Insights
-
Home
-
Insights
An AI agent can draft an invoice, match an incoming payment to an open receivable, or flag a likely duplicate charge for review; it should not set a new customer's credit limit, resolve a disputed deduction, approve a collections escalation, or write off a balance, because those calls carry account and relationship risk a model cannot absorb alone.
MCP servers for enterprise systems: how to build, secure and govern them
An MCP server for an enterprise system exposes a small, named set of tools, not the whole API, over the ERP, CRM or finance system behind it: read-only tools run freely, write tools wait for human approval, every caller authenticates through OAuth 2.1, every call is logged, and a registry lists every agent connection so it can be reviewed.
AI Agent Observability and Cost Control: Traces, Budgets and Kill Switches
AI agent observability and cost control means tracing every step of a running agent (prompt, tool calls, tokens and cost), capping spend with a tiered budget ceiling that warns before it stops, and giving a named person a kill switch that halts the run, refuses further calls, and records why it stopped.
A2A and MCP in a multi-agent design: when a mid-market company needs both
A mid-market company needs MCP alone when one agent calls tools and systems on its own behalf; it needs A2A as well once a second specialised agent must discover, delegate to or hand off work to another agent, with each agent keeping its own governed MCP connections underneath for the systems it touches.
A mid-market company should name four owners for every AI agent — business, technical, risk and operations — and set a three-tier escalation ladder by decision class, so a dispute, a blocked action or a decision outside the agent's mandate always reaches a named person within a stated time, not whoever is online when it pages.
Agentic AI platform vs custom-built agents for mid-market companies
Put a candidate AI-agent use case on a configurable platform when it is a standard, single-system workflow with no spare engineering capacity to own it; commission a custom-built agent, or a governance platform layer, when the workflow is proprietary, multi-system, or must produce a full audit trail the product cannot export.
Vendor lock-in and exit planning for agentic AI platforms: what to negotiate before you sign
Before signing with an agentic AI platform vendor, negotiate six clauses while you still have leverage: data-export scope and format, who owns the source code and agent configurations built on the platform, the notice period before a model change, rate-card stability, termination assistance, and a contract term matched to the use case's switching cost.
AI Agent Incident Response and Rollback: What to Do When an Agent Acts Wrong
When a live AI agent takes a wrong action, contain it first: stop it, revoke its credentials, and confirm it cannot repeat the action. Then preserve the run as evidence, test whether the action is technically reversible, compensable, or neither, and only then decide whether the agent returns, is restricted, rolled back, redesigned or retired.
Every change to a live AI agent — a new prompt, a model version, an updated tool or a wider permission — is a release: give it a version identifier, gate it with the same regression set that cleared the agent for launch, roll it out in stages, and keep the prior version ready to restore if it regresses.
Choosing an API integration agency or consultancy: partner, in-house team or platform
Once a business decides its systems must be connected by API, the next decision is who builds and runs the connection: an external agency, an in-house team, or an iPaaS/no-code platform. Score the choice by strategic value, delivery capacity, security and IP ownership, and total cost over time, not price alone.
AI Agent Drift Detection: Catching Silent Behaviour Degradation Between Releases
Agent drift is a live agent's behaviour changing even though nobody deliberately edited its prompt, model, tools or permissions; catch it by replaying a fixed golden set on a schedule, alerting when the score crosses a threshold you set, and diffing output against the frozen baseline from launch, instead of waiting for a complaint.
Healthcare workflow platform requirements: what to define before you scope, brief or send an RFP
Before a clinic or hospital scopes a workflow platform or sends an RFP, it needs a written requirement set in four categories: who the platform changes work for, what data must move in and out, which compliance rules apply, and which performance and access rules are non-negotiable, with a named owner for each before a vendor prices the work.