Responsible AI governed end to end

BRANCLE TECHNOLOGIES LLC (“Brancle,” “we,” “us”), 212 N. 2nd St. STE 100, Richmond, KY 40475, builds agentic workflows that operate inside your systems under joint governance. Five principles run every engagement: human accountability for consequential actions; least-privilege tool and data access by default; no autonomous high-stakes action without independent approval; measurement against the baseline captured at discovery; and no use of your content to train public models.

This page states our standard method. Threshold numbers, reviewer names, retention, and the model and tool stack for your workflow are fixed in your signed statement of work (“SOW”), which controls on conflict about that workflow. Our Terms of Service (Policy #2) sets the underlying duties; our Security page (Policy #4) sets safeguards; project personal-data handling is governed by our DPA (Policy #6).

This policy governs agentic workflows Brancle designs, builds, deploys, or supports for clients, including:

• discovery scoring of opportunities by hours, cost, risk, and feasibility;
• reasoning design, tool and data scoping, guardrail and threshold calibration;
• backtesting, acceptance, phased rollout, monitoring, and tuning;
• any future interactive agent demo on brancle.com, which will carry its own just-in-time disclosure naming the provider, data sent, retention, and review controls (per Privacy §6) and link here.

It does not govern changes you operate yourself after handoff unless under a support SOW, nor does it replace your counsel’s approval for regulated uses (see Terms §§6–7).

“Mid-level critical thinking” means bounded judgment within your rules and tools — never strategic, legal, or safety-critical autonomy. Anything with legal or similarly significant effect requires an SOW that expressly permits it plus your counsel’s written approval, per Terms §7.

Placement follows stakes × reversibility × volume, decided jointly at design and recorded in the SOW. Rollout promotes shadow → supervised → autonomous only after meeting the SOW’s acceptance thresholds on live volume; any tier can be demoted on drift, incident, or reviewer request.

Every workflow ships with layered controls: input validation, tool and data allowlists, policy checks, output validators, and confidence floors with escalation on miss. We publish method plus illustrative ranges here; your exact floors are calibrated jointly and versioned in the SOW so they never become accidental warranties.

Thresholds are retuned as edge cases surface in the wild, with author, reason, and effective version recorded. Model, tool, or API changes by third parties trigger re-validation before promotion continues.

Qualified reviewers — named in the SOW with role, coverage, and backup — approve Tier 2 drafts and decide Tier 3 escalations. An escalation package always includes the context read, rule or policy cited, confidence and why it missed (if applicable), recommended action, deadline, and one-click approve / reject / reassign. Logs support review; they do not replace it.

• Backtest on your history: run against real historical cases before anything goes live, measuring hours, accuracy, throughput, and cycle time versus the discovery baseline.
• Acceptance in the SOW: pass/fail criteria, measurement method, and reporting cadence fixed up front. If the workflow doesn’t move the agreed number, it doesn’t promote — iteration proceeds under change control, not as uncapped rework.
• Regression on change: model, prompt, tool, or API change re-runs the suite; promotion pauses until green.

Live workflows report against the day-one baseline with floors for accuracy and throughput. We monitor for drift (behavior, volume, exception rate), tune and harden as edge cases appear, and review quarterly with a next-workflow roadmap. Every action logs actor, action, timestamp, input/output reference, confidence or rule applied, and approver where review applied — reviewable by you, retained per the SOW and DPA.

Fallback paths are defined per workflow (retry once, queue, escalate, roll back to the last good version) so a failing step degrades to human handling instead of silent automation.

Contact, booking, and project content is never used to train public foundation models and is never fed to models for advertising profiling. This Site itself runs no AI agent that decides qualification, pricing, or hiring about your inquiry — a human reads and responds.

Client workflows call foundation models, tools, and APIs named in your SOW (provider, purpose, data sent, and retention fixed there; vendor roles summarized in Policy #7 Subprocessors). Because third-party model behavior can shift across versions, SOWs pin versions where feasible and require re-validation on change before autonomous tiers resume.

We decline or require redesign where a request would mean:

• autonomous high-stakes action without qualified human approval (see Section 3);
• solely automated decisions with legal or similarly significant effect without SOW permission plus counsel approval;
• uses requiring certifications we do not hold, or unlawful, discriminatory, deceptive, or surveilling uses;
• processing beyond the data rights, retention, or residency constraints you disclosed at discovery.

Screening runs at discovery via hours/cost/risk/feasibility scoring plus a prohibited-use check; failures are documented with the reason and, where possible, a safer scoped alternative.

We update this page as our method, models, tools, or law evolve by posting the revised version with a new Effective date; SOW-specific controls change only by written change control. The current version always lives at https://brancle.com/responsible-ai and companions Terms (Policy #2), Security (Policy #4), and the DPA (Policy #6).

This page describes our standard method and does not constitute legal advice or a warranty of any outcome. Workflow-specific protections are set in your SOW and DPA (Policy #6), which control on conflict about project work.