AI Across the Platform
AI is not a module in Solidlio. It is a metered capability wired into ticketing, changes, projects, billing, purchasing, the catalogue and the store — one credit ledger, one model-routing policy, one set of tenant boundaries.
The problem
“AI features” usually arrive as a separate product with a separate bill, a separate login and a separate idea of who may see what. The result is a tool that works beautifully in a demo and cannot be turned on in production, because nobody can answer what it cost last month, which client’s data it read, or what happens when the vendor declines a request.
What Solidlio does about it
Every AI operation in the platform goes through one service, one credit ledger and one routing matrix. Each operation declares its cost in credits and is charged before the model runs, so a plan limit is an enforced ceiling rather than a line in a brochure. Each operation also declares which model tier it needs — fast classification on Haiku, drafting and conversation on Sonnet — so the platform is not paying frontier prices for a category label. Cost in dollars is recorded per request against the paying account. And where a feature is deterministic rules rather than a model, the response says so.
Capabilities by area
| Area | What AI does |
|---|---|
| Ticketing | Auto-triage, duplicate detection, similar tickets, response drafting, deflection |
| Change management | Risk assessment, approver suggestions, plan review, checklists, conflict analysis, PIR analysis |
| Projects | Plan generation, SOW drafting, status reports, risk assessment, resource suggestions |
| Billing | Invoice extraction from documents, matching, categorisation, payment and credit triage |
| Purchasing | Order triage, reserve analysis, commission dispute triage and rule recommendations |
| Catalogue | Description generation, category suggestion, brand/vendor/distributor portfolio analysis |
| Store | Order triage, product recommendations, custom item descriptions, fee and pricing analysis |
| Assets & software | Health analysis, triage, licence optimisation, compliance risk |
| Knowledge | Indexing, categorisation, semantic retrieval |
| Reporting | Report triage and narrative generation, with a rule-based fallback |
| Branding | Colour palette extraction from an uploaded logo |
| The assistant | Sol, in all four portals |
How the money works
Every operation has a credit price, charged before the model call:
| Operation | Credits |
|---|---|
| Categorisation | 1 |
| Summaries | 2 |
| Retrieval-augmented query | 3 |
| Sol chat turn | 3 |
| Change analysis | 4 |
| Invoice extraction | 5 |
| Sol vision / tool-use turn | 5 |
| Brand palette extraction | 5 |
| Each tool call inside a Sol turn | +1 |
Plan tiers set the monthly allowance. When the allowance is exhausted the request is refused with a clear code, not silently degraded. Where a plan does not include a capability at all, the refusal carries a structured upgrade offer.
Actual model spend in dollars is recorded per request — model, token counts, cache reads and writes — against the paying account, so cost is attributable rather than a single platform-wide number.
Built for MSPs and their clients
AI credits follow the money. By default an operation is charged to the acting account. An MSP can grant a managed client access to its own credit pool per feature group — ticket AI, invoice AI, the assistant, knowledge AI, change AI — so a provider can switch on triage for one client without handing over its whole allowance. Where an MSP runs an analysis about a client, the client’s own policy is consulted, and the caller must be able to prove it services that client before anything runs.
Honest failure
- A model outage returns a 200 with the analysis fields null and a
fallbackflag — never a plausible-looking invented number. - A refusal by the model’s safety classifiers is reported as a refusal, not as a parsing failure.
- A truncated response is reported as truncated rather than partially parsed.
Limits
- AI is an enhancement, not a system of record. Analyses are advisory and are applied to real fields only on acceptance.
- If metering itself is unavailable, AI requests are refused rather than run unmetered.
- Batch processing is not currently offered.