Why your software should work with AI agents
Tun KelteschPublished
A growing share of everyday work now starts with a request to an AI assistant. A buyer asks ChatGPT to compare two suppliers. An accountant asks Claude which invoices are still open. A sales lead asks Copilot to prepare a customer meeting from the CRM. A warehouse manager asks Gemini which deliveries are late this week. When the assistant can use the software involved, it completes the task. When it cannot, it answers from whatever it finds, or it turns to a product it can use.
For a company with a product, a customer portal, a shop or internal systems, the business question is whether those systems should be usable by an agent, and when.
More work starts in an assistant
The usage figures come from the companies and statistical offices themselves, all checked on 27 September 2026:
| Measure | Figure | Source |
|---|---|---|
| ChatGPT weekly users | over 900 million | OpenAI, Feb 2026 |
| Gemini app monthly users | 950 million | Alphabet, Jul 2026 |
| EU firms (10+ staff) using AI | 20.0% in 2025, 13.5% in 2024 | Eurostat |
| Luxembourg firms using AI | 33.6% in 2025, 14.5% in 2023 | Eurostat |
Luxembourg sits well above the EU average, and the share of its enterprises using AI more than doubled in two years.
The more telling change is that assistants are moving from answering to doing. McKinsey's State of AI 2025 survey found 23% of respondents' organisations scaling an AI agent system somewhere in the business, with another 39% experimenting. The Model Context Protocol (MCP), the open standard most assistants use to connect to other software, had 97 million monthly SDK downloads and over 10,000 active public servers in December 2025. By July 2026 its maintainers reported "close to half-a-billion downloads a month". In one developer market the shift is already the majority: when Databricks announced it was buying the database company Neon, it wrote that agents created over 80% of new databases on the platform, up from 30% (all checked 27 September 2026).
What it means for your product
Your customers will increasingly reach you through their assistant. Retail shows the pattern first. Stripe and OpenAI launched Instant Checkout in ChatGPT in September 2025, built on an open protocol for agent purchases. Shopify announced in March 2026 that millions of its merchants can sell inside AI chats. In July 2026 Visa reported live purchases by AI agents in Europe, with over 30 European card issuers and merchants including lastminute.com and Frasers. Adobe measured traffic from AI tools to US retail sites up 693% during the 2025 holiday season, while noting that the base of users is still modest (all checked 27 September 2026).
The same logic applies well beyond shops. If a procurement manager asks an assistant to reorder the usual quantities from a supplier, the supplier whose portal the agent can use gets the order placed in a minute. The one that needs a login and a PDF form gets put off until later. For a software product, an assistant that cannot read your data will suggest exporting it, or recommend a competitor it can work with.
Customer experience changes too. A client who wants the status of a claim or next week's delivery date may never open your app. They will ask their assistant. If the assistant can fetch the answer from your system with the client's permission, the answer is correct and your service gets the credit. If it cannot, the client gets a guess or a link to your support page.
Where the value is
Staff time. Many teams already use assistants at work, then copy results by hand between the assistant, the CRM, the ERP and a spreadsheet. Connecting the assistant directly to internal tools removes that step, along with the typing errors it produces. The value is highest where people repeat the same lookup or update many times a week: order status, stock levels, open tickets, customer history.
Timing within your category. In several categories the leading products already ship official agent access: Stripe for payments, Notion for documents, Linear for project tracking, Atlassian for Jira and Confluence, and in Europe the French business bank Qonto (checked 27 September 2026). Once an assistant works well with one product in a category, its users have less reason to try another. Most categories in Luxembourg and the wider EU have no such product yet, which leaves room for whoever builds it first.
Lower integration cost. Until recently, every partner integration was a separate project: one for this marketplace, one for that ERP, each with its own maintenance. Agent access through MCP is built once and then works with ChatGPT, Claude, Gemini, Microsoft Copilot, Cursor and other clients that support the standard, according to Anthropic's list of adopters (checked 27 September 2026). The same work can later serve partners who build their own agents.
A view of what customers want. Requests arriving through agents show, in the customer's own words, which tasks they try to get done with your product. Few companies have that data today.
When it is too early
Demand is uneven. For many products, the people asking for agent access today are a few technical customers. A consumer product used twice a year, such as a tax-filing tool or a moving checklist, gains little, because nobody delegates a task they barely repeat.
The risks are real. An agent that misreads an instruction can change the wrong record, and text written by outsiders, such as a customer note or a support ticket, can steer an agent into doing something its user never asked for. These are solvable problems, but they cost engineering time: permissions, confirmation before changes, a record of every action and a way to reverse it. We cover that in a separate article on how to build it safely.
The technology is also young. MCP broke compatibility with earlier versions in July 2026, and Gartner predicted in June 2025 that over 40% of agentic AI projects will be cancelled by the end of 2027, naming cost and unclear business value among the reasons (checked 27 September 2026). Projects that start from a vague goal ("we need an AI agent") are the ones most exposed to that outcome.
Where to start
- List the tasks people repeat most. Ask sales, support, finance and operations which lookups and updates they do every day, and ask a few customers what they would hand to an assistant. Pick three to five.
- Start read-only. Let an assistant look up orders, balances, documents or status before it is allowed to change anything. Reads carry privacy questions but rarely break anything.
- Open it to your own staff first. Internal users give fast feedback and forgive rough edges.
- Measure. Track which tasks people actually use, and whether support requests and manual steps go down. Add write actions one at a time, where the numbers justify it.
TK MEDIA builds agent access into existing products and internal tools. If you are weighing it for yours, write to us.