Skip to content
Aelsha Innovations

Agentic-AI consulting

Agentic AI consulting for teams taking agents to production.

If you run a business or a technology team and want AI to take on real work, not only answer questions: before an agent acts in your systems you need to know what it may touch, who approves the steps that matter, and how you will know it still works next month.

  • A clear answer on where AI agents are worth using, and where not
  • A first workflow built, where the assessment says build, with a person approving what matters
  • Reports, code and tests handed to you, with an owner named on your side
Talk to us about your projectStart here: the readiness assessment →

Tell us what you want to automate. We reply by email.

Aelsha Innovations provides agentic AI consulting for organisations moving AI agents from pilot to production. Work spans six areas: readiness, architecture and integration, build, evaluation and observability, security, and governance. We recommend starting with a fixed-scope readiness assessment, with its scope and schedule agreed at kick-off; the client keeps the written output.

Standards: we map our deliverables to named public standards and produce the evidence for them.

Technical detail: standards, methods and what we check ↓
Terms used on this page
AI agent
software that reads a request, decides what to do and acts in your systems.
Agentic AI
AI that takes actions, not only gives answers.
Pilot / production
a pilot shows it can do the task once; production is when it runs on real work every day.
Tracing / observability
a record of every step an agent takes, so you can see what it did and what it cost.
Red-teaming
attacking a system, with written permission, to find weaknesses before someone else does.
Governance
the rules for what may run, who approved it, and the evidence kept.

The consulting map: an agent system, with each of our six areas of work placed where it applies.

Gold marks the one step where a person approves a consequential action before it runs.

Six areas of work, across the life of an agent system

We work on six areas of an agent system. Each is a fixed-scope piece of work that ends in a written deliverable, and each names the public standard its output is mapped to. You can start with any of them; we suggest starting with readiness.

  • Strategy and readiness

    Find where agents are worth using in your business, and whether your systems, data and people are ready to run them.

  • Architecture and integration

    Connect agents to the systems you already run through open protocols, with each agent given its own narrow access. MCP (Model Context Protocol) is an open standard for connecting agents to tools and data; A2A (Agent2Agent) lets agents call one another.

  • Build and productionisation

    Take one workflow from pilot to production, with a person approving every consequential action, a tested way to roll back, and an owner on your side.

  • Evaluation and observability

    Prove the system behaves, notice when it stops, and see what each run costs. Tracing records every step an agent takes; an evaluation set is a fixed collection of real examples the system is tested against before each release.

  • Security and red-teaming

    Map how an agent could be attacked or misused, then test it the way an attacker would. Prompt injection is text in a document or web page that tries to give the agent new instructions; red-teaming means attacking a system, with written permission, to find weaknesses first.

  • Governance and compliance

    Make the system auditable under the rules you answer to: what runs, who approved it, and what evidence exists.

Strategy and readiness: what you receive and what it maps to

The work coming in: which tasks are worth giving to agents, and whether your systems and data are ready.

What you receive

  • A scored list of candidate use cases
  • A review of data and system readiness
  • A build, buy or platform memo
  • A 90-day roadmap
  • A pilot, defer or stop decision for each use case

Mapped to

  • AELSHA™ ASTRA maturity levels L1 to L5
  • NIST AI 600-1, Generative AI Profile · Source: NIST

Not included: Building anything, legal opinions and tool procurement.

Readiness assessment →

Architecture and integration: what you receive and what it maps to

The tool layer: agents reach your systems and data through MCP servers, without a rewrite.

What you receive

  • A reference architecture
  • MCP server design, build and security audit
  • A2A design for agents that work together
  • An access model: each agent its own identity, with short-lived, narrowly scoped tokens
  • A plan for making older systems reachable by agents

Mapped to

Not included: Rewriting your core systems; we connect at agreed points.

MCP engineering →

Build and productionisation: what you receive and what it maps to

The agents and the approval step, taken from pilot to production as one workflow.

What you receive

  • A production agent, or a set of small single-purpose agents working together
  • An approval design: which actions a person approves, and what they see
  • A rollback path and an off switch
  • A release pipeline with evaluation gates
  • Runbooks and a signed handover

Mapped to

Not included: Model training, and new workflows beyond the one agreed.

Composable agents →

Evaluation and observability: what you receive and what it maps to

The trace lane underneath: every step recorded, measured and costed.

What you receive

  • An evaluation harness and a baseline on your own traffic
  • OpenTelemetry tracing for every agent and tool call
  • Drift and regression alerts
  • A cost model per task
  • Testing before a model is swapped

Mapped to

  • OpenTelemetry GenAI semantic conventions (Development status) · Source: OpenTelemetry
  • AELSHA™ ASTRA: five signal pillars across three planes

Not included: Accuracy or savings promises; targets are measured and agreed with you.

How AELSHA™ ASTRA frames this work →Framework · delivered through engagements

Security and red-teaming: what you receive and what it maps to

The boundary every tool call crosses: who the agent is and what it may touch.

What you receive

  • An agent threat model
  • Prompt-injection and tool-misuse testing
  • An MCP security review
  • Ranked findings, a remediation plan and a retest checklist

Mapped to

Not included: Claims that a system cannot be attacked, and testing of systems you cannot authorise in writing.

Agent security →

Governance and compliance: what you receive and what it maps to

The layer over everything: what runs, who approved it, and the audit trail.

What you receive

  • An AI inventory with risk tiers
  • A review of personal-data flows and consent
  • Alignment with the regulation and sector guidance that apply to you, such as the EU AI Act or Indian sector guidance
  • An audit-trail specification
  • A gap register with owners and dates

Mapped to

Not included: Certification, legal advice, and decisions on how a law applies to you.

How engagements work: Validate, Scale, Operate

Engagements run in three steps. Each step has a fixed scope and ends with something you keep, and you can stop after any of them. The usual starting point is a readiness assessment.

  1. 01

    Validate

    Decide where agents earn their keep, before you commit to a build, with a readiness assessment, an MCP audit, a threat-model review or an evaluation baseline. How long each takes depends on its scope, and we agree the schedule with you at kick-off.

  2. 02

    Scale

    Build the chosen workflow and harden it for production, one agent at a time. At a decision point agreed at kick-off, early in the build, you choose to continue, re-scope or stop.

  3. 03

    Operate

    Run and improve it from real results, by arrangement, or hand it to your team with the runbooks, evaluation sets and dashboards to do so.

Start here: how a readiness assessment works →

Four ways to engage

Every engagement has a fixed scope and a fee agreed in writing before work starts, and every one is delivered with AELSHA™ ASTRA, our framework for taking AI agents to production and keeping them there: lifecycle gates, tracing, evaluation and governance. You choose the model that fits the question you need answered.

  • Fixed-scope assessment

    A short piece of work that ends in a clear written answer. Four entry offers: a readiness assessment, an MCP server security audit, an agent threat-model review and an evaluation baseline. You keep every artefact.

  • Fixed-bid pilot to production

    One workflow built to production for an agreed fee, on a scope and schedule agreed at kick-off. An early decision point lets you continue, re-scope or stop, with the work so far handed over.

  • Monthly operations or advice

    Managed AgentOps (ongoing evaluation and improvement of agents already in production, with your team) or a fractional Chief AI Officer (senior advice each month). Offered by arrangement, with scope and response times agreed in the contract.

  • Expert sessions

    Short, fixed-scope expert work such as an architecture or code review, an MCP server build, or agent debugging. Each session is scoped and quoted before it starts; there is no public hourly rate.

Go deeper

Five parts of this work have their own pages, with the detail a technical reviewer will want. If you are not sure where to begin, begin with the readiness assessment.

  • Start here

    Readiness assessment

    Where agents are worth using, what it takes to run them, and a pilot, defer or stop decision for each workflow.

  • Specification

    Composable agents

    Small, single-purpose agents with typed contracts, scoped tools and their own evaluations, connected by MCP.

  • Agent security

    Ten threats to AI agents, the control for each, and how threat-model reviews and red-teaming test for them.

  • MCP engineering

    MCP servers built around your existing systems, and audited against the MCP security best practices and the OWASP MCP Top 10.

  • AELSHA™ ASTRA

    The framework every engagement is delivered with: a six-phase lifecycle with quality gates, observability, evaluation and governance.

What you keep at the end

Handover is a deliverable, not a courtesy. At the end of every engagement the artefacts are yours, and a person on your side is named as their owner.

  • Reports and decisions

    The readiness report, threat model or audit findings, with the evidence behind them.

  • Code and configuration

    Repositories, agent contracts and MCP servers, transferred to you.

  • Evaluation sets

    Versioned test cases and a harness you can run on every release.

  • Dashboards and alerts as code

    Tracing, dashboards and alert rules you can deploy and change yourselves.

  • Runbooks

    How to start, stop, roll back and handle an incident.

  • A named owner

    Someone on your side who has run a release unaided before the handover is signed.

Common questions

Short answers on price, results, ownership, compliance and data. In short: a fixed scope agreed in writing, no promised returns, artefacts that stay with you, evidence rather than certificates, and work that can run on your own infrastructure.

Do you publish prices?

Not on this page. Every engagement has a fixed scope and a fee agreed in writing before work starts. Tell us what you want to automate and we will reply with a proposed scope.

Can you promise a return on investment?

No. Results depend on the workflow and the data. What we offer is a fixed scope, an early decision point in a build, and measurements agreed with you in advance.

Will we depend on you afterwards?

Only if you choose to. The handover is part of the work: runbooks, evaluation sets, dashboards as code and a named owner on your side. Ongoing operations are optional and can be ended.

Do you certify compliance?

No. We map your system to the standards you name and produce the evidence an auditor will ask for. We do not issue certificates, and how a law applies to you is a decision for you and your counsel.

Which frameworks and protocols do you use?

We are framework-agnostic and standards-first. We choose the agent framework per client and per agent, for example LangGraph, Google ADK, the OpenAI Agents SDK, the Claude Agent SDK, Microsoft Agent Framework, CrewAI or Strands Agents. Connections use MCP and A2A, and tracing uses OpenTelemetry.

Our data is sensitive. How is it handled?

Engagements can run on your own infrastructure, with keys you control. Our trust page describes how we handle client data.

How we handle security and data →
What do you not do?

No chatbot reselling, staffing or resale of third-party platforms. No hourly billing for entry offers. No agent that replaces a required human decision: irreversible, financial and external-facing steps keep a person approving them. No testing of systems you cannot authorise in writing, and no publishing of client names or results without written approval.

Further reading

Tell us what you want to automate

Describe the workflow and the systems involved. We reply by email with a suggested first step and a proposed scope.

Talk to us about your project