Nothing invented.
Nothing leaked.
Nothing locked in.
Six commitments, each with a mechanism behind it. Not a values statement. The things you can check for yourself at every stage of an engagement.
Nothing invented
Every fact traces to a source.
A document and page, an API record, a web page and timestamp. When the sources are silent, the answer is “not stated.” We test this harder than anything else, because a confident wrong answer costs more than no answer.
Nothing leaked
Rules enforced in code, not left to the model.
Supplier pricing, internal rates, personal data, anything you designate: what must never be shown is enforced in code and configuration, tested with adversarial questions, and logged when triggered.
Nothing duplicated
Every record carries a stable key.
Run it twice, get the same records, updated. Re-runs update rather than duplicate, and every merge we make is recorded so it can be undone.
Nothing hidden
Counts before and after every stage.
Including what failed and why. You see the dry run before every write. You can pull the audit log of any assistant conversation: which sources it read, which tools it called, what it answered.
Nothing locked in
Your data lives in accounts you control.
Your extracted data, your ontology, your graph, and your indexes export in open formats. The workflows are on platforms your team can open. If we disappear tomorrow, your systems keep running and your data is yours.
No promises we cannot keep
We say so, and propose the alternative.
If a document type carries no usable data, we show you and leave it out rather than charge to process noise. If a target system lacks a field, we tell you which one. If a data source is legally or technically off limits, we say so.
Nothing touches your systems unseen.
Writing back to a CRM, ERP or asset register is where most data projects do damage. Ours cannot skip a step.
Dry run
Exactly what will be written: how many records, to which parents, which fields, and which rows could not be matched and why. You read it before anything moves.
Write with a stable key
Upsert semantics on an external key. The target platform’s own authentication and credential store. No secrets in code, files or workflow definitions.
Verify from the target
A verification query against the target system. The count you see comes from the system itself, not from our logs.
Report the failures
What could not be processed, what was flagged for review, what was skipped. In the summary, not in an appendix.
Compliance is built in from the start.
Bolting it on later never works.
Logged decisions and human oversight
Agents log every tool call with who asked and what came back, and stop for a human when the action is irreversible or the confidence is low. This is what auditors and, from August 2026, European regulators expect of AI systems used in operations.
No training on your data
Commercially available models under contract terms that exclude training on your inputs. Model choice is per task, on accuracy and cost, and swappable.
Data residency
Storage, databases and indexes in the client’s accounts and region. Compute-heavy stages scaled to your volume and switched off when idle.
Lawful data collection
Public data only, no bypassing of technical barriers, terms of service respected, personal data handled under the applicable privacy regime. A written record of every data source and its terms.
Disclosure
Where AI-generated content reaches the public, it is disclosed.
Handover
Documentation a competent engineer can follow. Every dataset exportable as plain text, JSON, CSV, or a graph export.
Trust is earned, not claimed.
If you have questions about how we source, secure, write or hand over data, our technical team will walk through it in detail, with the pipeline open.