Solutions
Two ways to put AI to work inside an organisation without handing it the keys.
Agentic systems
Software agents that take on a defined job, end to end, under rules and approvals you set.
Good fits
- Recurring reporting and summaries
- Document intake, classification and routing
- Research and monitoring briefs
- Back-office workflows across several tools
Scoped access
An agent only sees the systems and documents it has been granted. Access is a list, not a default.
Readable plans
Before it acts, the agent writes its plan in steps. You can read it, change it, or stop it.
Approval points
Anything that leaves the organisation, spends money or changes a record waits for a person.
A full record
Every read, every draft and every action is logged, so a result can be traced back to its sources.
Anatomy of an agent
What we actually build when we say agent: five parts, and the person is one of them.
Granted sources
Only what is on the access list. Calendars, documents, exports.
The agent
Perceives, reasons in readable steps, acts through tools.
Your tools
Mail, files, calendars, records. Nothing new to install.
Approval point
A person releases anything consequential.
The log
Every read, draft and action, kept per run and traceable.
Set the level of autonomy
Every agent we build has a dial like this one. You choose where it sits, per job, and you can turn it down at any time.
- Observe
- Suggest
- Act with approval
- Act and report
Act with approval
The agent does the work. A person releases it.
Runs the whole job end to end, then waits at an approval point before anything leaves.
Guardrails, as standard
Access lists
An agent's reach is a list you can read and shorten.
Approval points
Sending, spending and changing records wait for a person.
Stop switch
Any agent can be paused in one action, mid-job.
Full log
What it read, decided and did, kept for every run.
Data boundaries
Private models and in-house storage where the data requires it.
Reversible first
We start with jobs that can be undone, and widen from there.
Private AI assistants
The imperative of private large language models. We are dedicated to the development of generative AI assistants that are based on private and secure large language models.
Talk to usPersonalised user experiences
Private LLMs can be fine-tuned to cater to the specific needs and preferences of individual users or organisations. This customisation enhances the relevance and effectiveness of AI assistants, providing more accurate responses and personalised interactions. By leveraging private LLMs, organisations can offer bespoke AI solutions that align closely with their operational objectives and user expectations.
Privacy concerns
Public LLMs, while highly effective, often require vast amounts of user data to function optimally. This data collection can inadvertently lead to privacy breaches, where sensitive personal information is exposed to unauthorised entities. Private LLMs, on the other hand, can be tailored to operate within strict privacy guidelines, ensuring that user data remains secure and confidential.
Data security
The centralised nature of public LLMs makes them attractive targets for cyberattacks. A breach in a public LLM's security infrastructure can have far-reaching consequences, potentially compromising the data of millions of users. Private LLMs mitigate this risk by decentralising data storage and processing, making it more challenging for malicious actors to access sensitive information.
Ethical use of data
The ethical implications of AI and LLMs cannot be overlooked. Public LLMs, governed by broader organisational policies, may not always align with individual user ethics and values. Private LLMs allow for greater control over data usage policies, ensuring that the deployment and operation of AI assistants adhere to ethical standards that respect user autonomy and consent.
Talk to us
Describe the job you want done. We will tell you honestly whether an agent, an assistant, or neither is the right answer.
- WhatsApp or iMessage
- +39 370 373 55 99