, action and error
- Documented limits: what it cannot do
- A supervision guide for your team
What we need from you
- For the first enquiry, describe one recurring task, the tools you search and the result you want to obtain
- Separate looking up, drafting and executing: identify the step needing approval and the actions that must stay out of scope
- Identify who can authorise each source and who would review the result; do not send passwords or customer data in the initial enquiry
- Bring a fictional case or one anonymised with authorisation, including an exception: for example, a requested date that is not yet confirmed
Scope and costs
Scope is agreed in writing: sources, authorised actions, approvals and an observation period. We start with one small, well-delimited agent.
Agents consume AI models and hosting: those recurring costs are estimated separately and depend on usage. We promise neither unsupervised autonomy nor results without prior assessment.
Frequent questions
Does the agent act alone?
Not without limits. It executes only authorised low-risk actions; material decisions need human approval and stay recorded.
Which tools can it work with?
Those with an authorised API or a usable export. We assess each system before promising a connection.
What if the agent makes a mistake?
The error stays recorded, the case escalates and the cause is fixed: sources, instructions or limits. Without records there is no improvement.
Does it replace my employees?
No. It prepares work and answers routine matters under supervision; the decisions and relationships that matter stay human.
How would we check the first scope before using it?
We can agree acceptance cases like the example: distinguish a request from a confirmation, show the sources and leave communication to the responsible person. If a source is missing or the states conflict, the expected result is to identify the gap and request review. This is a proposed project criterion, not a test already performed or a guarantee of error-free results.
Does reviewing a draft also authorise sending it or changing records?
These are separate steps. For the illustrated scope we propose only lookups and a draft; the person sends through their usual channel. If you later need sending or writing, the specific action, permissions, approval and result checks need to be defined separately. Read access alone does not authorise those steps.
Keep exploring
API or export: assess access before connecting an agent
Scope depends on the process, data and existing systems. We assess feasibility, risk and applicable obligations before proposing a solution.
What the project delivers
- Bounded agents
- Supervised actions
- Records and escalation
