AI Workflow Assessment

One workflow, ten business days, four questions — safety, security, efficiency and economics. This page answers the three that matter before you commit: what you receive, what we need from you, and what decision it helps you make.

Last reviewed: October 2026.

Who the assessment is for

Organizations running — or about to run — an AI agent workflow that has business consequences.
Technology leaders who need to show the workflow is controlled, not just impressive.
Risk and compliance leaders who must sign off on an operating workflow they can audit.
Transformation sponsors deciding whether a pilot deserves a budget — or a redesign, or neither.

Typical problems it investigates

"The output looks right, but nobody can say what it is allowed to do."
"We found out it emails customers only when something goes wrong."
"Model spend tripled last quarter and nobody can explain it line by line."
"Every failure gets a prompt patch. We have no tests and no baseline."
"We need to say yes or no to this pilot — defensibly — by next month."

Who does the assessment

Every assessment is carried out by a Banbreach principal, not delegated. Our background is in the two disciplines this work depends on:

Cyber security. [VERIFY: one concrete sentence — e.g. years leading application and cloud security, threat modelling and red-teaming] We bring the attacker's view: what an agent can be tricked into doing, what a compromised tool can reach, and where permissions exceed need.
Product development. [VERIFY: one concrete sentence — e.g. led design, launch and operation of N products from concept to production] We bring the product owner's view: what "done" means, what it costs per accepted unit of work, and when to ship, redesign or stop.

Meet the principal

Scope and commercial terms

Scope	One defined AI or agentic workflow
Duration	10 business days
Interviews	Up to three stakeholder interviews
Artefacts	Up to 10 representative, non-sensitive or redacted artefacts
Deliverables	Current-state and target-state maps, threat model and control analysis, evaluation plan, acceptance criteria and release gates, economic model and recommendations
Exclusions	Production implementation · Penetration testing · Certification · Legal opinions · Guaranteed savings

The fee is indicative and subject to engagement scoping. Economic estimates depend on workflow volume, loaded labour cost, quality requirements, exception rates, integration costs and operating costs — we state every assumption rather than promise savings.

Assessment, not penetration test. We apply the same adversarial method — threat modelling, abuse-case analysis, tool-boundary review — but we do not attack your production systems, and the output is not a certification. If you need hands-on penetration testing, we will say so and help you scope it.

The engagement process

1

Charter and boundaries

We agree the workflow in scope, the decisions the assessment should inform, and the boundaries of the analysis. You name the workflow owner; we agree what is out of scope and why that is safe to leave out.
2

Current-state mapping

Interviews and redacted artefacts in, a map out: every step, permission, tool call, data boundary, review gate and failure path as the workflow actually runs today — including the informal workarounds people don't put in diagrams.
3

Evidence and testing

We examine what the workflow's claims rest on. Security: we threat-model the workflow — untrusted-input paths, prompt-injection exposure, tool permissions and the blast radius of each tool call. Evidence: we find where outputs are verifiable and where they are guesses. Testing: we run the properties that matter, including adversarial and failure-mode checks, on redacted or synthetic data.
4

Economics

A measured baseline of the current path (human-and-compute) and a model of the agentic path: volumes, loaded cost, exception rates, model spend, integration and operating cost — with every assumption stated.
5

Decision dossier and readout

Current and target state, the threat model and controls, the evaluation plan, the economic model, and a recommendation: pilot, redesign, or decline. The readout is a product decision, not a presentation — it comes with the acceptance criteria a product owner would demand before launch, and we walk your team through what would have to be true to change it.

Deliverables

Current-state and target-state maps — the workflow as it runs, and as it should.
Threat model and control analysis — permission gaps, abuse cases, evidence weaknesses and failure paths, with severity.
Evaluation plan — what to test, which properties, which acceptance criteria.
Acceptance criteria and release gates — what "good enough to ship" means, in testable terms.
Economic model — baseline, agentic-path costs and stated assumptions.
Recommendations — pilot, redesign or decline, with the decision criteria.

Want to see the shape of the dossier before committing? Read a synthetic sample dossier — illustrative data throughout, not a client engagement.

Inputs required from you

One consequential workflow, named and bounded.
Up to three stakeholder interviews (60–90 minutes each).
Up to 10 representative artefacts — redacted or synthetic where sensitive.
Volume and cost figures for the baseline, where available.
A workflow owner empowered to act on the findings.

After the assessment

Pilot

We define the pilot's entry and exit criteria, the instrumentation it needs from day one, and the review gate that decides whether it grows. Your team runs it; we stay available for the first readout.

Redesign

A principal leads the design of the target state — boundaries, permissions, policies and release gates — and hands your engineers (or a partner you choose) a specification to build. We do not sell the build.

Decline

You keep the dossier: the threat model, the economics and the written reasons why not. That record is often the most valuable output — it is the argument you replay when the same workflow returns with a better proposal.

The recommendation names the path; these are the ways we can help you walk it. We are as valuable when the answer is "no" as when it is "go".

Security and confidentiality

How we protect your material

We practise the data minimisation we recommend: non-sensitive or redacted artefacts only, no production credentials, no customer data. Working material stays in access-controlled locations, is deleted at the end of the engagement (or on request), and a mutual NDA is available before any material is shared. Findings reference your workflow only with your written permission.

Why you can trust the security analysis

The assessor's background: [VERIFY: relevant credentials or track record, with links] . Our own handling follows the practices we recommend: minimal data, no production credentials, access-controlled working material, deletion at close.

Prefer to talk first? Book a 30-minute scoping conversation — the quickest way to find out whether an assessment is the right next step.

Discuss a Workflow Assessment

Tell us about one consequential workflow. We will reply with a concrete next step — usually a short call to confirm the workflow is a good fit for an assessment. (Reply times are stated on the contact page.)