AI Agents for Business in 2026: Where They Pay Off and How to Start

From chatbots to agents

For the last few years, most companies met AI as a chat window: you ask, it answers. An AI agent goes a step further. It can read data, decide on the next step and take action in other systems: create a draft invoice in your ERP, update a lead in your CRM, reply to a customer on WhatsApp, or prepare a weekly report.

That shift is happening quickly. Gartner forecasts that 40% of enterprise applications will include task-specific AI agents by the end of 2026, up from less than 5% in 2025. In other words, many of the business tools you already pay for are about to ship with agents built in.

But the same research firm adds a warning. Gartner also predicts that over 40% of agentic AI projects will be cancelled by the end of 2027, because of unclear business value, rising costs or inadequate risk controls. You can read more in Gartner’s Hype Cycle for Agentic AI.

So agents are coming either way. The real question is how to get value from them without becoming part of that 40%.

Adoption is wide, scale is rare

The numbers show a big gap between trying AI and getting results from it.

• According to McKinsey’s State of AI 2025, 88% of organizations use AI in at least one business function.
• Yet only about 23% report scaling agentic AI somewhere in the enterprise.
• PwC’s 2026 CEO Survey, as reported, found that only around 12% of CEOs see both revenue gains and cost reductions from AI.

Almost everyone is experimenting. Few have turned experiments into measurable business results. The difference is rarely the model itself; it is usually the process, the data and the way the agent is connected to real systems.

Where agents actually pay off

Agents work best on tasks that are repetitive, rule-based enough to check, and connected to systems you already run. Typical use cases we see in the region include:

• Invoice and receipt data entry. Reading supplier invoices (often a mix of Arabic and English) and preparing entries in the ERP for an accountant to approve.
• Customer service on WhatsApp and the website. Answering common questions, checking order status, and handing the conversation to a human when it gets complex or sensitive.
• Sales lead qualification. Asking the right first questions, scoring the lead and booking a call with the right salesperson in the CRM.
• HR requests. Leave requests, salary certificates and policy questions handled without emails going back and forth.
• Report generation. Pulling figures from the ERP every week and writing a first-draft summary for management.
• Internal knowledge search in Arabic and English. Letting staff ask questions about procedures, contracts or product specs and get answers with a link to the source document.
• Procurement follow-ups. Chasing suppliers for quotes, delivery dates and missing documents.

None of these are science fiction. They are ordinary, time-consuming tasks where saving even a few hours a week across a team adds up quickly.

Why so many pilots fail

When an agent project stalls, the cause is usually one of these:

• No clean data or integrations. The agent cannot do useful work if it cannot reliably read from and write to your ERP or CRM.
• No clear KPI. “Let’s try AI” is not a goal. Without a number to improve, nobody can say whether the pilot worked.
• Too much autonomy, too soon. Letting an agent send payments or change records without guardrails or approvals is how small errors become expensive ones.
• No human in the loop. Customers and staff need a clear path to a real person.
• Ignoring Arabic quality. An agent that handles English well but writes awkward Arabic, or misreads Gulf or Levantine dialect, will lose customer trust fast.
• Weak security and permissions. An agent must respect the same role-based access (RBAC) as the user it acts for. If an employee cannot see payroll, neither should their assistant.
• Untracked costs. Model usage, integrations and maintenance add up. If you do not track cost per task, the business case can quietly disappear.

Good first agent vs. bad first agent

Good first agent	Bad first agent
One repetitive, high-volume task	“An AI that runs operations”
Clear before/after metric (time, errors)	No defined measure of success
Reads and drafts; a human approves	Acts on money or data with no approval
Connected to the ERP/CRM via APIs	Works on copy-pasted spreadsheets
Respects existing user permissions	Has admin access “to make it work”
Tested in Arabic and English	Tested only in English

How a mid-size company should start

1. Pick one process. Choose something repetitive and measurable, such as invoice entry or first-line customer questions.
2. Measure the baseline. Record how long it takes today and how often errors happen.
3. Connect to systems of record. Integrate with your ERP and CRM through proper APIs, not screen scraping or exported files.
4. Keep humans on the important actions. Require approval for anything that moves money or changes master data.
5. Log everything. Keep an audit trail of what the agent read, decided and did, so you can review and improve it.
6. Apply the same permissions. The agent sees and does only what the responsible user is allowed to.
7. Test the language. Check answers in Arabic and English with real users before going live.
8. Track cost per task. Compare it with the time and errors you are saving.
9. Expand gradually. Once the first agent proves its value, move to the next process and reuse the integrations you have built.

A note on standards

Agents are also becoming easier to connect. Open protocols are emerging, such as MCP (Model Context Protocol) for connecting models to tools and data, and Google’s A2A (Agent2Agent) protocol for agents talking to each other, backed by more than 50 partners including Salesforce, SAP, Atlassian and PayPal. You do not need to master these on day one, but it is worth choosing systems that support open standards rather than locking yourself into a single vendor.

How we approach it at Cortex

Because we build and maintain ERP systems, we see the same pattern again and again: the agent is the easy part, and clean integration with the system of record is what makes it useful. Our AI and automation work starts with one measurable process, human approval on sensitive actions and full logging, then grows from there. If you have a process in mind, that is usually the best place to begin the conversation.