Role-scoped AI assistant
Five specialists in one chat
Per-specialist skill gating means an assistant cannot edit data outside its role. Forty-seven tool-bound skills, zero cross-domain hallucinations.
Read the caseAll case studies
Case study · Account onboarding
Meridian Operations · new-account setup automation2.5 hours → 18 minutes per new account
A folder of historical packets goes in. A fully seeded account record, product catalog, and configuration set come out. The operator reviews and approves; they don't reconstruct from scratch.
0 min
New account setup vs. 2.5 hr
0x
Faster catalog seeding vs. manual
0.00
Configuration surfaces accepted without edits
Built with
The problem
A morning per new account, and 30% of them went live with no catalog at all
A new account arrived with a folder of historical packets — PDFs, Excels, occasionally scans. Setting up the account meant opening every one, pulling party names, product lines, regulatory categories, then re-typing them into the system. Different operators ran the steps in different orders. The catalog seeded fully sometimes, partially other times, and not at all when the operator ran out of time.
- The folder problemA new submitting account arrived with a ZIP of historical packets — PDF, sometimes Excel, occasionally scanned images. Someone had to open each one, pull party names, product lines, and regulatory categories, and re-enter them into the system. Forty packets took a morning.
- Configuration drift between operatorsThere was no enforced sequence. One operator configured the catalog before the account defaults; another did it after. The result was inconsistencies that surfaced weeks later as wrong outputs on live submissions.
- The 30% catalog gapBefore the specialist, 30% of new accounts went live without a starter catalog. Seeding was supposed to happen during onboarding but got skipped when the operator ran short on time. Those accounts generated manual catalog-lookup overhead on every early submission.
- Expert time on mechanical workThe operator who knew the regulatory categories well enough to configure them correctly was also the person who could have been reviewing live submissions. New-account setup consumed expert hours on tasks that were 90% mechanical.
historical-packets-northbay.zip · 42 files
one morningMO-2024-0142.pdfparsed8:14am
MO-2024-0143.xlsxconvert pending8:22am
MO-2024-0144_scan.jpgocr pending8:29am
MO-2024-0145.pdfparsed8:37am
MO-2024-0146.pdfparsed8:42am
...37 more—
Operator opens each manually2.5 hr total
The pipeline
From inbox to verified record in one pass
Six steps from folder drop to committed configuration. The operator reviews surface-by-surface; nothing commits without explicit approval.
01Pick
Onboarding specialist selected
Operator opens the assistant and selects the onboarding persona; grant required to see the option.
02Intake
Folder of historical packets parsed
Operator drops the folder in. Specialist identifies file types, runs CloudConvert on non-PDF formats, and parses each packet for structured fields.
CloudConvert
03Seed
Account record created from samples
Account record built with parties, address, tax identifier, and defaults derived from the historical sample. Cross-packet conflicts surfaced for operator resolution.
04Catalog
Products and aliases consolidated
Products extracted across the historical packets. Alias candidates consolidated. Classification codes proposed with confidence and cited sources.
Claude
05Configure
Per-account settings derived
Default consignee, regulatory category defaults, auto-routing toggles. Each setting is shown with the source packet(s) that drove the proposal.
Claude
06Confirm
Surface-by-surface review and commit
Operator reviews each surface and approves, edits, or rejects individually. Commit writes the approved configuration with the operator's identity on the audit row.
Clean historical packet set
18-minute onboarding with complete catalog
Specialist parses all packets, resolves aliases, proposes configuration with clear sourcing. Operator approves with minor edits. Account goes live with a complete catalog.
Sparse or inconsistent packets
Partial setup with checklist for first-live action
Fewer than five historical packets means thin catalog coverage. Specialist proposes only fields it can source; auto-routing toggles are all flagged for operator review rather than proposed as defaults. Account goes live in 'partial setup' state with a checklist.
Validation review
First-pass accuracy per setup surface
Confidence is per surface, not per field. The auto-routing surface is intentionally low — those toggles depend on operator preference that historical packets cannot reveal.
Field-level confidence
Pass 2 — Claude self-review
HighMediumLow
Account recordName, address, tax identifier
99%High
Default consigneeMost-frequent across packets
97%High
Product catalogCodes, aliases, descriptions
91%High
Regulatory category settingsDefaults derived from product types
88%High
Auto-routing togglesPer-rule routing preferences
61%Low
Routed to human review. Auto-routing depends on operator preference that doesn't appear in historical packets. Specialist proposes the most conservative default and surfaces every toggle for review — by design, not by failure.
4 of 5 fields cleared the 0.85 threshold
model: First-pass accuracyThe stack
Boring tech, glued together well
Each vendor handles what it's best at. Aisyst owns the orchestration layer in between.
Claude
Parses packets, consolidates aliases, proposes codes and configuration
CloudConvert
Converts non-PDF historical files (Excel, Word, scans) to a parseable format
Azure Blob
Historical packet storage and retrieval during the onboarding session
PostgreSQL
Account records, product catalog, configuration, audit rows
Drizzle ORM
Type-safe queries against the configuration tables
Third-party logos are trademarks of their respective owners and appear here only to indicate integration.
Outcomes
What changed when onboarding became a wizard
The expert still reviews and approves. They stop reconstructing. The 30% no-catalog gap closed because the catalog seeds before the operator can run out of time.
0 min
New account setup time
0x
Faster catalog seeding
0.00
First-pass accuracy
0%
Accounts onboarded with starter catalog (vs. 30%)
Watch first-pass configuration accuracy
Currently 0.93. If it drifts below 0.88, the historical-packet parser is missing a new common shape — a new file format, a new field arrangement, a new consignee pattern. The fix is almost always a prompt addition with a representative case, not a model retrain.