AI Agent Readiness for SMEs: Is Your Company Ready for AI Agents?

What this is about

AI agents shift the focus from generative assistance to software-supported action. They can plan tasks, analyse data, use tools and — where they are authorised to — trigger actions in real systems. That is why NIST launched a dedicated AI Agent Standards Initiative in 2026, naming secure interoperability, open protocols, agent security and identity as the central fields of action.

For SMEs, though, agent readiness does not mean rolling out MCP, llms.txt or an autonomous agent as fast as possible. What matters is whether the specific use case is clear, whether processes and data are suitable, whether information is machine-readable, whether systems offer controlled interfaces, and whether permissions, governance and evaluation grow along with the level of autonomy.

A useful technical reference comes from Switzerland: AI architect Thierry Gilgen has built Engawa, an open-source toolkit for agent-native websites. His approach cleanly separates the human web from the agent interface, without claiming that agents cannot read HTML. An additional agent surface makes the same publicly visible content available in a cleaner, smaller and more deterministic form — with explicit boundaries and read-only defaults.

AHEAD OF TIME frames this more broadly: a company is only genuinely agent-ready when business use case, processes, data, website, interfaces, security, governance and accountability line up.

What does AI agent readiness mean?

AI agent readiness describes a company's ability to deploy AI agents sensibly, reliably and under control — while also enabling its digital systems to interact with external agents in defined ways. The term therefore covers two perspectives: internal organisational and system readiness, and the external agent readiness of websites and digital services.

That distinction matters. A website can be very well structured for AI search and machine reading without any external agent being allowed to trigger a transaction. Conversely, a company can run capable agents internally even though its public website has no dedicated agent surface.

The AHEAD OF TIME definition: an SME is agent-ready when a clearly defined agent can do its intended job with the right information and tools, within controlled permission, security and accountability boundaries — and when the effect is measured.

From chatbot to agent: why the requirements change

A classic chatbot mainly produces an answer. An agentic system can go further: break goals into steps, process external information, choose tools and trigger actions. NIST describes AI agent systems as systems that can plan and carry out autonomous actions with effects on real systems or environments. It is precisely the combination of model output with software functions that raises additional security and control questions.

A marketing assistant writing a campaign proposal is therefore something quite different from an agent that reads customer data from the CRM, builds audiences, creates campaigns in ad platforms and changes budgets. Every additional capability does not automatically increase the benefit — but it does increase the possible impact of a mistake.

What that looks like in practice is shown by the concrete setups in our hands-on series on AI agent automations: the point there is not model choice, but process scope, data access and control.

The AHEAD OF TIME 5-level agent readiness model

For SMEs, a graded model is more useful than the question "agent-ready: yes or no?". AHEAD OF TIME distinguishes five readiness levels. They are meant as a strategic orientation model, not as a formal standard.

Important: level 5 is not automatically the goal. For many SMEs, a cleanly structured, easily understood website plus a few controlled internal agent processes makes more economic sense than maximum possible autonomy.

How agent-ready is your website?

Websites have traditionally been optimised for people and search engines. Browser-based agents add another mode of use: software should not only find information, but understand it, compare it and in part prepare or perform actions. Current research on agent-ready websites distinguishes accordingly between interpretability, actionability and decision reliability.

That leads to practical questions. Are products and services described semantically clearly? Are price, availability, validity and restrictions unambiguous? Do forms have clean labels and predictable states? Are actions clearly named? Can an agent tell whether an action succeeded?

Agent readiness is therefore closely tied to good web quality: semantic HTML, accessibility, structured data, stable URLs, clear content and deterministic states help people, search engines and agents at the same time.

Engawa: a controlled threshold to the agent web

Thierry Gilgen takes a particularly pragmatic approach with the open-source project Engawa. In Japanese architecture the name describes a transitional space between inside and outside — and that is exactly how Gilgen understands the agent interface: as a controlled threshold between an existing website and the agent web.

Engawa replaces neither the website nor the CMS, schema.org, sitemap, robots.txt or OpenAPI. Instead, it can additionally expose the same human-public information as deterministic markdown resources, llms.txt and bounded read-only MCP. The core governance principle: the agent corpus should match what an anonymous human visitor can already see in public.

Gilgen's sober discovery rule is especially important: an agent surface can exist without any given provider actually fetching it, using it or producing better output because of it. Implementing something is no substitute for measuring it.

Placing llms.txt, markdown, MCP and APIs correctly

The agent web debate quickly becomes protocol-centric. For SMEs it helps to understand the function first.

The right question is therefore not "do we need MCP?" but: which information or actions must an agent be able to use reliably for our use case — and which existing or new standard solves that task most simply?

Security: the more agency, the more the boundaries matter

As long as an agent only reads public information, the risk profile is comparatively limited. Once it may send emails, change files, trigger payments or edit CRM data, security becomes a design component rather than a checkbox added later.

OWASP describes excessive agency as the risk that arises when LLM-based systems are given too much functionality, overly broad permissions or too much autonomy. Misbehaviour can be triggered by hallucinations, direct or indirect prompt injection, compromised tools or other unexpected inputs.

In 2026 NIST focuses particularly on identity and authorisation: agents need a reliable identity, clearly defined authority and limited access to data, tools and applications. Classic cybersecurity principles remain relevant, but have to be adapted to agentic systems.

Practical security principles for SMEs

• Read-only first: where possible, start new agent surfaces without mutations
• Least privilege: only release the functions and data the specific use case needs
• Explicit tool allowlist: no "the agent may do anything", only clearly defined functions
• Validation and limits: restrict inputs, result sizes, search scope and action parameters
• Human and policy gates: critical actions need approval or deterministic rules, depending on risk
• Audit and recovery: log actions, detect errors and be able to undo their effects

Which processes are suitable for agents?

Not every automatable process needs an agent. Tasks that require several information sources, variable intermediate steps or tool decisions are particularly suitable. Processes that are fully deterministic and reliably covered by classic rules are less suitable.

Good starting points:

• Research and monitoring across several sources with a clearly defined output
• Internal knowledge search with a bounded corpus and source references
• Pre-qualifying leads or service requests before a human takes over
• Producing and checking content drafts with defined brand and approval rules
• Coding and DevOps tasks with limited repositories, tests and review gates

Poor starting points:

• Unclear processes that nobody can describe properly even without AI
• Unrestricted access to production systems "so the agent stays flexible"
• Highly critical decisions without reliable data, control mechanisms or accountable owners
• Autonomy as an end in itself, without measurable quality, time or cost targets

If you are unsure here, the biggest gain comes from making the brief clearer first. The same discipline that makes a good prompt — role, context, boundaries, output format — decides whether an agent works usefully. The entry point is in our article on prompt engineering for SMEs.

Agent readiness check for SMEs

Ten management questions are enough for a first assessment. The more of them cannot be answered clearly, the more likely it is that the fundamentals need work before an agent pilot.

1. Is the specific business use case and its desired outcome defined?
2. Is today's process documented, and are the exceptions known?
3. Is the required data available, current and legally usable?
4. Can the agent access only the information it genuinely needs?
5. Are the website and public facts machine-readable and consistent?
6. Are there stable APIs, MCP tools or other controlled interfaces for the required actions?
7. Are identity, authentication and authorisation regulated?
8. Are there human or policy gates for critical steps?
9. Can agent actions be logged, evaluated and, in case of errors, stopped or rolled back?
10. Are the business owner, the technical owner and accountability for risks clear?

Where to get the answers in a structured way

Those ten questions are a self-assessment around the table. If you want something more robust, use the two checks that actually exist at AHEAD OF TIME — and that cover exactly the two halves of agent readiness:

• Digital Readiness & Potential Check — the company side. Around 35 scored questions across ten dimensions, from strategy through processes and data to technology and security. Particularly relevant for agent readiness is the AI discoverability dimension: it checks whether crawlers and agents are granted access at all according to robots.txt, whether an llms.txt exists, and whether your own entities are marked up in machine-readable form. That is levels 1 to 3 of the model above — measured rather than guessed.
• AI Skills Check — the personal side. Six standalone levels, from AI Beginner through AI Practitioner, AI Advanced, AI Integrator and AI Architect to AI Strategist, tested across nine topics — among them LLM fundamentals and agents, data protection and sovereignty, prompt and context engineering, failure modes and limits, and building your own agents. Agent readiness rarely fails on technology alone; it fails when nobody on the team can judge where an agent is reliable and where it is not.

Both checks sit side by side in the overview of the digital and AI checks. If you would rather read first, the AI & automation topic hub collects the relevant articles.

Swiss implementation expertise: Thierry Gilgen, Engawa and Gnocchi & Friends

Agent readiness is a good example of why strategy, web architecture, AI engineering and security have to come together. AHEAD OF TIME combines these topics in its own AI transformation, education and engineering mandates and works with specialised partners where that helps.

One such cooperation context is Gnocchi & Friends, the Swiss network of senior experts for digital transformation, AI and marketing. In the public network, Michael Schranz is listed as a digital business innovation expert and Thierry Gilgen as an AI architect. The service model covers tech integration, MVP development, custom software, API integration and automation, among others.

For agent readiness projects this combination is particularly relevant: AHEAD OF TIME covers business readiness, use cases, customer experience, skills building and transformation logic; technical specialists such as Thierry Gilgen bring deeper architecture and agent web perspectives. Engawa's open-source character is valuable here, because principles and implementation are publicly verifiable and not tied to a proprietary platform.

The cooperation does not mean that every project needs Engawa or MCP. The Engawa documentation itself warns against unnecessary additional layers. What counts is the use case — and the simplest architecture that fulfils it safely and measurably.

From readiness check to pilot project

If you can answer most of the ten questions, do not immediately build a company-wide agent platform. A narrow pilot is usually the better next step.

1. Choose the use case: a frequent, clearly bounded task with recognisable benefit
2. Limit the scope: define data, tools and permitted actions explicitly
3. Start read-only: where possible, research, classify or generate proposals before mutating anything
4. Define evaluation: measure quality, error types, time saved, cost and security events
5. Expand under control: only release additional tools, data or autonomy after solid results

Agent readiness is therefore not a one-off technical audit. It develops with every new use case. More autonomy needs more control, more external data needs stronger protection, and new interfaces have to be measurably used in practice.

Conclusion: agent-ready is a company state, not a file

2026 is producing new agent standards, new protocols and new products all at once. That makes it tempting to define readiness through individual technical artefacts. For SMEs, that would fall short.

llms.txt can be useful. Markdown can make retrieval more efficient. MCP can provide a standardised agent interface. But none of these elements answers on its own whether the business use case makes sense, the data is suitable, the permissions are safe or the process is robust.

The better sequence is therefore: understand the use case → check processes and data → add agent surfaces and tools only where they serve a concrete purpose → scale security and governance with agency → measure the effect.

Do it that way and you are not simply building "an agent". You are creating the conditions under which agents can actually work reliably inside the company or at digital touchpoints.

Q&A — the most common questions about AI agent readiness

What does AI agent readiness mean? AI agent readiness describes whether processes, data, content, systems, interfaces, permissions and governance are prepared so that an AI agent can execute a defined use case reliably and under control.

Does an agent-ready website necessarily need llms.txt? No. llms.txt can be useful as a compact agent handoff, but it is not a prerequisite for every website and does not guarantee use by AI providers. Semantically clean HTML, structured data and a sitemap may already be enough, depending on the goal.

Does an agent-ready website need MCP? No. MCP becomes relevant when explicitly connected agents are meant to use resources or defined tools in a standardised way. For pure information websites it can be unnecessary.

What is the difference between GEO and agent readiness? GEO focuses primarily on visibility and citability in generative search and answer systems. Agent readiness goes further: agents should understand information reliably and — where intended — carry out actions under control.

What risks come with AI agents? Additional risks arise in particular from tool access, prompt injection, overly broad permissions, identity and authorisation problems, and too much autonomy. The more an agent is allowed to do, the more important least privilege, audit and approval mechanisms become.

How does an SME start with AI agents? With a clearly bounded use case, limited data and tools, read-only wherever possible, defined quality and security metrics, and human oversight. Autonomy should only be extended after successful evaluations.

How do I measure how agent-ready my company is? On two levels. The Digital Readiness & Potential Check measures the company side including AI discoverability — crawler access according to robots.txt, llms.txt and entity markup. The AI Skills Check measures across six levels how well the people on the team can judge where an agent works reliably.

Is Engawa a standard for all websites? No. Engawa is an open-source toolkit by Thierry Gilgen for deliberately designed agent surfaces. The documentation itself emphasises that many websites need no additional layer and that existing standards such as HTML, schema.org, sitemap or OpenAPI remain relevant.