Agent Economy: Your ERP Needs to Speak AI

While the Spanish business software industry spends its marketing budgets advertising regulatory compliance, something far bigger is happening. And almost nobody is paying attention.

This is not about a new feature or another regulation. This is about a structural shift in how businesses interact with their software. A shift where the user does not open an application, does not navigate menus, and does not click buttons. Instead, they tell an AI agent what they need and the agent gets it done.

This is not science fiction. It is happening now, it has real numbers behind it, and it is going to change the rules for every small business that uses invoicing software.

The numbers nobody is talking about

In October 2025, McKinsey published a report that should have shaken the entire business software industry. Their conclusion: AI agents will mediate between 3 and 5 trillion dollars in global commerce by 2030. In the US alone, the B2C market could reach one trillion dollars in agent-orchestrated revenue.

This is not a projection from an AI blog. This is McKinsey.

The adoption data supports the thesis:

• 97 million monthly downloads of the MCP SDK, the protocol that connects agents to external tools. A year ago it was 100,000 per month.
• Claude Code, Anthropic's product for developers, generates 2.5 billion dollars in annualized revenue. Its usage doubled in January 2026.
• Cursor, the AI-powered code editor, surpassed 1 billion dollars in annual recurring revenue. It is the fastest SaaS product to reach that milestone in history.
• Cline, a VS Code extension, has 5 million active developers. Samsung is testing it company-wide.
• Adobe reported a 4,700% year-over-year increase in traffic to retail websites from AI browsers and chat services.

Gartner adds context: by 2028, 90% of B2B purchases will be intermediated by AI agents. Not assisted. Intermediated. The agent searches, compares, negotiates, and executes.

By 2028, 90% of B2B buying will be AI-agent intermediated, channeling 15 trillion dollars through automated exchanges. — Gartner, November 2025

What is an agent and why it matters for invoicing

An AI agent is not a chatbot. A chatbot answers questions. An agent executes tasks.

When a freelancer tells Claude Code "create an invoice for TechStart Ltd for 40 hours of consulting at 75 EUR per hour, due March 1st," the agent does not generate a nice paragraph. It finds the connected invoicing tool, calls the correct API, sends the data in the right format, and receives confirmation with the invoice number, calculated VAT, and total.

The freelancer did not open any application. Did not navigate any menu. Did not click anything.

This works because there is a standard protocol that lets agents connect with external tools: MCP (Model Context Protocol). Created by Anthropic, adopted by OpenAI, backed by Google, and now governed by the Agentic AI Foundation under the Linux Foundation. With platinum members including AWS, Microsoft, Cloudflare, Bloomberg, and Block.

MCP won. It is the USB port for AI agents. If your software has an MCP server, any agent can use it. If it does not, your software does not exist for them.

The protocol ecosystem in 5 minutes

The protocol landscape as of February 2026 has stabilized into complementary layers:

Protocol	Function	Backed by
MCP	Agent connects to tools (APIs, databases, ERPs)	Anthropic, OpenAI, Google, AAIF
A2A	Agent talks to agent (multi-agent coordination)	Google, Linux Foundation
ACP	Agent buys products (commerce)	OpenAI + Stripe
UCP	Universal commerce (retail, marketplaces)	Google, Shopify, Walmart, Target

For an ERP, MCP is the relevant protocol. It is what allows an agent to create an invoice, record an expense, or query clients. The other protocols orchestrate higher levels: an agent using A2A to coordinate with another agent that uses MCP to access your ERP.

Meanwhile, in the Spanish market

The Spanish business software industry is fighting a battle it already won. Certified invoicing software (Verifactu) becomes mandatory in 2027, and every vendor is investing in positioning themselves as "your Verifactu solution." It is a necessary battle, but it is yesterday's battle.

Verifactu forces 2.9 million freelancers and small businesses to adopt certified software. That is real and important. But it is a regulatory requirement, not a differentiator. Everyone will comply. It is like advertising that your car has seatbelts: mandatory, not optional, not special.

The interesting question is not who complies with Verifactu. It is who is ready for what comes after.

And what comes after is a world where your invoicing software is not used by you. It is used by your agent.

Here is what each Spanish competitor offers for that world:

ERP	Public API	Webhooks	MCP Server	AI Skill
Holded	Undocumented	No	No (third-party only)	No
Quipu	Limited	Basic	No	No
Anfix	No	No	No	No
Billin	Limited	No	No	No
Contasimple	No	No	No	No
Sage	Partial	No	No	No
Frihet	Yes (5 resources, docs)	Yes (14 events, HMAC)	Yes (official, npm)	Yes (Claude Code)

The gap is not incremental. It is structural. You cannot bolt an MCP server onto a product that has no public API. You need to rebuild the architecture.

How it works in practice

A real scenario. Maria is a freelance designer. She uses Claude Code to draft proposals and review contracts. Today, when she finishes a project, she tells Claude:

"Create an invoice for the GreenTech branding project for 3,200 EUR."

Claude accesses Frihet's MCP server, finds the client GreenTech in the database, creates the invoice with the correct tax details (21% VAT, 15% income tax retention), assigns the sequential number, and confirms. Maria did not open Frihet. Did not search for the client. Did not calculate taxes. The agent did.

Another scenario. A development team works in Cursor. They need to check if a client has paid before activating a premium feature. They ask directly in the editor: "Show unpaid invoices for the API-Premium client." Without switching tabs, without losing context, without interrupting their workflow.

And the most interesting one: an autonomous agent — running on a next-generation agent platform — manages an e-commerce business. Every Friday it generates a weekly report: invoices issued, expenses recorded, net cash flow. It sends it via Telegram to the owner. Zero human intervention. Zero dashboards to check.

The regulatory moat nobody sees

Agents can create invoices. But invoices must comply with the law.

In Spain, a valid invoice requires the correct VAT rate (21%, 10%, or 4% depending on the type), income tax retention when applicable (15% or 7% for new freelancers), sequential numbering without gaps, complete tax details for both issuer and recipient, and starting in 2027, a certified digital signature under the Verifactu regulation.

An AI agent knows none of this. It does not know that the VAT on a book is 4%. It does not know that a first-year freelancer's income tax retention is 7%. It does not know that skipping an invoice number is illegal.

But it does not need to. That is what the ERP is for.

When an agent calls create_invoice on Frihet, Frihet applies all fiscal rules. The agent provides the business data (client, concept, amount). Frihet provides regulatory compliance. It is a division of responsibility that works.

This is a moat that cannot be replicated with a generic MCP server. You need a real ERP behind it, with fiscal logic, Verifactu certification, and knowledge of the local tax context. A wrapper over an API does not solve this.

What is coming in the next 18 months

Sequoia Capital describes three phases in the agent economy:

1. Assistants (2023-2024): ChatGPT, Copilot. The user asks, the AI answers. Reactive.
2. Agent swarms (2025): Agents collaborating with each other, continuous monitoring, multi-step automations.
3. Agent economy (2026 onwards): Agents as economic actors with identity, payments, and reputation.

We are in the transition from phase 2 to phase 3. What is coming:

Stripe x402 — a micropayment protocol where an agent calls an API, receives an HTTP 402 (payment required) response, pays automatically, and receives the data. Zero subscriptions, pure pay-per-use. Currently in developer preview.

Standardized MCP registries — over 8,000 MCP servers are already listed in directories like Smithery, PulseMCP, and the official registry. Agents search for tools in these registries the way you search for apps in an app store.

Agent identity (KYA) — a16z and Catena Labs are working on "Know Your Agent" frameworks, the equivalent of banking KYC but for AI agents. Cryptographic credentials linking an agent to its owner.

Google UCP — launched in January 2026 with Shopify, Walmart, Target, PayPal, and Visa. Enables AI agents to buy products natively within Google Search and Gemini.

What a small business should do today

You do not need to build an agent infrastructure tomorrow. But you need to choose software that is ready for it.

Three questions you should ask your business software provider:

1. Do you have a documented public API? If the answer is no, your software is a black box that no agent can use.
2. Do you have an MCP server? If the answer is no, your software does not exist for Claude, Cursor, Windsurf, or any current AI assistant.
3. Do you have webhooks? If the answer is no, your software cannot integrate with automation platforms like n8n, Zapier, or Make.

If your current provider cannot answer yes to at least two of these three questions, you have a problem that regulatory compliance will not solve.

The opportunity for first movers

McKinsey is clear: first movers will have a decisive advantage. Not because the technology is inaccessible, but because inertia is enormous. Most small businesses will keep using spreadsheets until regulation forces them to change. And when they change, they will choose the software that was recommended to them.

The difference is who recommends it. It used to be the accountant or the tech-savvy relative. Now it is the AI agent the business owner already uses to draft emails, summarize documents, and analyze data.

If that agent can connect with your ERP, it recommends it. If it cannot, it recommends another one.

That simple. That brutal.

---

Frihet is the first Spanish ERP with an official MCP server, published on npm as @frihet/mcp-server. 31 tools, documented REST API, 14 webhook event types, and a Claude Code skill with Spanish tax intelligence. Try it free at app.frihet.io.