Bring one workflow and one document set. We will show you where your run stops.
Talk to Us →Most vendors show you an AI that finishes. The interesting part is where ours refuses to — when two of your documents disagree, when a figure will not reconcile, when the answer is not in the record. Scroll a run and the meter moves with it.
One of the six was executed and scored against a published answer key. The other five are deployment patterns, marked as such.
Facilities · maintenance contract Scored run · answer key published
Four documents and a 245-row register. This is the one run on this page that was executed and scored against a known answer key — the working is published.
Reads the rate card at Table 7.1 and the interval table at Table 3.1.
Contract p.4 · Code p.30Matches all 245 register rows to their contracted rate.
19 lines billed above contracted rate — AED 3,436. 5 visits invoiced twice — AED 6,980.
Cited to both sources13 inspection intervals exceed the code maximum.
Code Table 3.1 p.30Gate 09 · Arithmetic returns the report. The draft rounded each line before summing and read 1.23%. Recomputed from rows: 1.21%.
Draft rejected — not annotated, rewrittenReport rewritten and re-gated. The reader never sees the failed draft; the audit log does.
Gate 10 · Contradiction fires. The code allows 60 days for AC-01 and AC-03; MS-114 permits 90. Both cited. The agent does not pick one.
Code p.30 vs MS-114 p.2Gate 12 · Question for you. No reason for the longer interval exists in the record, and no variation under Clause 10.2 was found. Who approved MS-114 Table 4.1, and against which edition?
Held for a named person| # | Step | Grounding | Result |
|---|---|---|---|
| 01 | Reads the rate card at Table 7.1 and the interval table at Table 3.1. | Contract p.4 · Code p.30 | Ran |
| 02 | Matches all 245 register rows to their contracted rate. | — | Ran |
| 03 | 19 lines billed above contracted rate — AED 3,436. 5 visits invoiced twice — AED 6,980. | Cited to both sources | Ran |
| 04 | 13 inspection intervals exceed the code maximum. | Code Table 3.1 p.30 | Ran |
| 05 | Gate 09 · Arithmetic returns the report. The draft rounded each line before summing and read 1.23%. Recomputed from rows: 1.21%. | Draft rejected — not annotated, rewritten | Returned |
| 06 | Report rewritten and re-gated. The reader never sees the failed draft; the audit log does. | — | Ran |
| 07 | Gate 10 · Contradiction fires. The code allows 60 days for AC-01 and AC-03; MS-114 permits 90. Both cited. The agent does not pick one. | Code p.30 vs MS-114 p.2 | Blocked |
| 08 | Gate 12 · Question for you. No reason for the longer interval exists in the record, and no variation under Clause 10.2 was found. Who approved MS-114 Table 4.1, and against which edition? | Held for a named person | Needs you |
AED 12,118 of AED 284,456 billed is not payable on the terms of the agreement. Nothing was disputed, credited or paid — the run produced a draft and stopped.
See the full working and the score against the answer key →Executed run — figures computed from four reference documents built to a published answer key, reproducible from seed 20260803. Full working at servai.ae/evidence/. · ServAI is a product of EADPAG, Dubai. servai.ae/cases/
Government · plan review Deployment pattern
A regulatory authority reviewing design submissions on hardware inside its own building, with no external network dependency.
Reads the submission and identifies the 34 code clauses that govern it.
Assesses each clause, citing code section and page beside the submission page it tests.
Every assessment citedTwo clauses cite an amended annex. Gate 11 flags both as superseded rather than quoting them as current.
Gate 06 · Data egress blocks a write. A scanned drawing was about to leave the site inside an outbound summary. It does not.
Nothing leaves the buildingThree clauses cannot be assessed from the submission as filed. Returned as precise questions to the applicant, not assumptions.
HeldDrafts the determination — unsigned.
The reviewing officer decides| # | Step | Grounding | Result |
|---|---|---|---|
| 01 | Reads the submission and identifies the 34 code clauses that govern it. | — | Ran |
| 02 | Assesses each clause, citing code section and page beside the submission page it tests. | Every assessment cited | Ran |
| 03 | Two clauses cite an amended annex. Gate 11 flags both as superseded rather than quoting them as current. | — | Ran |
| 04 | Gate 06 · Data egress blocks a write. A scanned drawing was about to leave the site inside an outbound summary. It does not. | Nothing leaves the building | Blocked |
| 05 | Three clauses cannot be assessed from the submission as filed. Returned as precise questions to the applicant, not assumptions. | Held | Needs you |
| 06 | Drafts the determination — unsigned. | The reviewing officer decides | Needs you |
ServAI prepares the assessment and shows its working. It does not issue determinations. The authority to decide never leaves the officer.
Deployment pattern — token counts and timings are representative of how the gates bind a run in this sector, not measurements of a live deployment. · ServAI is a product of EADPAG, Dubai. servai.ae/cases/
Healthcare · administrative only Deployment pattern · no clinical role
A hospital group's revenue-cycle function. Scoped so it cannot reach clinical decision-making at all.
Groups 12 months of rejections by payer, code and stated reason.
Three rejection reasons account for most of the value. Each traced to the payer clause it breaches.
Payer manual, cited to pageGate 10 fires. Two payer manuals contradict each other on one requirement. Both surfaced, cited. The agent does not pick one.
Drafts corrected documentation guidance for the coding team.
Prepares resubmissions. Every one waits for the revenue-cycle lead.
Held| # | Step | Grounding | Result |
|---|---|---|---|
| 01 | Groups 12 months of rejections by payer, code and stated reason. | — | Ran |
| 02 | Three rejection reasons account for most of the value. Each traced to the payer clause it breaches. | Payer manual, cited to page | Ran |
| 03 | Gate 10 fires. Two payer manuals contradict each other on one requirement. Both surfaced, cited. The agent does not pick one. | — | Blocked |
| 04 | Drafts corrected documentation guidance for the coding team. | — | Ran |
| 05 | Prepares resubmissions. Every one waits for the revenue-cycle lead. | Held | Needs you |
This deployment handles claims, coding, procurement and scheduling administration. It has no clinical role and is not a medical device. Diagnosis, treatment and clinical judgement remain entirely with licensed clinicians.
Deployment pattern — token counts and timings are representative of how the gates bind a run in this sector, not measurements of a live deployment. · ServAI is a product of EADPAG, Dubai. servai.ae/cases/
Construction · non-conformance Deployment pattern
A contractor running four concurrent packages. The QA manager needs an answer defensible to the client before the next progress meeting.
Reads the NCR, the governing spec clause, and the method statement that should have prevented it.
Recomputes recorded values against the specified tolerance. The method statement permits what the spec forbids.
Both documents citedChecks the other three packages for the same method statement — two use it.
Two inspection records for the period are missing. Gate 12 names them rather than assuming they passed.
HeldDrafts the corrective action and client NCR response in English and Arabic — figures and clause references carried across untouched.
Issues nothing. The QA manager signs.
Held| # | Step | Grounding | Result |
|---|---|---|---|
| 01 | Reads the NCR, the governing spec clause, and the method statement that should have prevented it. | — | Ran |
| 02 | Recomputes recorded values against the specified tolerance. The method statement permits what the spec forbids. | Both documents cited | Ran |
| 03 | Checks the other three packages for the same method statement — two use it. | — | Ran |
| 04 | Two inspection records for the period are missing. Gate 12 names them rather than assuming they passed. | Held | Needs you |
| 05 | Drafts the corrective action and client NCR response in English and Arabic — figures and clause references carried across untouched. | — | Ran |
| 06 | Issues nothing. The QA manager signs. | Held | Needs you |
The answer was not “someone made a mistake.” It was a document conflict affecting two other packages — with both documents cited, so the client can check it.
Deployment pattern — token counts and timings are representative of how the gates bind a run in this sector, not measurements of a live deployment. · ServAI is a product of EADPAG, Dubai. servai.ae/cases/
Retail · 62 branches, daily Deployment pattern · runs 06:00 daily
A supermarket chain across 62 sites. Head office currently learns about a cold-chain deviation when someone escalates it, which is usually late.
Reads yesterday's checks across all 62 branches, plus the temperature logs behind them.
4 branches with incomplete cold-chain checks. 1 logged excursion.
Quotes the SOP requirement each one breached, with section and page.
Branch manager gets the clause, not a complaintNotifies five branch managers and their regional lead. Gate 02 shows the recipient count before sending.
5 recipients, boundedThe excursion crosses the disposal threshold. That decision is held for the quality lead.
Not automated| # | Step | Grounding | Result |
|---|---|---|---|
| 01 | Reads yesterday's checks across all 62 branches, plus the temperature logs behind them. | — | Ran |
| 02 | 4 branches with incomplete cold-chain checks. 1 logged excursion. | — | Ran |
| 03 | Quotes the SOP requirement each one breached, with section and page. | Branch manager gets the clause, not a complaint | Ran |
| 04 | Notifies five branch managers and their regional lead. Gate 02 shows the recipient count before sending. | 5 recipients, bounded | Ran |
| 05 | The excursion crosses the disposal threshold. That decision is held for the quality lead. | Not automated | Needs you |
The office finds out at 06:00, not next week.
Deployment pattern — token counts and timings are representative of how the gates bind a run in this sector, not measurements of a live deployment. · ServAI is a product of EADPAG, Dubai. servai.ae/cases/
Logistics · quarterly performance Deployment pattern
A freight and last-mile operator. Included because two systems disagreed and the run refused to average them.
Pulls delivery performance from the TMS and the finance workbook.
Gate 10 fires. The two systems disagree on the quarter's volume. Both figures surfaced, cited to source. Neither is averaged away.
TMS vs workbookGate 09 returns the draft. A percentage in the summary does not reconcile with the table beneath it.
Rewritten, re-gatedRoot cause traced to one lane and one carrier, with the SLA clause and penalty terms quoted.
Drafts the carrier escalation and a reconciliation request to finance.
Held| # | Step | Grounding | Result |
|---|---|---|---|
| 01 | Pulls delivery performance from the TMS and the finance workbook. | — | Ran |
| 02 | Gate 10 fires. The two systems disagree on the quarter's volume. Both figures surfaced, cited to source. Neither is averaged away. | TMS vs workbook | Blocked |
| 03 | Gate 09 returns the draft. A percentage in the summary does not reconcile with the table beneath it. | Rewritten, re-gated | Returned |
| 04 | Root cause traced to one lane and one carrier, with the SLA clause and penalty terms quoted. | — | Ran |
| 05 | Drafts the carrier escalation and a reconciliation request to finance. | Held | Needs you |
A disagreement between two systems is a finding, not a rounding error. The run reported it and stopped.
Deployment pattern — token counts and timings are representative of how the gates bind a run in this sector, not measurements of a live deployment. · ServAI is a product of EADPAG, Dubai. servai.ae/cases/
Banking · onboarding and screening Deployment pattern
A UAE commercial bank. Included because this run stops before it reads anything — and not because the evidence was weak.
Stopped before the first document. Layla holds screening.alert on this seat. Whoever brings a customer in must not clear the alert raised against them.
onboarding.case ⊗ screening.alertSplits the request. The screening half is refused; the onboarding half is not, and continues.
Lists what the onboarding file is still missing: two beneficial-owner declarations and a board resolution.
onboarding.caseRoutes the alert to a second officer, and names why it could not be handled here.
| # | What happened | Cited | Outcome |
|---|---|---|---|
| 01 | Refused before reading. The seat holds screening.alert; onboarding and clearing the alert against the same customer may not sit together. | Separation | Blocked |
| 02 | Splits the request — the half it may do continues. | — | Ran |
| 03 | Lists what the onboarding file is missing. | File | Ran |
| 04 | Routes the alert to a second officer. | Held | Needs you |
At step one, on four hundred tokens. The other six runs on this page stop because a document is wrong or a figure will not reconcile — this one stops because of who asked. The evidence was fine. The separation is not a warning somebody clicks past: the capability is not on that seat.
Deployment pattern — token counts and timings are representative of how the gates bind a run in this sector, not measurements of a live deployment. · ServAI is a product of EADPAG, Dubai. servai.ae/cases/
Case 01 is a real execution: figures computed from four reference documents built to a published answer key, reproducible from seed 20260803, with the full working on the evidence page. Cases 02–06 are deployment patterns showing how the gates bind a run in each sector; their token counts and timings are representative, not measurements of a live deployment. We will run case 01's audit against your own documents under NDA.
Bring one workflow and one document set. We will show you where your run stops.
Talk to Us →