A live FinOps cost governance dashboard — spend by provider, model, user and team, budgets with burn-rate alerts, spend reconciliation, anomaly detection and cost-to-value ROI — plus an in-process adversarial self-test harness, reversible PII tokenization, a policy atomizer, and a lighter agent tool-call efficiency toolkit. The same gateway that enforces policy now tells you what it's costing, and helps you prove it's safe.
Enforcement without visibility eventually gets questioned by whoever signs the AI bill, and visibility without enforcement is just a nicer invoice. Smartflow 1.10 puts a real cost dashboard directly on top of the same per-request telemetry the gateway already produces for compliance — so the numbers finance sees and the traffic Shield is enforcing are the same data, not two systems that quietly drift apart. Alongside it, three governance additions close gaps a security review will ask about (adversarial self-testing, reversible redaction, policy-document onboarding), and a tool-call efficiency toolkit gives agent builders fewer, cheaper round trips without changing how they call tools.
| # | Feature | Category | Who it's for |
|---|---|---|---|
| 1 | FinOps cost governance dashboard — spend by provider/model, department & group attribution, week/month/year cards, top spenders, cache-savings, CSV export & PDF report | New Feature | Finance, platform admins, procurement |
| 2 | Scheduled spend digest — recurring "what changed & where to save" briefing, delta vs. prior period, top movers | New Feature | FinOps leads, recurring reporting |
| 3 | Budgets & burn-rate alerts — monthly caps with forecasted burn rate; over-pace budgets can auto-stage a policy proposal | New Feature | Cost owners, budget holders |
| 4 | Spend reconciliation — provider billing connectors vs. gateway-metered spend; surfaces ungoverned spend | New Feature | FinOps, procurement, security |
| 5 | Anomaly & spike detection — daily spend timeseries with flagged anomalies and a one-click drafted policy | New Feature | On-call, FinOps |
| 6 | Allocation, concentration & ROI — provider concentration (HHI), allocation coverage, cost-to-value ratio | New Feature | Leadership reporting |
| 7 | Smarter policy proposals — scope/target picked from live groups, users, providers & models; model-to-model routing proposals | Enhancement | Finance & policy approvers |
| 8 | Adversarial self-test harness — in-process red-team suite across 6 categories with risk scoring and suggested block rules | New Feature | Security, compliance |
| 9 | Reversible PII tokenization — opt-in redact-with-key / unredact, ephemeral per-request key | New Feature | Data protection teams |
| 10 | Policy atomizer — turns policy text into machine-checkable rules for onboarding | New Feature | Compliance onboarding |
| 11 | Control & policy ID traceability — per-request control/policy IDs carried through the log and sealed in the audit chain | New Feature | Auditors, examiners |
| 12 | Agent tool-call efficiency toolkit — reduction benchmark, discovery cache, quality-delta check, trajectory caching, lighter smartflow-sdk | New Feature · SDK | Agent builders |
A live dashboard tab built on the same VAS telemetry Smartflow already produces for every request — no separate export pipeline to keep in sync. Spend rolls up by provider, model, user, and by department or group where that hierarchy exists in the identity provider; where it doesn't, admins can tag users into FinOps groups directly in the dashboard.
Cards at the top answer the questions finance actually asks first: spend this week / month / year (toggle in place), the highest-spend provider and model, and the top 5 spenders over the trailing 30 days. A dedicated card shows cache and token savings by provider over the same selectable window, so the savings story sits next to the spend story rather than in a different tool.
Export to CSV for spreadsheet work, or generate a print-optimized PDF report for a stakeholder who wants a document, not a login.
Finance stops asking platform teams for a one-off spend pull every month — the dashboard is the pull, always current, and exportable on demand.
A generated briefing — what changed since the last period, which provider or model moved the most, and where the recommendation engine sees savings — on a schedule you configure rather than a report someone has to remember to run. This is the "define a report for next cycle" button made real: pick a cadence, and the dashboard renders the same digest logic into a shareable HTML/PDF automatically.
A recurring FinOps update that doesn't depend on someone remembering to build the slide.
Set a monthly cap for a provider, model, group or user, and the dashboard forecasts burn rate against it — not just "you're at 80%" but "at this pace you hit the cap on the 19th." A budget that's on track to breach can auto-stage a policy proposal (route to a cheaper model, cap a group, tighten a rate limit) into the same approval flow described below, rather than just sending an email nobody actioned in time.
Budget overruns become a proposed fix waiting for approval, not a surprise on next month's invoice.
Point the dashboard at a provider's own billing export (read-only) and it reconciles that figure against what the gateway actually metered for the same window. The gap is ungoverned spend — usage that happened outside Smartflow entirely, bypassing policy, caching and audit. Connectors are managed the same way as everything else in the dashboard: add, update, or remove them, and trigger a policy proposal directly from a reconciliation gap.
"Are we actually seeing all our AI spend, or just what's routed through the gateway?" gets a number instead of a shrug.
A daily spend timeseries with statistically flagged anomalies plotted directly on the chart, so an unusual day is visible at a glance rather than buried in a monthly total. Clicking a flagged spike drafts a policy proposal pre-scoped to whatever provider, model, user or group drove it — the fix is one approval away from the anomaly, not a separate investigation.
A spend spike gets caught the day it happens, with a drafted fix already scoped to the cause.
Three numbers leadership tends to ask for and rarely gets in one place: provider concentration (an HHI-style score — are you dangerously dependent on one provider), allocation coverage (how much spend is actually attributed to a user, group or cost center vs. unattributed), and a cost-to-value ratio once you enter an estimated monthly value the AI usage is producing.
A ROI conversation with a number behind it, instead of "it feels worth it."
The propose-a-policy drawer (triggered from recommendations, budgets, anomalies or reconciliation) no longer asks anyone to type a scope by hand. Scope and target now pull from live backend data — the actual groups, users, providers and models the gateway has seen — so a finance user picks from real options instead of guessing a string that might not match anything. A second section lets you propose model-to-model routing: pick a source model and a target model from dropdowns (for example, route a chat-heavy group from a frontier model to a cheaper one for low-complexity calls) as a distinct proposal type alongside caps and blocks. Proposals are staged for a finance or admin approver; who exactly approves is being scoped separately as part of entitlement work.
A cost recommendation turns into an accurate, approvable policy in one drawer, not a support ticket asking someone to build it.
An in-process red-team suite that runs a corpus of adversarial prompts against your own deployment — prompt injection, jailbreak attempts, data exfiltration framing, and more — across 6 attack categories, and scores what got through. Every run produces a risk score, a heuristic rollup against common control frameworks, and a set of suggested block rules for whatever slipped past current policy. Runs entirely in-process against your own gateway; nothing leaves the deployment.
Blocking every risky-looking pattern outright trades false positives for false confidence — legitimate security research, red-team exercises and support workflows routinely use language that looks adversarial out of context. A self-test harness gives you the same coverage as a hard block, but as an evidence-producing exercise you run on your own terms, with a suggested rule you choose to adopt rather than a rule that silently changed your production traffic.
Standard redaction is a one-way trip — once PII is masked, nobody, including the original requester, can get it back. Reversible tokenization is opt-in per request: sensitive spans are replaced with stable tokens before the request reaches the model provider, using an ephemeral key generated fresh for that request and held only in memory. If the caller needs the original value back — to complete a workflow that unavoidably needs it — it can be un-redacted with that same key. The key is never persisted, so a compromised log or export still can't recover the original text.
Workflows that need PII downstream can still send it through Smartflow's scanning and audit path — instead of routing around redaction entirely because it was one-way.
Turns a policy document's plain-language text into a set of atomic, machine-checkable rules — an LLM-first pass with a heuristic fallback so it degrades gracefully rather than failing outright. Instead of a compliance team hand-writing rule syntax from a 40-page acceptable-use policy, they paste the relevant text in and get back a structured, reviewable rule set to refine and publish.
Onboarding an existing acceptable-use or data-handling policy into enforceable rules drops from a manual rewrite to a review pass.
Every request log entry can now carry the specific control IDs and policy IDs that acted on it, not just a pass/fail decision. That mapping is sealed into the tamper-evident audit chain and included in observability exports, so an auditor asking "which control caught this, and which policy defined it" gets a direct answer traced back to the exact request — not a reconstruction from timestamps and best guesses.
A receipt-level answer to "prove which rule did what" for any single request in the log, not just an aggregate compliance percentage.
Agents that repeatedly call the same or similar tools waste round trips re-fetching schemas and re-running work they've already done. This release exposes that existing MCP caching layer as usable infrastructure for agent builders, in four parts:
All of it is reachable from a lighter smartflow-sdk (Python, v0.5.0): a new @smartflow_tool decorator and smartflow_task context manager wrap existing tool functions with almost no code change, wiring them into trajectory caching automatically.
# pip install --upgrade smartflow-sdk (0.5.0) from smartflow import smartflow_tool, smartflow_task # wraps an existing tool function; participates in trajectory caching @smartflow_tool def list_files(path: str) -> list[str]: ... # groups a sequence of tool calls under one cacheable task with smartflow_task("summarize repo structure"): list_files("/src") ...
Fewer, cheaper round trips for agents that repeat the same discovery and task patterns — measured, not assumed, via the benchmark endpoint.
Runtime image — pull the 1.10 runtime images as usual. Contact your Aperion account team if you need the exact tags for your deployment channel.
FinOps dashboard — appears automatically as a new tab once the dashboard image is updated; it reads existing VAS telemetry, so there's no backfill step. Group tagging and provider billing connectors are optional and off until configured.
Reversible PII tokenization — opt-in and off by default. Enable per deployment when a workflow genuinely needs the original value back; ephemeral keys are never persisted to disk or logs.
Policy atomizer — LLM-first with a heuristic fallback, so it degrades gracefully if no LLM is configured for the atomizer path. Always review an atomized rule set before publishing it.
Python SDK — upgrade with pip install --upgrade smartflow-sdk for v0.5.0. Existing SDK code is unaffected; @smartflow_tool and smartflow_task are additive.
No database migrations. All 1.10 additions are additive — existing VAS logs, policies, and audit chain entries are fully compatible; control/policy ID traceability populates on new requests going forward.