AI Training baseline
The company baseline is the upper layer of Wisteria’s AI Training Profile — the context that applies to every department, every AI evaluation.
For the concept and per-department layer, see AI Training Profile.
Where to set it
Settings → AI Training → Company tab. Super admin only.
What to fill in
Five plain-text fields:
What does your organisation do?
The elevator pitch. Two or three sentences. Claude reads this the same way a new hire would.
Example:
Cedarbrook Memorial Health is a regional hospital network with 12 sites and 6,800 staff. We provide acute care, outpatient services, and specialty programmes in oncology and cardiac care across the Pacific Northwest.
Who are your customers / who do you serve?
Patients, retail consumers, enterprise clients, students, employees — whoever your team’s work is in service of.
Example:
Patients in inpatient, outpatient, and emergency settings. We also serve referring physicians and care coordinators at partner clinics.
What kinds of training matter most?
The categories of training your organisation prioritises.
Example:
Clinical compliance (HIPAA, joint commission), patient safety, infection control, equipment certification, customer service for non-clinical staff, and onboarding for new hires.
What’s NOT training-relevant?
The exclusion list. This is where you keep noise out.
Example:
Financial reports, marketing materials, executive memos, vendor invoices, individual patient records, internal IT documentation, board meeting minutes.
Tone and culture notes
Anything that should colour the AI’s framing.
Example:
We use plain language whenever possible — avoid medical jargon in non-clinical training. Cedarbrook’s culture is “compassionate competence” — training should reflect both. We don’t refer to patients as “customers.”
When to update
- After the first scan — look at the suggestions Wisteria produced. If anything obvious is wrong (irrelevant categories surfaced, relevant ones missed), edit the baseline.
- When your org changes scope — new line of business, new compliance regime, new acquisition.
- Once a quarter as a maintenance pass.
The baseline isn’t fire-and-forget. It’s a living description.
What happens if it’s empty
Claude falls back to generic reasoning. Workable but noisy. You’ll spend more time dismissing irrelevant suggestions.
Recommended approach: fill in the elevator pitch and the exclusion list before the first scan. Add the other fields after a week or two of seeing what the AI surfaces.
Privacy
The baseline text is sent to Anthropic’s API (Claude) along with each candidate file’s text. Anthropic doesn’t retain API data for training; it’s used to produce the evaluation and discarded.
If your baseline contains anything you wouldn’t want Claude to see (proprietary trade secrets, sensitive personnel info), keep it out of the baseline. The baseline is “context the AI needs,” not “everything secret about us.”