Skip to main content
Security-engineered, product-minded consultancy for agentic AI

Turn promising AI agent workflows into controlled, testable, economically justified business operations.

Led by principals with a background in cyber security and product development. We treat an AI agent the way a security team treats any powerful new system: define what it may do, test how it fails, and keep a named human accountable.

Discuss a Workflow Assessment Read the methods

AI Workflow Assessment

Decision dossier RFP-12

RFP response workflow · owner: Bid Desk

Objectives

Raise win-rate on qualified RFPs — without unsupervised submissions

Boundaries

Drafts and assembles; never submits; customer data stays in-tenancy

Authority

Reviewer sign-off before send; pricing or PII escalates to a named owner

Release pipeline

Intake
Agent drafts
Review gate
Submit

Illustrative dossier for a synthetic RFP workflow — shown to demonstrate the assessment's structure, not a client engagement or measured result.

Run by people who have secured and shipped real products.

Cyber security
Experience of leading incident-response and application-security programmes for business systems handling millions of transactions.
Product development
Full lifecycle ownership of building and operating B2B products where release gates and unit economics were standing board-level topics.
Published work
Writing on AI, security workshops, university partnerships, published public guidance on databreaches, and participation in security focused events.
How we work
Principals perform every engagement personally.
Meet the principal

The business problem

Agents can make plausible decisions — and still be wrong for your business.

Agentic AI is moving from answering questions to taking actions: drafting, sending, deciding, spending. Four failure patterns show up again and again in workflows that looked fine in a demo.

Plausible is not permitted

An agent can produce a fluent, confident decision that its permissions never allowed — a submitted proposal, a priced quote, an email to a customer — and nothing in the output tells you it stepped outside its authority.

Evidence is not guaranteed

Generated answers cite sources that may not support them, or omit the fact that the model guessed. Downstream, someone makes a business decision on evidence nobody verified.

Cheap per call, expensive per year

A workflow that retries too eagerly, over-loads context, or routes every step to the strongest model quietly converts a per-call cost into an operating cost nobody budgeted.

Failure to escalate

When confidence drops or a policy requirement is not met, the workflow should hand a named human a decision — not smooth over the gap and keep going.

None of this is new. Security teams learned to assume untrusted input, over-privileged accounts and silent failure decades ago. Product teams learned that a demo is not a release. Agentic AI has to relearn both, faster.

Why a better prompt is not an operating-control strategy

Instructions in a prompt are suggestions the model usually follows. They are not boundaries, because nothing checks them — and they say nothing about what happens when the model is wrong anyway. Selecting a stronger model raises the ceiling of quality but not the floor of behaviour. Real control over a business workflow comes from engineering: explicit permissions, machine-checkable policies, tests over the workflow's properties, decision economics for model and review choices, and named humans who own the material decisions. That is the discipline we bring — and it is auditable, which a prompt is not.

What the firm does

We do not sell AI adoption. We make specific workflows defensible.

Our goal is not to maximize AI usage. It is to maximize useful business outcomes while making risks and operating costs visible.

Assess AI workflows

Take one consequential workflow — an RFP response, a materiality review, a support triage — and examine its permissions, evidence quality, failure handling, release criteria and unit economics against a measured baseline.

Design controls, evaluation and decision policies

Define what the workflow may do, how its outputs are tested, when it must escalate, and how choices between models, tools and human review trade off quality, cost, latency and risk.

Develop reusable patterns and implementation plans

Provenance, bounded tool use, draft–verify–approve, risk-tiered routing, retries and idempotency: patterns your engineers can implement again on the next workflow, not one-off fixes.

Reason about cost, assurance and governance

Make the operating cost visible, the evidence standards explicit and the residual risk stated — so leadership can decide to deploy, redesign or walk away with the trade-offs in the open.

Flagship service

The AI Workflow Assessment: one real workflow, one written decision.

Ten business days, one consequential workflow, four questions. The output is a decision dossier — current and target state, risks and controls, an evaluation plan, an economic model, and a clear go / no-go recommendation you can defend to a board.

Safety

What can this workflow be trusted to do without a human?

It establishes: Boundaries, review gates and escalation paths that are enforced, not suggested — and the failure modes testing has actually covered.

It does not give: Not a guarantee of safe operation in all conditions; untested cases remain untested.

Security: threat model first

What can it be tricked or manipulated into doing?

We model the workflow as an attacker would: untrusted inputs, tool blast radius, data boundaries, prompt injection and abuse cases.

It establishes: Permission gaps, data boundaries, prompt-injection exposure and the blast radius of each tool it can call.

It does not give: Not a penetration test or a certification; findings are scoped to the assessed workflow.

Efficiency

What does each accepted task actually cost?

It establishes: Model calls, context size, retries and compute per unit of useful output — measured against a baseline, not estimated.

It does not give: Estimates depend on your volumes, exception rates and integrations; we state the assumptions.

Economics: product decision

Is this workflow worth operating at all?

A build, redesign or stop decision with the acceptance criteria a product owner would demand.

It establishes: A decision model with quality, cost and risk explicit — supporting pilot, redesign, or a well-argued 'no'.

It does not give: No guaranteed savings; the output is a defensible decision with stated residual risk.

Scope at a glance

One defined workflow · 10 business days · up to three stakeholder interviews · up to 10 redacted artefacts · indicative fee ₹1,50,000 plus taxes. Full scope, exclusions and commercial terms on the service page.

See the full assessment

Also from Banbreach

AI Incident Response and Forensics (DFIR for AI)

When an agent does something it shouldn't: a scoped engagement that reconstructs what it did, what it was permitted to do, and which controls failed — with confidence levels and stated limits. Before anything goes wrong, an Incident Readiness Review makes the workflow investigable by design.

The incident response service

Live incident?

Live incidents use a monitored urgent channel, not the general form. Coverage, first-response time and what the service does not cover are stated exactly — never implied.

Urgent incident contact

Methods

The working methods behind every assessment.

Each method has a page of its own — what it addresses, what it can establish, and, just as important, what it cannot.

Worked demonstration — illustrative

ESG materiality intelligence: AI-assisted evidence, not automated judgment.

A bounded, observable workflow we use to demonstrate the methods in practice: a double-materiality review where the agent gathers and structures evidence, and every material decision stays with a named analyst. All documents, figures and outputs below are synthetic.

Inputs

  • A corpus of disclosures and policy documents (synthetic, publicly styled)
  • A materiality framework with thresholds and definitions
  • A permitted-actions list: the workflow may collect, cluster and cite — never rank, score or recommend

Controls

  • Bounded retrieval: only the approved corpus, with provenance recorded per fragment
  • Every claim in the draft output must carry a citation; uncited sentences are stripped and logged
  • Findings above a threshold are quarantined for analyst review — the workflow cannot escalate them onward

Decisions

  • Agent proposes topic clusters and evidence mappings; the analyst accepts, merges or rejects
  • Materiality scoring stays human: the workflow supplies evidence, not judgment
  • Disagreement between clusters is surfaced as a question for the analyst, not averaged away

Stated limits

  • Coverage is limited to the corpus provided; absence of evidence is not evidence of absence
  • Citation validity is checked structurally, not semantically — a source can be real but misread
  • This is an illustrative demonstration with synthetic data, not a client deployment

Sample output (synthetic)

Cluster Evidence fragments Citations valid Routed to
Water stewardship 31 29 Analyst review (threshold)
Supply-chain labour 18 14 Analyst review — 4 fragments re-sourced
Board diversity 7 7 Below threshold — logged, not escalated

Illustrative figures only. A real run reports the same structure with measured numbers, failure cases and the evaluation plan that produced them.

See how each control is built

Interactive tool

The Workflow Readiness Check

Seven questions, answered honestly — about a minute each. At the end you will see which of the assessment's four questions — safety, security, efficiency, economics — is least answered in your organisation today.

Indicative and anonymous: your answers stay in your browser and are never submitted.

Insights

How we think about agentic AI.

Notes and worked thinking on workflow assessment, design patterns, decision economics and governance — written to be useful, not to chase keywords.

All insights

Who you will work with

The principal does the work — not a sales layer.

Banbreach is a boutique practice. You deal with the principal who does the work, from the first conversation to the delivered dossier.

Suman Kar, Principal

Principal — security engineering and AI workflow assurance

Kolkata, India (IST)

Background

  • Cyber security. 20 years in product security and incident response for firms; handled multiple databreaches; expert witness.
  • Product development. 15 years of product experience from device drivers, to desktop apps, as well as SaaS with installed userbase in millions.
  • MBA (Strategy), HEC-Paris | ex-Adobe | BE (Computer Science), IIEST Shibpur, India; Bayesian Data Analysis, Aalto Univerity.

At Banbreach

Leads AI Workflow Assessments and the DFIR for AI service; first responder on the urgent channel.

LinkedIn profileWriting

Measured results, not estimates presented as results

Every number we publish is either measured, sourced, or labelled as an estimate or hypothesis — and we keep those categories visibly separate.

Limitations stated where they could be missed

A test suite is evidence over tested cases. A formal proof covers its stated property under its assumptions. Neither makes a whole business process safe — and we say so where it matters.

No guarantees sold

We do not sell certainty. We sell the ability to make a defensible decision: deploy, redesign, or decline — with the residual risk in writing.

About the firm

Bring us one workflow. Leave with a decision you can defend.

Start with a 30-minute scoping conversation with a principal. If an assessment isn't the right next step, we'll say so.

Book a scoping conversation Discuss a Workflow Assessment

A workflow that saves time but introduces unacceptable risk is not a success. That is the standard we assess against.

The Workflow Readiness Check

Seven questions, answered honestly. You’ll see which of the four assessment questions — safety, security, efficiency, economics — is least answered in your organisation today.