Ship Shift

AI agents for businesses

An AI agent is a system that understands a task written in plain language, queries your systems (ERP, CRM, email, documents), decides the steps and executes them — asking a person for confirmation when there is risk. At Ship Shift we design and deploy custom AI agents that work inside your operations, not in a demo.

Updated: 17/11/2026

Automatización de procesos

What is an AI agent — and how is it different from a chatbot or an automation?

A chatbot converses; a classic automation follows fixed rules (“if X arrives, do Y”); an agent combines both and adds judgement: it reads an ambiguous email, understands it is an order, checks stock, notices the delivery address is missing and writes to the customer before creating the order.

That ability to choose the next step is what people mean when they search for “AI agents for businesses” or “intelligent agents”. A FAQ bot is not enough, and neither is a Zapier flow that breaks when the email subject changes. The agent reasons inside limits you define.

Chatbot, automation and AI agent
ChatbotAutomation (RPA, Zapier, Make)AI agent
Understands natural languageYesNoYes
Uses your systems (ERP, CRM…)LittleYes, with fixed rulesYes, decides what to query
Handles unforeseen casesNoBreaksReasons or escalates to a person
ExampleAnswers website FAQsCopies a field from A to BProcesses a full order end to end

What tasks can an AI agent run in your company?

A useful agent targets a process with volume and real examples. These tasks already fit Ship Shift projects or the task-level solutions we publish:

  • Orders: read orders from email, PDF or WhatsApp and create them in the ERP (automate orders with AI). We do this in production with LogIA.
  • Supplier invoices: extract data, match to the order and delivery note, and leave them ready to post (automate invoices with AI).
  • Quotes: turn a measurement PDF into a quote with lines from a price bank. In a real public project we typed 159 uncoded lines with 83 % first-hit accuracy and 91.8 % in the top-5, on a curated bank of 1,045 items; we export to .bc3 and the AI never invents prices. See the installations quoting case.
  • Customer support: answer queries with live system data (order status, availability) and hand off sensitive cases to a person.
  • Email: classify shared inboxes, draft replies and assign tasks (reply to emails with AI).
  • Back office: prepare weekly reports, reconcile Excel with the ERP and flag anomalies.

If the pain sits in a specific department, cross this list with our pages for industry, construction or logistics. The agent is not a generic product: it is designed around your exceptions.

How we design an AI agent that does not fail on what matters

The AI interprets; the system calculates. That principle stops a model from “inventing” a price, a VAT rate or a quantity. We use five controls that belong in the design, not as a patch at the end:

  1. Scoped tools: the agent may only do what we grant permission for (read stock, create a draft order), never “everything”. Without the tool, it cannot improvise a dangerous action.
  2. Deterministic validation: numbers (prices, quantities, VAT) are computed by code or the ERP, not by the model. That is how we work on quoting: the AI never invents prices.
  3. Confidence threshold: if the agent is unsure, it does not execute: it parks the task in a review queue with its proposal and the reason.
  4. Human in the loop: irreversible actions (send, pay, delete) require human confirmation.
  5. Traceability: every decision is logged (what it read, what it decided and why) for audit and for improving the agent.

These controls fit process automation with AI and ERP/CRM integrations: without reliable connectors, the agent has nothing to work with.

What technology do we use?

We pick components for cost, latency and privacy — not for fashion. The agent is the product; the model is a swappable component.

  • Agent orchestration: orchestration frameworks that define tools, states and human-review checkpoints. Stack details are published when the team confirms them.
  • Models: language models from multiple providers, chosen for cost, latency and privacy. EU-hosted models are an option.
  • Integrations: your ERP or CRM API and, when there is no API, connectors to database, files or email.

We do not lock you to one vendor: if a better or cheaper model appears tomorrow, we swap it without rebuilding the agent. When the case only needs an assistant that queries internal documents (without executing ERP actions), the usual path is the internal docs AI assistant.

How much does an AI agent cost and how long does it take?

Cost depends on the number of systems to integrate, case volume, data quality and autonomy level. We publish ranges on pricing when the team sets them; until then we quote fixed-scope phases.

Autonomy levels for an AI agent
LevelBehaviourBest fit
Assisted agentProposes; a person approvesNew or high-risk process
Semi-autonomousRuns safe cases; escalates the restHigh volume with clear exceptions
Autonomous with samplingRuns with sampled reviewMeasured, stable accuracy

A typical prototype is ready in 2–4 weeks on your data. The pilot and the reduction of review depend on measured accuracy, not on a marketing date.

Most back-office agents (orders, invoices, email) are not high-risk under the EU AI Act. Even so, GDPR still matters: data minimisation, a legal basis and contracts with model providers.

Watch Annex III uses (recruitment, creditworthiness assessment…): those obligations apply from 2 December 2027, after the deferral in Regulation (EU) 2026/1744. If the agent talks to customers, you must tell them they are speaking with AI (Article 50 transparency). From the design we apply decision logging, data minimisation and contracts with model providers.

For the inventory, classification and technical-measures plan, see AI Act readiness. We are not a law firm: we cover the technical and organisational side alongside your legal counsel.

What does an AI agent project with Ship Shift look like?

We follow the five steps of the shared methodology, adapted for agents:

  1. Pick one process with volumeOne measurable work queue (orders, invoices, measurements), not five at once.
  2. Collect 50–200 real cases, including odd onesWithout edge cases the agent only learns the happy path and fails on Monday morning.
  3. PrototypeScoped tools, accuracy metrics and a review queue visible to your team.
  4. Pilot with 100% human reviewNothing is sent or posted without human eyes at first.
  5. Lower review based on measured accuracyMove to semi-autonomous or sampling when the numbers hold.

More detail in methodology. If you still do not know which process to attack first, the AI diagnosis prioritises by return.

Related links

FAQ about AI agents

What is the difference between an AI agent and ChatGPT?

ChatGPT converses; an agent uses your company tools to complete tasks, with permissions and logging.

Can an AI agent replace an employee?

It replaces repetitive tasks, not people: it frees hours for work that needs judgement.

What if the agent makes a mistake?

Sensitive actions go through human review; everything is logged and can be undone.

Does it work with my ERP even if it is old?

Yes: if there is no API, we connect via database, files or email. We have done this with systems without an API.

Do my data leave the company?

Only to providers you approve, under contracts that exclude training. EU-hosted models are an option.

How long until an agent is working?

A typical prototype is ready in 2–4 weeks. Full rollout depends on integrations and the agreed autonomy level.

Do I need clean data beforehand?

Not entirely: part of the project is organising the minimum needed.

Have a process an agent could run?

Tell us the volume, systems involved and where it fails today. We will say whether an assisted or semi-autonomous agent fits.

Tell us your process