Will closegate replace my analysts?

No. closegate moves the repetitive part of close + reconciliation off your analysts' plates so they spend time on judgment calls (real exceptions, FX commentary, audit prep) instead of tying-out spreadsheets. Most pilot teams reallocate 30–40% of close-week hours toward higher-value work — the headcount math is reallocation, not reduction.

Will my external auditor accept it?

The audit narrative is the same as for any controls system: identity, segregation of duties, materiality, escalation, evidence. closegate's append-only SQLite audit log carries the verbatim policy clause text on every event — auditors quote that text verbatim. We ship SOC 2 + SOX 404 + PCAOB AS 1215 control mappings with file.py:line citations.

What if the LLM hallucinates a confirmation?

It can't reach production-confirmation. The policy gate is server-side, between the LLM and the GL. The LLM's tool call passes context; the gate decides Allow / RequireHumanApproval / Deny based on materiality, sensitive accounts, and SoD rules — none of which the LLM can override from the prompt.

What does the pilot look like?

Typical shape: one workflow (close, recon, or AP), 6–12 weeks, against a snapshot of last quarter's GL data. We seed your closegate instance from the snapshot, port your existing policy thresholds into policy.yaml, and run the deterministic eval harness before any LLM calls. Then you graduate to a parallel run on the next close.

closegate for finance teams — controllers, FP&A, accounting

Your team can recognize 95% of GL ↔ subledger matches in seconds. The remaining 5% — the exceptions, the above-materiality cases, the cross-entity intercompany clearing that needs FX commentary — is where your time goes. closegate is the layer that lets the AI agent safely handle the 95% so your team owns the 5% with judgment.

What changes after closegate

Today (manual)	Vendor "agentic" black box	closegate
Controller ties out GL/SL line-by-line in spreadsheets	Agent auto-matches; you trust vendor's RCA if something breaks	Agent proposes matches; policy gate auto-confirms below materiality, routes above to your HITL inbox
Above-materiality match → controller manually escalates	Above-materiality match → vendor decides routing internally	Above-materiality match → server-side rule fires, your named approver gets a Slack/Teams card with the verbatim policy clause
Audit prep = 2 weeks of spreadsheet screenshots	Audit prep = vendor PBC bundle (their format, their definitions)	Audit prep = closegate-engine audit-evidence-export → one ZIP with the SQL audit log, actor identities, and policy versions
You can't reproduce what happened in close 3 months ago	You can ask the vendor; they'll get back to you	git log on policy.yaml + replay the audit log — anyone with the data can reconstruct any decision
Fraud risk = manual vigilance	Fraud risk = vendor's threat model (closed)	Fraud risk = sensitive-account routing forces HITL on cash, legal accruals, vendor bank changes, regardless of materiality

The four pieces, in finance-team language

1. Policy gate — the chokepoint

Every state-changing action (confirm match, approve invoice, submit payment run) passes through one transactional gate. The gate decides: Allow (mutate + write audit), RequireHumanApproval (route to HITL inbox), or Deny (write violation event, raise to controller). There is no second confirmation API anywhere in the system.

2. Audit log — the auditor's friend

Append-only SQLite table. BEFORE UPDATE and BEFORE DELETE triggers refuse the operation at the database layer — not in application code, so an insider with code-write access can't quietly tamper. Every blocked event carries the verbatim policy clause text (auditors quote it verbatim) and a JSON-pointer to your policy.yaml source.

3. HITL approval inbox — your queue, not the vendor's

When the gate routes a state-change to HITL, the approval lands in a web inbox + a Slack/Teams card with deeplink-back to the canonical web record. The approver confirms via a different actor identity; the gate enforces SoD server-side. The LLM cannot impersonate a human approver, no matter what the chat history says.

4. Eval harness — your audit committee's friend

Four dimensions run continuously: matching accuracy, policy enforcement, adversarial robustness, latency. Output is a reproducible JSON artifact your audit committee can read. Pairs with the soc2-monitor nightly CI job so you have SOC 2 CC4.2 evidence on autopilot.

Is closegate right for our team?

Strong fit:

You're piloting an AI agent against close, recon, or AP and your internal auditor is the bottleneck
You self-host or want to (Docker, Kubernetes, fly.io)
You're scoping a SOC 2 Type 2 readiness program and want operational evidence on autopilot
You're multi-entity / multi-jurisdiction and need per-entity materiality overrides
You've been quoted $100K+/yr for a vendor and want to evaluate the open alternative

Probably not the right call yet:

You haven't done a manual close in 2026 — start with the workflow before adding the agent
You want a turnkey SaaS UI with white-glove onboarding for non-technical staff — closegate is a reference architecture, not a managed product
Your policy.yaml shape is "we'll figure it out as we go" — closegate works best with explicit materiality + SoD rules

What a pilot looks like

Week 1: Clone the repo, run the SaaS seed pack locally, walk your CFO through the workspace UI. (Tutorial: install in 60 seconds.)
Week 2–4: Port your existing policy thresholds to policy.yaml. Test against the 11 starter jurisdictional libraries.
Week 5–8: Wire a real ingestion adapter (Stripe, QuickBooks, NetSuite, etc.). Run the deterministic eval against your data.
Week 9–12: Parallel-run alongside your existing close for one full cycle. Compare outputs. Surface the deltas to internal audit.

The 30/60/90 pilot plan with stakeholder briefings, control-mapping walkthroughs, and a security review template lives in the technical docs. The CFO pitch deck and ROI calculator live on this site — For CFOs and ROI.

Put an AI agent in your close cycle. Keep the controls.