A real stake
Enough volume, or enough lost time, for the gain to be counted in days. Below that, simple automation does the job, at a fraction of the cost.
Peregrine AI
AI moves a ceiling. It fixes nothing.
When the volume outgrows what your teams can absorb, or a simple question takes half a day of digging, AI has its place. Connected to your data, it reads, sorts, prepares and answers. Your teams keep the decision, and your data stays under your control.
Let's see what it does with your data.
Uses
What AI would do in your company.
Six common uses, each with its safeguard: what stops an error from slipping through unnoticed. Every project starts from your process, not from a catalogue.
01
An assistant that knows your company
Management asks a question in plain language: what margin on this quarter's files, which customers have ordered nothing in six months, which salesperson converts best. The assistant queries your data, runs the calculations and answers, with figures and sources.
The safeguardRead-only access. Everyone only gets what their permissions allow them to see, and every question is logged.
02
Document reading and extraction
Invoices, certificates, statements, letters: the information is extracted, attached to the right file and checked against what you already know.
The safeguardEvery field carries a confidence score. Below the threshold, the document goes to human review, automatically.
03
Case preparation
The file reaches the case handler already prepared: documents checked, inconsistencies flagged, a reasoned proposal.
The safeguardA proposal, never a decision. The handler approves, corrects or rejects, and the correction is recorded.
04
Sorting and routing
Incoming mail, requests, reports: sorted by type, priority and responsible team, as soon as they arrive.
The safeguardNothing vanishes into an invisible queue: everything stays visible and reassignable in one click.
05
Search across your archives
Years of files, contracts and letters searchable in plain language.
The safeguardNo answer without a source document. Access rights apply to the search.
06
Assisted drafting
Letters, summaries, reasoned documents: a first draft based on the file and your templates.
The safeguardAnything that commits the company goes through a person; routine messages go out on their own, logged and traceable.
Conditions
What separates useful AI from a gadget.
A model placed on a fuzzy process produces errors faster, not fewer. We only deploy where these five conditions are met.
01
02
A structured process
Data entered once, written rules, logged approvals. Otherwise, AI amplifies the mess.
03
A human who decides
A commitment, a refusal, a payment: these are acts. AI proposes, a person signs.
04
Measured quality
A success rate measured on your real files before go-live, and tracked afterwards.
05
Your data protected
Where it is processed, by which provider, without training any model: written down at scoping.
No showcase chatbot: an assistant that answers your customers vaguely saves no one an hour. An assistant connected to your data does.
Method
We measure before we deploy.
The same approach every time, whatever the use. The second step may conclude it is not worth it: that is an acceptable outcome.
Where the ceiling is
We measure the real processing: how many documents, how many minutes each, where time is lost. From that we draw the candidate use and the expected gain, in hours and euros.
The trial on your real files
Your experts annotate a real sample. We measure what the model gets right, what it misses, and at what cost.
Result: it is not worth it. We tell you, with the figures.
Integration into your tools
AI is connected where your teams already work, with its confidence thresholds, its human review queue and its log. Your teams are trained to work with it.
Quality monitoring
The success rate is tracked over time; your teams' corrections improve it. If quality drifts, the use is switched off and human processing resumes, with no interruption.
Questions we are asked
On AI, precisely.
The questions management teams ask before committing to anything.
Where should a small business start with AI?
With the process, not the tool. We find where your teams read, sort, copy or search by hand, and measure the time it costs.
Then we test the most promising use on a sample of your real cases. If the gain does not cover the cost, we stop there, and you know it before committing to a rollout.
We already use ChatGPT. What would an assistant on our data add?
ChatGPT answers from what it learned elsewhere, and only sees of your company what someone pastes into it. A business assistant queries your own data (cases, sales, contracts) within each person's access rights, and cites its sources.
It answers "what margin on this quarter's cases?", with the figures to back it. And your teams stop pasting client data into a tool no one has vetted.
Is our data used to train an AI model?
No. Before any rollout, the scoping sets down in writing where your data is processed, by which provider, with which clause ruling out training, and what is retained.
For the most sensitive data, processing can be hosted on your premises.
Is AI compatible with the GDPR?
It can be, provided it is designed in from the start: a defined purpose, data limited to what is necessary, a provider bound by a data processing agreement, access rights by role and a log of every query.
These choices are written down at scoping, with your data protection officer if you have one.
What happens if the AI gets it wrong?
It sometimes does, and everything is built for that: a confidence threshold on every result, human review below the threshold, a full log of what was proposed and decided.
A reviewed error costs a few minutes; an invisible one costs a dispute.
How much does an artificial intelligence project cost?
The cost of an AI project depends on the use: the volume to process, the documents or data to read, the tools to connect, the level of human review. The trial on your real cases prices it before any rollout.
After that, a cost per item processed, visible every month. If it exceeds the human time saved, the use should not exist.
Will AI replace our teams?
In a growing company, the problem is rarely surplus staff: it is a processing backlog.
AI absorbs data entry and sorting; analysis, judgement and customer relationships stay human.
Can we use AI without custom software?
For the assistant on your data, yes, if it exists somewhere in a structured form.
For the rest, start by structuring: data entered once and written rules often free up more hours than a model.
What happens during the one-hour call?
You describe what your teams handle by hand; we look with you at whether AI has a place there, or whether simple automation is enough. You leave with a first candidate use, and what would need measuring to confirm it.
The call is by video, free of charge and with no commitment.
First call
Tell us what you want to hand over to AI. We will tell you whether it is worth it.
Reply within two hours, with a one-hour video call slot.