Menu
ENESPT-BRPT-PT

Consulting · AI agents in production

Your agent works in the demo. We take it to production.

In two weeks we review your AI agent's architecture and hand you a concrete plan to make it reliable, secure, observable and cheaper to run. A fixed price, agreed before we start.

Book a 45-minute diagnosis What we review

The problem

Agents fail between the pilot and production

A prototype that shines in a demo does not survive real traffic, real data and real users. The problems show up in production, when they already cost money or trust.

  • Cost grows with every user: the same context is paid for on every call, and the same questions are answered again and again.
  • The agent's tools have more permissions than anyone reviewed.
  • When something breaks, nobody knows at which step or why: there are no traces.
  • A traffic spike or a crashed process loses requests, or handles them twice.

What we review

Six fronts, one report

We review your agent the way a platform team would before letting it into production.

  • Reliability

    Queues, retries, idempotency and durable state: what happens when a process dies in the middle of a conversation.

  • Cost and latency

    Prompt caching, response reuse, deduplication and model choice, with a savings estimate on your own volumes.

  • Tool security

    Who can run what, least privilege, human approval for anything sensitive, and an audit log.

  • Observability

    End-to-end traces, token, cost and error metrics, and alerts that warn you before your users do.

  • Answer quality

    Evaluations on your real cases, limits on what the agent may do, and how it behaves when it doesn't know.

  • Operations

    Deployment, usage limits, prompt and model versions, and who responds when something fails.

Isometric model of an AI agent's architecture in production: alert sources, a queue, a correlation function, databases, the agent's server, the language model and the analyst's console.
Illustrative example: an alert-triage agent with correlation, response caching and prompt caching, each piece in its place.

How it works

Two weeks, a fixed price

You know what it costs and what you get before we start.

  1. 01

    A 45-minute call

    You tell us what your agent does and where it gets stuck. If the review won't help you, we say so.

  2. 02

    The review

    With read-only access to your code, architecture and metrics, plus two or three interviews with your team.

  3. 03

    The hand-off

    A report with prioritised findings, and a session to walk your team through it and answer questions.

  4. 04

    Implementation (optional)

    If you want, we take the plan to production with you in 4 to 6 weeks.

What you get

  • Findings prioritised by impact and effort.
  • A diagram of the current architecture and of the target one.
  • A monthly cost estimate and the savings from each improvement.
  • A plan for tool permissions, approvals and auditing.
  • An observability plan: traces, metrics and alerts.
  • A phased roadmap, marking what your team can do on its own.

Experience

What we have built

  • Security alert triage

    Vervux Rock: deduplication of alert storms, a decision cache, rules before AI, human approval below 70% confidence, idempotent actions and a trace for every run.

    See Vervux Rock

  • Company knowledge

    Internal documents turned into answers with citations you can check.

  • Customer service

    Technical questions answered from the catalog, with a hand-off to a person.

FAQ

What technical teams ask

Do we need the agent already built?
Yes: a working pilot or prototype, even with few users. If you are still choosing what to automate, start with our consulting's opportunity diagnosis.
Which technologies do you work with?
The ones you already use: AWS, Google Cloud or another cloud; LangGraph or your own orchestration; and models from Anthropic, Google or OpenAI. The review doesn't make you switch providers.
What access do you need?
Read-only access to the code, the configuration and aggregate metrics. We don't need your customers' personal data, and we sign a confidentiality agreement before starting.
How much does it cost?
A fixed price based on the size of the system, agreed after the 45-minute call and before we start. No extra hours, no surprises.
What if we want you to implement it?
That is the optional phase: we take the plan to production with you in 4 to 6 weeks and leave your team running it.

Contact

Book a 45-minute diagnosis

Tell us what your agent does and where it gets stuck. We'll reply within 24 hours with a time that works.

Preferred contact method