AI in a traditional business starts with one workflow.

The existing system is not the obstacle

Many tour operators, logistics companies, insurance agencies and clinics run on a booking system, an ERP, a policy management system or a practice management system that has been in place for years. It holds the data, the business rules and the habits of the team. Replacing it to make room for AI is expensive and rarely necessary.

AI can work alongside that system. It can read documents, sort incoming messages, draft answers and search internal knowledge while the system of record stays where it is. The useful question is not which AI platform to buy, but which piece of daily work would benefit first.

Broad surveys point the same way. McKinsey's 2025 global survey on the state of AI found that most organizations now use AI in at least one business function, while nearly two thirds have not yet begun to scale it across the enterprise. The organizations reporting the most value were more likely to redesign workflows and to define when a person must validate model output.

Start with one repeated workflow

Pick a task that happens many times a week, follows a recognizable pattern and takes real staff time today. It should also be easy to check: when the AI gets something wrong, a person can see the mistake and fix it before it causes harm.

Leave decisions with direct financial, legal or medical consequences for later. Pricing, claims decisions and clinical advice need more controls than a first project should carry.

• Documents: reading supplier invoices, booking confirmations or policy documents and filling fields for a person to check
• Email triage: sorting a shared inbox by type and urgency and drafting replies for review
• Customer questions: answering recurring questions about bookings, delivery status or opening hours from approved content
• Internal knowledge: letting staff search procedures, supplier terms and past answers in plain language

Connect to the existing system, read-only first

The information that makes an AI workflow useful usually lives in the existing system: the booking, the shipment, the policy, the appointment. Connect through the most stable route available, such as an existing API, a reporting database, a scheduled export or the integration options the system vendor already offers.

Begin with read access. An assistant that can look up and summarize but cannot change records limits the damage a wrong answer can do. Use a dedicated service account with only the permissions the workflow needs, keep credentials in the organization's normal secret management and log every request.

Add write access later, one action at a time, once the read-only stage has shown that the data and the answers are reliable.

Keep a person approving consequential actions

Decide in advance which actions count as consequential. A useful rule: anything that moves money, commits the business to a customer, changes a record other people rely on or sends a message in the business's name. For these actions, AI prepares the work and a person approves it.

Make approval fast enough that people actually do it. Show the source document or message, the proposed action and exactly what will change. Record who approved what. If you later let some low-risk actions run automatically, make that an explicit decision with an owner, not a gradual drift.

Measure against a baseline, then expand

Before switching anything on, measure how the workflow performs today: time per item, open backlog, response time and error rate. Without that baseline, nobody can say whether the AI helped.

Run the new workflow for a defined period and measure again. Track how often staff correct the AI output and what kind of corrections they make, because that shows where the instructions or the data need work. If the numbers do not improve, adjust or stop. If they do, choose the next workflow and reuse the integration, the permission model and the review screen you already built.

What to avoid

The common mistakes are organizational as much as technical. Each of them is easier to prevent at the start than to undo later.

Privacy deserves particular attention in Israel. Amendment 13 to the Privacy Protection Law, in force since August 2025, expanded the enforcement powers of the Privacy Protection Authority. Check with your legal advisor what your obligations are before personal data starts flowing into a new tool.

• Big-bang projects that try to automate a whole department or replace the core system at the same time
• Tools nobody owns, such as an assistant one employee connected to a shared mailbox before leaving the company
• Sending passport details, medical records, claims files or financial data to a tool without a data processing agreement and settings you have checked
• Automating a process nobody can describe clearly, which only makes an unclear process run faster

How a dedicated engineering partner approaches it

At UVISION we treat AI as part of the systems a business already depends on. We start with an assessment: which workflow to choose, where the data lives, who owns the process and what could go wrong. Then we connect carefully, with read-only access, scoped service accounts and logging, and we keep a person approving the actions that matter.

The work does not end at launch. Models, vendor terms and business needs change, so prompts, data sources and integrations need ongoing maintenance and development. We work as dedicated developers or as a managed engineering team, with at least one full-time developer on every project, so the knowledge of how your workflow runs stays with the people who keep developing it.