AI agents in 2026: how to build systems that actually work
A practical guide to AI agents: MCP, durable work, tracing, evaluations, human approval, pilot economics and the ArcCloud five-gate test.
The full article is available below. Enable JavaScript to use quizzes and simulations.
A practical AI implementation guide
AI should not replace the business. It should remove repetition and return decisions to people.
Picture an ordinary Monday. The inbox holds 86 messages, three invoices still need to be described, sales staff are copying data into the CRM, and an engineer is investigating the same recurring failure again. Everyone is busy, yet little of that work moves the company forward.
A sound AI implementation does not start with choosing a model. It starts by separating work into repeatable steps, decisions that require accountability, and exceptions the system must not resolve on its own. This guide shows how to do that in a small business and a large organisation without building an expensive, aimless demo.
Technology status: September 2026
An agent is no longer just a chat interface. It is a controlled process operator.
The important shift is not simply better text generation. Modern agent platforms can keep long-running work alive, delegate to specialised subagents, use tools through shared protocols and retain an end-to-end execution trace. That makes useful automation possible, while making permissions, data provenance and accountability for every action non-negotiable.
MCP and A2A
Agents can use approved tools and cooperate with other agents without a bespoke integration for every model.
Long-running work
State, context and results can survive a longer workflow. The agent does not have to finish in one answer or hide a failed step.
Tracing and evals
Every step can be replayed, measured and compared with an expected result. Without this, autonomy is expensive guesswork.
Identity and gates
An agent receives its own identity, least privilege, limits and approval queue instead of access to the whole company.
The 20/80 model without marketing magic
20% human attention, up to 80% of routine work delegated to the system
This is not a guarantee or a target for every process. It is a pilot hypothesis: people retain judgement, relationships, approvals and exception handling, while the system takes over classification, retrieval, data preparation, first drafts and follow-up coordination.
The real automation ratio depends on data quality, exception frequency, legal constraints and the cost of failure. One workflow may reach 30%, another 85%. Measure the baseline first and never call time “saved” when nobody measured it.
Fit the technology to the work
The best AI process is frequent, documented and verifiable
High volume is not enough. You need a clear input, rules, an owner and an unambiguous signal that the result is acceptable. The more expensive an error and the less reversible the decision, the stronger the human gate must be.
Customer enquiries
Identify intent, collect context, propose a reply and route it to the right person. Send only after approval.
Documents and data
Read invoices, compare fields, detect missing information and prepare a system entry. Posting and payment remain controlled.
Sales and CRM
Company research, call summaries, record completion and follow-up tasks. The agent neither impersonates people nor sends spam.
Finance and risk
Document and signal analysis can support an expert, but credit, rejection and movement of funds require policies and accountable people.
HR and recruitment
AI may organise schedules and documents. It should not autonomously assess people, filter by sensitive traits or make hiring decisions.
Health and safety
Diagnosis, critical infrastructure control and rights-impacting decisions require qualified experts, validation and the appropriate regulatory regime.
| Sector | Sensible first use case | Human gate | Pilot metric |
|---|---|---|---|
| Professional services | Intake, classification and response draft | before sending | first-response time, routing accuracy |
| E-commerce | Product copy, returns and missing-data analysis | pricing exceptions | handling time, correction rate |
| Manufacturing | Anomaly detection and maintenance planning | engineer decision | downtime, alert precision |
| Finance | Extraction and document-completeness checks | expert decision | analysis time, false positives |
| Healthcare | Administration and scheduling, not diagnosis | clinical staff | completeness, time, incidents |
| IT | Log triage, tests, runbooks and fix drafts | review and deploy | MTTR, regressions, incident cost |
Decision 1 / process selection
You receive 400 similar messages per month with 12 exceptional cases. Where do you start?
Let the agent answer every message immediately
Classify and create a draft for approval
Buy the largest available model first
Choose the pilot entry point and inspect the consequence.
Two different scales
A small business needs a short path. An enterprise needs controlled boundaries.
A five-person service company
Goal: stop losing enquiries and shorten response preparation.
- Forms and email enter one queue.
- The agent identifies service, urgency and missing information.
- It searches only approved company knowledge.
- It prepares a response draft and a CRM task.
- The owner edits or approves; corrections become evaluation material.
Do not automate: a quote without data, a delivery-date promise or a sensitive response.
An organisation with 500+ people
Goal: accelerate documents without opening cross-department access.
- Every agent has its own identity, scope and operating budget.
- Data comes only from permitted sources with tenant and role filters.
- Outputs include a schema, sources, confidence and model version.
- Financial, legal and employment actions enter an approval queue.
- Evaluations, logs, costs and incidents remain visible to the process owner.
Do not automate: lateral data access, self-escalation of privileges or untraceable decisions.
The agent contract
A good agent is not a “smart chat”. It is a constrained system worker.
Before its first run, specify the contract as precisely as you would specify an API.
public sealed record AgentDecision(
string Action,
decimal Confidence,
IReadOnlyList<string> Sources,
bool RequiresHumanApproval,
string Reason);
public async Task<AgentDecision> PrepareReplyAsync(Request request)
{
var context = await knowledge.FindApprovedAsync(request.Topic);
if (context.Count == 0)
return Stop("No approved source");
var draft = await model.CreateStructuredDraftAsync(request, context);
var requiresApproval = draft.Confidence < 0.88m
|| request.ContainsPersonalData
|| request.EstimatedValue > 5_000m;
await audit.WriteAsync(request.Id, draft, requiresApproval);
return draft with { RequiresHumanApproval = requiresApproval };
}Decision 2 / permissions
The agent prepares an invoice and checks it against an order. Which permission does it receive?
Read documents and save a draft, without payment access
Full bank and ERP access
A shared team account without a separate audit
Choose the least-privilege scope.
Economics before models
(volume × time × cost) − AI cost − oversight cost − error costIf you do not know volume, baseline time and correction cost, you do not yet have a business case. You only have an idea.Minimum pilot acceptance
- 100–300 representative test cases
- comparison with the non-AI process
- quality accepted by the process owner
- cost per case and cost of exceptions
- manual corrections and incidents
- stop, rollback and data-deletion plan
Exercise / order the rollout
Click the stages in a safe sequence
Small-cohort pilot
Process and baseline
Scale
Data and risk
Evaluation and acceptance
Agent contract
Start with the process, not a tool purchase.
Reset sequence
Controlled-run simulation
Run the agent preflight before the first case
Run preflight
The system is waiting for five control layers.
Decision 3 / scaling
The pilot reduced handling time by 42%, but 7% of outputs needed correction. What now?
Enable autonomy for every case
Classify errors and scale only the safe segment
Delete the solution because it did not reach 100%
Choose the decision after the first measurement.
Law, security and accountability
The greater the impact on people, the less room there is for autonomy
The EU AI Act follows a risk-based approach. Systems affecting health, safety or fundamental rights may face strict duties covering risk management, data quality, logging, documentation, human oversight, robustness and cybersecurity. NIST structures the operational work around four continuous functions: Govern, Map, Measure and Manage. That remains a useful pattern even when a particular use case is not legally classified as high risk.
A practical rule: do not delegate an action you cannot reproduce, explain, stop and reverse. Do not place customer data in an unapproved tool. Do not treat model output as truth without sources and verification rules.
Sources and further work
- European Commission: AI Act and risk levels
- NIST AI Risk Management Framework
- NIST AI RMF Playbook: Govern, Map, Measure, Manage
- OpenAI: Agents API and durable agent workflows
- OpenAI: agent tracing and evaluation
- Model Context Protocol interoperability roadmap
- AWS AgentCore observability
- Microsoft Foundry Agent Service overview
This material provides technical and organisational guidance. It does not replace legal, security or compliance analysis for a specific implementation.
Your score
Correct implementation decisions
Solve the three implementation decisions.
Clear answers
From process to a working pilot
Is one process consuming too much time? Before buying a tool, verify whether it is suitable for AI.
We will map the baseline, risks, data access, human gates, economics and acceptance criteria. Then we decide whether building is justified.