We asked Claude the same question twice, once with context
Eighteen months of financials, practice analytics, time records and roster for a seven-practice dental group. One run got the data alone. The other got the same data plus the group's context brain, built stage by stage in this console and exported as markdown — definitions, ownership, thresholds, approved exceptions and dated history. Same model, same settings, separate sessions. Nothing else differed.
“Brief us on portfolio performance: which practices are underperforming, what is driving it, and what we should do about it.”
Anyone can use AI. Without context, it is guessing.
The gap is not model quality. It is that the model has your numbers and none of the judgment that makes them mean anything.
Anyone can use AI now
Everyone on your team can paste a spreadsheet into a model and get a confident, well-written answer back in seconds. That part is solved.
Without context, it is guessing
The model does not know how you define margin, what counts as bad, who owns the number, what already happened last quarter, or where a decision gets made. So it fills the gaps — fluently, and often wrongly.
A context layer closes the gap
A context brain, written once by the people who already carry this knowledge, sitting next to your data. The same model, the same question — a usable answer instead of a plausible one.
What a context brain actually is
A structured, validated representation of your organization — its people, locations, entities, systems, terminology, relationships, processes, rules, metrics, goals and history — written down so AI can reason about your group instead of dentistry in general. Not a place where information is stored: a model of how the organization works.
Organizational identity
What are the parts, and how do they relate?
Entities, practices and departments, and which belongs to which. Without it, five practices get read as one average.
People and responsibility
Who answers for this, by name?
Leadership, reporting lines, accountability, authority and escalation. Without a name, a finding is addressed to “management” — which means it is addressed to no one.
Location context
What is this practice actually capable of?
Operatories, providers, hours, specialties, patient base, staffing, capacity and constraints. Six chairs with four in use is a plan, not a problem.
Systems
Which system is authoritative for what?
What each platform does, who uses it, and which number wins when two disagree.
Operations
How does the work actually run here?
Scheduling, billing, insurance verification and treatment coordination as performed, not as written in the policy binder.
Semantics
What does this word mean in your group?
Revenue, production, active patient, new patient, cancellation, provider. Margin on collections, not production — otherwise a model computes something defensible and wrong.
Metrics
How is this calculated, and how bad is bad?
Formula, target, owner, source and review cadence. A number with no threshold is an observation; a number with one is an alarm.
Rules and exceptions
What is supposed to happen, and where is it different?
The standard, plus the practices that run under an exception and why. Exceptions are the part no report carries.
Strategy
What are we actually trying to do this year?
Targets, which practices are expected to grow, what constrains them and what is out of scope. This is what makes findings rankable.
Institutional memory
What already happened that explains this?
A planned leave, two resignations, a doctor starting in sixty days — what was decided and, when it matters, why. Without it every variance is investigated from scratch.
Case Overview
Cedar Ridge Dental Partners is a fictional seven-practice group. Here is what the data alone says about it — every line arithmetically correct, and every line pointing a leadership team at the wrong work.
Riverbend
The second-best profit margin in the group, at 23.5%.
MisleadingMaple Court
Production up year over year, so nothing to look at here.
MisleadingLakeside
Weakest margin in the group at 5.3% — adjustments are out of control.
MisleadingCedar Hollow
Loses money every month. One of the group's five underperformers.
MisleadingSouthfield
Lost a month of production in February. Coverage is slipping.
MisleadingAll five of those sentences are misleading.
That is the whole demonstration.
Every one of them is arithmetically correct. Not one of them leads a reader to the right conclusion. What separates them is a context brain: the definitions, thresholds, ownership, approved exceptions and history written down by the people who run the business.
CASE STUDY · A SEVEN-PRACTICE DENTAL GROUP
The setup
Two uploads. The numbers in them are identical, byte for byte. The second adds the group's exported context brain — definitions, ownership, thresholds, rules, approved exceptions, the intended day at each location, and everything that already happened, dated.
How we ran it
We did nothing clever. We opened a normal Claude session, attached the files below straight into the message — the four spreadsheets on their own for the first run, then the same four spreadsheets plus the context brain for the second — and typed the same question both times.
No tools, no plugins, no fine-tuning, no special instructions. Two fresh chats, the same model, the same prompt. The only thing that changed between them is what was attached to the message.
Upload A · the data
4 files. 1,802 rows of financials, analytics, time records and roster. No context of any kind. Click any file to open it.
Upload B · the data, plus the context
The same 4 data files, plus the exported context brain — 28 markdown files across 9 folders, about 34,600 words, and not one additional number. Click any file to open it.
- Model & settings
- Identical
- Prompt
- Identical, verbatim
- Sessions
- Separate and fresh
- Data files
- Identical, byte for byte
- The only variable
- The context brain
Both runs were asked through Claude. Nothing about the result depends on which model you use — the same two uploads and the same prompt run the same way on any general-purpose model.
The question — portfolio performance
“Brief us on portfolio performance: which practices are underperforming, what is driving it, and what we should do about it.”
The four practices this half turns on
Same rows in both runs. Without the context, the obvious reading of the numbers sends the work to the wrong place in three of the four.
Both runs find the collections problem. Only B can name its cause — the central billing seat sat vacant, so claims went out late — and only B knows the escalation threshold that decides when this stops being a practice matter.
“The driver is cash leakage and aging receivables”
“Put Maple Court on a 60-day AR recovery plan”
“Central billing vacancy created late claims and AR backlog”
“until Maple Court is below the 12% escalation threshold”
Lakeside runs the group's weakest margin, and adjustments did rise. But the lab allowance and the adjustment increase are both approved exceptions during an out-of-network migration. A attacks the exception and orders a fee-schedule review; B leaves the exception alone and goes after labour, which is genuinely over target at 30.2% of collections against a 26% standard.
“The major driver is adjustment pressure”
“Conduct payer/fee schedule review immediately”
“Do not blame lab or adjustment rate: Lakeside has approved exceptions”
“staff labor in Jun was 30.2% of collections, above the 26% target”
A sees hygiene visits fall and asks the team to work out why. B already knows: two hygienists resigned, the utilization target is formally suspended while the seats are open, and reappointment — which is not suspended — has fallen from the 85% standard to 63%. That is the difference between a diagnostic exercise and a specific fix.
“Confirm whether this is staffing, schedule template, or provider productivity”
“Hygiene capacity fell after two hygienists resigned”
“the utilization target is suspended under X-06, but reappointment still applies”
“Jun reappointment was 63.2%, still far below the 85% standard”
Both runs catch that the flattering margin comes from an empty dentist seat. B adds the standing rule for how long the exclusion lasts and what has to be true before the practice re-enters the ranking.
“Do not manage to the artificially high 23.5% EBITDA margin”
“excluded from margin ranking until the seat is filled and 90 days of production are recorded”
It is a leadership structure before it is an AI input
A context brain is not a file system of your organization, and not just a second brain you hand to AI. It is where the questions, definitions and processes nobody has asked, reviewed or answered are finally written down — everything a leadership team should be able to say out loud and have understood the same way by everyone.
Signs the answers were never written down
- If no one can name the owner of your leading indicator — it does not have an owner. It has a mention in a deck.
- If two leaders define collection rate differently — your board pack already contains two answers to the same question.
- If the rule for “does this project serve our strategy” was never written down — every project passes it, and has for years.
None of these are AI problems. They are leadership questions nobody ever asked out loud — and every one of them gets answered naturally in the course of building a second brain.
What it looked like here
The context brain used in this run — exported straight out of the console as 28 markdown files, about 34,600 words, containing almost no data. Open any of it.
Build the same thing for your group.
We run the interviews, reconcile the answers and hand you the finished folder. Tell us about your group and we will take it from there.
Build yours