Conversational AI Consultancy in Israel: What to Expect and How to Evaluate a Partner

TL;DR: A conversational AI consultancy helps a business decide where chatbots, voice agents or AI assistants make sense, designs them around real customer conversations, connects them to company systems and data, and sets up the guardrails and measurement that keep them reliable. In Israel, a good partner must also handle Hebrew properly, including mixed Hebrew and English, spelling variations and right-to-left display, and must be comfortable with Israeli privacy rules. Evaluate partners on discovery work, live references, how they prevent wrong answers, who owns the data and prompts, and how they measure success after launch.

Since large language models became widely available, almost every software house and marketing agency in Israel offers "AI chatbots". The range of quality behind that phrase is enormous. This guide explains what a serious engagement looks like and how to tell the difference before you commit budget.

What a conversational AI consultancy actually does

A consultancy is not just a vendor that installs a chatbot widget. A complete engagement usually covers five areas:

1. Use-case selection. Reviewing your customer service, sales and internal support conversations to find where automation will pay back and where it will annoy people.
2. Conversation design. Defining what the assistant handles, what it refuses, how it hands over to a human, and how it sounds in your brand voice.
3. Knowledge and integrations. Connecting the assistant to your documents, product catalogue, CRM, booking system or order data, often through retrieval-augmented generation (RAG).
4. Guardrails and testing. Preventing invented answers, protecting personal data, testing against real and adversarial questions before launch.
5. Measurement and iteration. Tracking resolution rate, escalations, satisfaction and errors, and improving the system after launch.

If a proposal covers only the third point, you are buying a technical integration, not a consultancy.

Types of partners compared

Partner type	What they offer	Good fit	Watch out for
Chatbot SaaS platform with onboarding	A product plus setup help	Standard FAQ and lead capture	Limited customisation, data on their platform
Strategy consultancy	Roadmaps, business cases, vendor selection	Large organisations planning AI programmes	May not build anything themselves
AI development studio	Design, build and integration of custom assistants	Businesses with specific processes and systems	Quality varies widely; check live work
Large systems integrator	Enterprise-wide deployments, governance	Banks, insurers, government	Cost and speed
Freelance AI developer	Prototypes and small bots	Proofs of concept	Support, security and continuity

Aiterra sits in the third row: we design and build conversational agents and AI integrations for Israeli businesses. The criteria below apply to every row, including us.

Hebrew is where many projects fail

Large language models handle Hebrew far better than they did a few years ago, but real customer messages are messier than a demo. Israelis write with and without vowel letters, mix English product names and slang into Hebrew sentences, use heavy abbreviations, and often write in WhatsApp style with no punctuation. Voice adds accents, background noise and switching between languages mid-sentence.

A capable partner tests on your real historical conversations, not on clean sample questions, and checks how answers render in right-to-left interfaces such as your website widget and WhatsApp. Ask to see test results on Hebrew data before you sign off on a build.

How a typical engagement is structured

Discovery. Interviews, a review of conversation logs, and a short list of use cases ranked by value and risk. The output is a written recommendation, which may include "do not automate this".

Pilot. One use case, one channel, limited traffic. The goal is to measure real performance with real customers, with a human able to step in.

Production. Wider rollout, integrations with more systems, monitoring dashboards, and an agreed process for updating knowledge and handling errors.

Ongoing optimisation. Reading transcripts, fixing gaps in the knowledge base, adjusting prompts and handover rules as the business changes.

Questions to ask a conversational AI partner

• Can you show a live assistant you built, and let us try to break it?
• How do you stop the assistant from inventing prices, policies or promises?
• What happens when it does not know the answer?
• Where are conversation logs stored, and who can access them?
• Which models do you use, and can we switch provider later without a rebuild?
• Who owns the prompts, the knowledge base and the integration code?
• How will we measure success, and what results do you consider a failure?
• What does ongoing support cost and include?

A partner that cannot answer the second and third questions concretely is selling a demo.

Privacy and data in Israel

Conversations often contain names, phone numbers, ID numbers and order details. Israel's Privacy Protection Law, strengthened by Amendment 13 in 2025, applies to that data. Ask your partner where data is processed, whether conversation content is used to train third-party models, how long logs are retained, and how access is controlled. For regulated sectors, it is worth involving your legal and security teams from discovery, not at launch. This is general information, not legal advice.

Checklist: evaluating a conversational AI consultancy

1. Define the business goal in one sentence before talking to vendors.
2. Gather a sample of real customer conversations to share under NDA.
3. Ask each partner for a discovery phase with a written recommendation.
4. Request live references and test their assistants yourself.
5. Check Hebrew performance on your own messages, including messy ones.
6. Confirm how hallucinations are prevented and how handover to humans works.
7. Clarify data storage, retention and model training policies in writing.
8. Ensure you own the prompts, knowledge base and integration code.
9. Agree on success metrics and a pilot scope before the full build.
10. Budget for ongoing optimisation, not only for the initial build.

What success looks like after launch

Agree on how you will judge the project before it starts, because "the bot works" means different things to different people. Useful measures include:

• Containment or resolution rate: the share of conversations the assistant completes without a human, counted honestly, excluding customers who simply gave up.
• Escalation quality: whether conversations reach the right person with the full context, so the customer does not repeat the story.
• Error rate: answers that were wrong, made up or off-policy, found by reading a sample of transcripts every week.
• Customer satisfaction: a short rating at the end of the conversation, compared with your human channels.
• Business outcomes: booked meetings, qualified leads, orders handled or support hours saved, depending on the use case.

Expect the first weeks to reveal gaps in your own documentation more than flaws in the model. Questions customers ask that nobody wrote an answer for are normal and valuable; they are the backlog for the next iteration. A partner who reviews transcripts with you regularly will improve the system much faster than one who hands it over and disappears.

Also plan for change. Model providers release new versions and retire old ones, and prices change. An architecture that keeps your knowledge base, prompts and integrations separate from a specific model makes those transitions routine rather than a rebuild.

When conversational AI is not the answer

If you receive a handful of enquiries a week, if every question is unique and high-stakes, or if your underlying information such as prices and policies is not written down clearly, a conversational assistant will struggle to pay back. Sometimes the right first step is fixing the knowledge base or the process itself. A good consultancy will tell you that. Our AI consulting engagements start with exactly that assessment, and our article on AI agents for business explains when an agent goes beyond a chatbot. For phone-based use cases, see our voice AI agents page.