AI agents and SaaS: what changes for software buyers

In short: an AI agent is software that connects to your systems through APIs and decides on its own which steps to take to finish a task. It does not replace your CRM, ERP or e-commerce platform, which remain the home of your data, but it can make redundant the tools whose only job is to view, move or re-key that data.

Key takeaways

The early-2026 software sell-off hit big names too, but the long-term risk falls mainly on easy-to-replicate SaaS, not on the platforms that hold data and processes.
Agents run on top of existing systems: without open APIs and clean data they deliver nothing.
Software pricing is moving from per-seat licences to paying per conversation, per action or per outcome.
Replacing a subscription only pays off when the task is repetitive, measurable and its errors are easy to fix.
For smaller companies the first step is often not the agent but clean data in an ERP or CRM the agent can work on.

Between late January and early February 2026 the S&P 500 software and services index lost roughly a trillion dollars in market value. According to Information Age, the Australian Computer Society's publication, the trigger was the launch of industry plugins (finance, legal, marketing, data analysis) for Claude Cowork, Anthropic's agent. As of 10 February the index stood about 25% below its October 2025 peak. Some dubbed it the "SaaSpocalypse": if agents use software instead of people, companies may need fewer seats.

If you buy software rather than trade it, the question is more practical: which subscriptions can I cut, which must I keep, and where does an agent connected to my data do better than yet another tool? In Italy, Istat reports that 16.4% of companies with ten or more employees used at least one AI technology in 2025, double the 2024 figure, and 68.1% buy intermediate or advanced cloud services.

What is an AI agent, and how is it different from SaaS?

SaaS gives you a tool to use; an AI agent uses the tools for you, within the limits you set.

SaaS (software as a service) is an application you use in the browser for a subscription fee, usually per seat or per volume: Shopify, HubSpot, a helpdesk, a cloud ERP. The vendor looks after servers, updates and security.

An AI agent is a system in which a language model decides which steps to take and which tools to call in order to reach a goal. In its guide Building effective agents, Anthropic separates workflows, where models and tools follow code paths defined in advance, from agents, where the model directs the process itself. The same guide recommends starting with the simplest solution: often a workflow is enough.

Agents need connectors to reach company systems. The Model Context Protocol (MCP) is an open-source standard that lets AI applications access databases, files and external tools in a consistent way; the official documentation compares it to a USB-C port for AI. In practice, one agent can read an order from your store, update a contact in the CRM and email the customer, provided it has permission to do so.

One distinction runs through the rest of this article. A system of record is the software where your official data lives: ERP, CRM, e-commerce platform. The interface layer is everything used to read, move or present that data. Agents put pressure on the second, far less on the first.

A common misconception is that an agent "learns on its own" as it is used. In a business project, memory and improvement are designed: request history is stored in a database, mistakes are fixed by updating instructions, examples and rules, and every change is checked against the logs before it is rolled out further. Without that work, improvements do not happen by themselves.

Are AI agents killing SaaS?

No, but they are changing which SaaS is worth paying for. Software that holds critical data and processes stays; software that adds a single feature on top of someone else's data becomes easy to replace.

That is also the view of the Omdia analyst quoted by Information Age: a long-term correction for feature-focused or easily replicated SaaS, while high-quality platforms with deep integrations may even benefit from AI. In other words, the risk is not the end of software. It is the end of software that owns neither the data nor the process.

There is no shortage of radical views. In December 2024 Microsoft CEO Satya Nadella described business applications as databases with business logic on top, logic that he expects to move into agents; in the same podcast, as reported by Computerworld, he added that agents are more likely to complement applications than to eliminate them. More recently, Vercel CEO Guillermo Rauch wrote that almost every SaaS app used inside the company had been replaced by a generated app or an agent interface. Nordic APIs, which covered the post, points out that SaaS customers do not just pay for an interface: they pay for convenience, security, ease of use, service-level agreements (SLAs), ongoing maintenance and customer service. Those are exactly the items a home-built alternative has to cover.

The same goes for micro SaaS, small vertical tools sold on subscription to solve one problem: writing product descriptions, translating listings, syncing two apps. For years they were a handy shortcut. Today many of those features can be rebuilt with a language model and a custom automation. The micro SaaS that survives has proprietary data, an integration that is hard to rebuild, or a compliance burden the buyer would rather not carry.

For a company, the point is not to bet on a vendor's share price but to work out how much of its software spend pays for work an agent could do on top of systems it already owns. As I argue in my piece on enterprise AI and operational context, the value sits in data and processes, not in the model.

What impact are AI agents having on the software industry?

Three effects are already visible: pricing is changing, the user is changing, and integrations are worth more than ever.

Pricing moves from seats to usage and outcomes

Per-seat pricing works as long as people are the ones using the software. When an agent does the work, vendors charge for the activity. Salesforce's public Agentforce pricing lists, alongside per-user licences, a price per conversation ($2) and usage-based credits charged per action. Intercom pairs per-seat helpdesk licences with its Fin agent at $0.99 per "outcome", meaning a resolved request, and charges nothing if the customer asks for a human.

The user becomes an agent

When an agent uses the software, screens matter less and APIs, permissions and logs matter more. Vendors with closed APIs become a bottleneck.

Integration beats individual tools

The advantage shifts to whoever connects the systems and governs the flow between a model, a CRM and an ERP. That is the idea behind orchestrating multiple AI systems: the more AI tools you run, the more you need someone deciding who does what, with which data.

For buyers, the practical consequence is less predictable spend. A fixed fee is easy to budget; a per-conversation fee depends on volumes, seasonality and how well the agent is set up.

From Software as a Service to "Service as a Software"

In an analysis published by the Italian outlet Agenda Digitale, Mauro Colopi, a partner at Bain & Company, describes the move from Software as a Service to "Service as a Software": value no longer lies in helping a person work with data and processes, but in the work the software does itself. Repetitive, rule-based and predictable tasks go to AI; strategic and decision-making work stays with people.

The same piece describes three layers that are useful for buyers too:

Systems of record: the sources of truth that hold the data, control access and ensure compliance.
Systems that orchestrate agents: they decide what to do, in which order and with which tools, and keep track of context.
Outcome interfaces: the tools through which users ask, receive updates and check results.

Its examples show where the line falls. Managing a marketing list, classifying a ticket or drafting a standard email reply are tasks an agent can reproduce. Systems such as project accounting in complex industries, like construction or pharmaceuticals, involve intricate rules and high accountability: there AI strengthens the software but is unlikely to replace it.

When can an AI agent replace a software subscription?

When the tool performs a repetitive, rule-based task on data that already lives in another system, and a mistake is cheap or easy to fix. In every other case the agent works alongside the software rather than replacing it.

Software type	Example task	What to do	Why
ERP, CRM, e-commerce platform	Orders, customer records, invoices, stock	Keep	System of record: the agent works on top of it through APIs
Helpdesk and first-line chat	Order status, returns, product questions	Add an agent, then trim seats	The agent handles repetitive requests, people handle exceptions
Simple reporting	Weekly sales and campaign summary	Consider replacing	An agent queries the data and writes the summary; official numbers stay in the dashboard
Content micro SaaS	Product descriptions, translations, meta tags	Replace with a custom flow	Standard task, output can be spot-checked
"Bridge" tools and data entry	Leads from web form to CRM, B2B orders from email to ERP	Replace with automation plus agent	The most exposed part of the interface layer
Regulated software	Payroll, accounting, filings	Keep	Legal liability, regulatory updates, high risk

What if the tool you need has no API? Some agents use a browser the way a person would, clicking and filling in forms: I cover this in my article on AI browser automation. They are slower and more fragile than an API connection, because a single screen change can stop them, so I would treat them as a stopgap rather than a long-term architecture.

How much does an AI agent cost compared with a SaaS subscription?

An agent is not free just because it has no licence fee; it costs money in a different way. Compare it with the total cost of the tool it replaces, staff hours included, not with the sticker price.

Cost item	SaaS subscription	Custom AI agent
Set-up	Low: configuration and training	Design, integrations, testing, training for whoever supervises it
Recurring cost	Fixed fee per seat or volume	Model usage, automation hosting, any platform licences
Maintenance	Handled by the vendor	Handled by you or your builder: instructions, connectors, API changes
Supervision	Minimal	Spot checks, exception handling, log reviews
Predictability	High	Medium: tracks volumes
Exit	Depends on data export	Depends on how well the flow is documented and portable

The sum is easy to write down: licences you can drop plus staff hours saved, minus build cost spread over the months, monthly usage and supervision hours. In a hypothetical online store where half the tickets are about order status and returns, the team can run on fewer helpdesk seats; if tickets are few and all different, the agent costs more than it saves. To measure it you need indicators agreed upfront, as covered in my article on KPIs for measuring the value of AI.

Should you use the agent built into your SaaS or a custom one?

It depends on how many systems the task touches. If the work happens entirely inside one product, the vendor's agent is usually the fastest route; if the flow runs across your store, CRM, ERP and email, you need an orchestration layer that you control.

Large vendors are moving this way: according to Computerworld, Salesforce and ServiceNow are building no-code platforms for creating AI agents that automate parts of a workflow. Agentforce pricing and Intercom's Fin agent, mentioned above, are examples of agents sold by the software vendor itself.

Option	When it makes sense	What to watch
Agent built into the SaaS you already use	The task lives in a single system (helpdesk, CRM) and the vendor already holds your data	Usage pricing, vendor lock-in, little visibility into what happens outside that product
Custom workflow or agent on top of APIs	The flow spans several systems or needs company-specific rules	Connector maintenance, documentation, an internal owner
No agent	Low volumes, every request different, costly mistakes	Revisit when volumes or processes change

Questions to ask the vendor

Actions: which operations does the agent perform, on which systems, and which ones need approval?
Integrations: which APIs or connectors are available, including to other vendors' software?
Permissions: can read and write access be separated and limited by data type?
Logs: is every action recorded, searchable and exportable?
Pricing: do you pay per seat, per conversation or per outcome? Can you set spending caps?
Data: where is it processed, is there a data processing agreement (DPA), and is your data used to train models?
Exit: if you switch vendors, can you take your instructions, data and history with you?

Practical examples for e-commerce and B2B companies

Agents pay off most where someone currently copies data from one screen to another or answers the same questions every day.

E-commerce

Post-purchase support: the agent reads order status from the store and the carrier, replies to the customer, opens a return according to policy and hands anything outside the rules to a person.
Catalogue: product listings, translations and attributes generated from supplier data, reviewed before publishing. On translation, see also my overview of AI translation software.

B2B companies

Lead qualification: a website enquiry goes to an agent that enriches the company with public information, assigns a priority and creates the contact in the CRM with a note for sales.
Sales support: before a call, the agent summarises order history, open tickets and recent emails from the CRM and suggests a next step; the decision stays with the sales rep.
Quotes: email requests become a draft quote using price lists and terms from the ERP; the sales rep checks and sends it.
Account-based marketing: research on target accounts and draft personalised messages, reviewed before sending. More in my guide to account-based marketing.
Logistics: the agent reads emails from carriers and customers, updates shipment status in the ERP and only alerts the customer when there is a delay. See also logistics digitalisation for SMEs.

In every case the CRM, ERP and store stay where they are. What disappears is the copy-and-paste work and, sometimes, the tool that only existed to do it. Flows like these are built with n8n or Make for orchestration, an OpenAI model for the reasoning step and a database such as Supabase for logs and memory: it is the stack I use to design business automation systems.

What are the risks of replacing software with an agent?

The main risks are mistakes, overly broad data access and runaway costs. You manage them with minimal permissions, human approvals and spending caps, within clear AI governance.

Errors: a model can be wrong and sound sure of itself. Actions with financial impact, such as refunds, discounts or supplier orders, need human sign-off.
Data and privacy: the agent reads customer and supplier data. You need a GDPR legal basis, a data processing agreement with every vendor involved and access limited to what the task requires.
Security: malicious text hidden in an email or web page can push the agent into unintended actions (prompt injection). Keep read and write permissions separate.
Variable costs: with usage pricing, a spike in requests becomes a spike in the bill. Set spending limits and alerts.
Vendor lock-in: keep instructions and logs in systems you control.
Regulation: according to the European Commission, the AI Act entered into force on 1 August 2024 and applies from 2 August 2026; for high-risk systems in areas such as hiring, the rules apply from 2 December 2027. Chatbots carry a transparency duty: users must know they are talking to a machine.

Can SMEs benefit, or is this a game for large companies?

SMEs can benefit, but they start further back. In Italy, according to Istat, 15.7% of SMEs used at least one AI technology in 2025 (up from 7.7% in 2024), compared with 53.1% of large companies, and the gap widened to 37 percentage points. The same report finds that 48.8% of SMEs use an ERP against 85.9% of large companies, and 21.1% use a CRM against 56.5%.

For the topic of this article, that is the figure that matters: an agent works on top of a system of record. If customers, orders and quotes live in spreadsheets and inboxes, the first investment is getting that data into an ERP or CRM, not buying an agent.

Istat also reports that a lack of skills held back AI adoption in almost 60% of companies that considered an AI investment but did not go ahead, and that AI is used most in marketing and sales (33.1%) and in organising administrative processes (25.7%). For an SME this means three practical things:

start with a sales or administrative process, where examples and tools are most mature;
appoint an internal owner for the flow, even if an external partner builds it;
check whether grants or incentives are available for AI adoption: I cover this in my article on public funding for AI.

For the wider picture, see also the future of AI in small businesses.

What if you are the one selling software on subscription?

If you make an ERP, a vertical platform or a SaaS product for other businesses, agents are both a threat and a chance to evolve your offering. Nordic APIs argues that robust APIs become the execution layer for agents and that vendors with well-designed APIs will be better placed. The Agenda Digitale analysis adds that AI has to become part of the product rather than an add-on, and that pricing tied to the number of users stops making sense when a machine does the work.

In practice, the questions to ask are:

which product features a customer's agent can already handle alone, and which ones need your data or expertise;
whether your APIs let an external agent read, write and respect permissions without going through the interface;
whether your pricing can include a usage or outcome component alongside licences;
which simple, repetitive feature to automate first, to test the model with your current customers.

I discuss this more broadly in my article on how AI is changing software business models.

How to apply this in your company: first steps

List every subscription. For each tool, note annual cost, active users, the data it holds and whether it has an API.
Classify each tool using the table above: system of record to keep, tool to complement, replaceable task.
Pick one pilot process that is repetitive and measurable, such as order-status replies or lead qualification, and set two or three baseline metrics: response time, hours per week, error rate.
Build a workflow first and add an agent only where needed. Fixed rules where possible, model decisions where requests vary, human approval for anything with financial impact.
After 60 to 90 days, compare the metrics and decide whether to cut seats, extend the project or stop. To widen the scope you need a process map, covered in business process automation.

Mistakes to avoid

Cancelling the software before the trial ends: run both in parallel until the agent has proved itself on your data.
Buying chatbots rebranded as "agents": ask the vendor which actions it performs, on which systems and with what logs.
Granting write access everywhere: the agent should only be able to change what its task requires.
Comparing the agent with the licence fee alone: maintenance, supervision and usage belong in the sum.
Starting without an internal owner: someone has to read the logs, handle exceptions and decide when to change the instructions.
Connecting an agent to messy data: if customer records and price lists are inconsistent, the agent multiplies errors instead of reducing them.

Frequently asked questions (FAQ)

Will AI agents completely replace SaaS?

No. Software that holds your official data, such as ERP, CRM and e-commerce platforms, stays essential because agents work on top of it through APIs. What loses ground is the tool that only acts as an interface or a bridge between systems, along with part of the per-seat licences.

Which software subscriptions can an AI agent replace?

Usually the ones that perform a repetitive, rule-based task on data that already lives somewhere else: data entry between apps, simple reporting, product descriptions or translations, first-line support answers. Payroll, accounting and regulated software should stay, because the risk and the legal liability are high.

How much does it cost to deploy an AI agent?

It depends on how many systems you connect and on volumes. The cost items are design and initial integrations, model usage, automation hosting, maintenance and supervision time. Compare it with the total cost of the tool it replaces, staff hours included, not with the licence fee alone.

Can an AI agent work with Shopify, WooCommerce or my CRM?

Yes, as long as the software exposes APIs or connectors and the agent gets permissions limited to its task. Standards such as the Model Context Protocol make these connections more uniform. A tool with no API is the first sign that it will need to be replaced or complemented by something more open.

What is the difference between an automation and an AI agent?

An automation, or workflow, always runs the same predefined steps, for example in n8n or Make. An agent decides which steps to take and which tools to use based on the request. For well-defined tasks a workflow is more predictable and cheaper; an agent earns its place when requests vary a lot.

Are AI agents compliant with GDPR and the EU AI Act?

They can be, if you process data on a valid legal basis, use vendors with data processing agreements and keep access to the minimum. The AI Act requires transparency: people talking to a chatbot must know they are dealing with a machine. Customer service agents are usually not high-risk systems, while agents used in hiring, for instance, are.

Is an AI agent worth it for a small or mid-sized company?

Yes, if there is a repetitive process with enough volume and the data already sits in an ERP or CRM. In Italy, Istat reports that 15.7% of SMEs used AI in 2025, against 53.1% of large companies: start with a single, measured pilot process and an internal owner.

Sources

If you want to know which of your subscriptions would work better with an agent and which should stay as they are, start with a free diagnosis.