Horus Eye · The Fair Question
“Why can’t I just use Claude directly?”
If you’re evaluating Claude for Salesforce work — or any AI layer connected to your org — this is the right first question. The honest answer: Horus Eye is not a UI on top of an AI model. The AI is the optional last 10%; the product is the 90% underneath it. Here is exactly what that 90% is.
Nine things a chat can’t give you
Kinds of Tool, Not BrandsFindings you can rely on
A general AI assistant: generated per conversation — ask twice, get two answers;
it can invent a problem that isn’t there.
Horus Eye: 202 deterministic rules
in code with published thresholds. Same org → same findings, every time. Works fully with
AI switched off.
The whole org, not a sample
An assistant: reads what fits in a context window, one component at a time.
Horus Eye: captures the org’s metadata and dependency graph in one scan, then
reasons over the whole population — percentiles (is this object a 95th-percentile
outlier?), blast radius across every dependent component, cross-package coupling.
A change that really happened
An assistant: writes through the API and reports what the API said.
Horus Eye: reads every change back from the org. Real case: Salesforce returned
success for withdrawing Modify All from a permission set — and changed nothing,
because the set also held “Modify All Data”. Horus Eye caught it by reading back. A
chat assistant would have told you it was fixed.
Safe writes
An assistant: can do whatever the connected user can do.
Horus Eye: 58 fix operations, each proven live on a real org before it may run.
Validate with zero writes → your approval → apply → read back → targeted
re-check → closure receipt → undo to the exact prior state.
Salesforce’s hidden rules
An assistant: has to rediscover them each time, if it notices at all.
Horus Eye: encoded from hundreds of live investigation cycles — a standard profile
cannot be edited in place; “System Administrator” is Admin to the Metadata
API; global permissions silently keep object grants on. These live in the product, not in a
prompt.
AI that can’t overstate
An assistant: free text, confident tone whether or not the data supports it.
Horus Eye: every AI answer must cite the evidence it rests on and pass a hedging check,
or it is discarded — and not charged. You see exactly what the AI was sent. Record data is
never sent.
Memory and accountability
An assistant: each chat starts from nothing.
Horus Eye: scan history, trends, regressions; your decisions remembered —
intentional, snoozed, answered; who approved what and when; closure receipts that say how each
finding was verified.
A team, not a chat
An assistant: one person’s conversation.
Horus Eye: shared orgs, a Fix list with owners, the report card and Impact Pack for
stakeholders, governed credits.
Predictable cost
An assistant: token bills that scale with how much you paste.
Horus Eye: scans and findings need no AI; AI actions are fixed-price credits (a Consult is
5); a rejected read is free. The pricing page →
The honest close
Use BothUse Claude directly for what it is great at — explaining a concept, drafting one Apex class, a one-off question. Horus Eye uses Claude too (Anthropic is our model provider). What Claude on its own does not give you is a complete, deterministic picture of your org, fixes that are proven before they run and verified after, an undo, and a record of all of it. That is what Horus Eye is.
This comparison is about kinds of tool, not any vendor’s product; statements about general assistants describe the category. Last reviewed October 2026.
Protected AI — the Trust Layer
Every AI read in Horus Eye is metadata-first: built from an allowlist, sanitized by pattern, made through one audited gateway and checked by an output firewall on the way back — and AI is off by default, per org. Every request carries an audit id and an AI Data Scope manifest you can open.
Frequently asked questions
FAQIs Horus Eye just a UI on top of an AI model?
No — the AI is the optional last 10%. Findings come from 202 deterministic rules in code with published thresholds; the same org produces the same findings every time, with AI switched off entirely.
Can't Claude connected to my org do the same analysis?
A connected assistant reads what fits in a context window, one component at a time, and its answers vary between conversations. Horus Eye captures the whole org's metadata and dependency graph in one scan and reasons over the full population — percentiles, blast radius, cross-package coupling.
What about making changes?
An assistant writes through the API and reports what the API said. Horus Eye runs only operations proven live before they may run, with validation, your approval, read-back from the org, a targeted re-check, a closure receipt — and an undo to the exact prior state.
See the difference on your own org
Connect read-only, run the 202 rules at no cost, and judge the evidence yourself — Claude for Salesforce questions, Horus Eye for the org.
Try Horus Eye Talk to an Architect All Horus Eye pages