Part of our guide hub: AI solutions and automation
Every vendor selling AI support software will tell you it deflects seventy percent of tickets. That number is achievable and it is achieved by a minority of deployments. The difference is almost never the model quality.
The volume threshold
Before anything else, check whether this is worth doing at all.
- Under 100 conversations a week: automation is not your bottleneck. A well written FAQ page, saved replies and a pinned message solve most of it. Building anything here is a distraction.
- 100 to 500 a week: off the shelf tools with saved replies and simple automation are usually enough.
- Above 500 a week with repetitive content: the economics turn quickly and AI support becomes genuinely worthwhile.
- Above 2,000 a week: a custom build connected to your own systems starts to justify itself.
Most businesses asking about this are in the first two bands and would get more from fixing their documentation.
Buy: what off the shelf gives you
Modern support platforms with AI features handle the common case well:
- Ingest your help articles and answer questions from them.
- Draft replies for a human to review and send.
- Route conversations by topic and detect frustration.
- Work across email, web chat and increasingly WhatsApp.
This covers a large share of value at low risk. If your questions are answerable from documentation you could write, buy rather than build.
Cost: typically a per agent subscription plus usage based AI charges. For a small team, budget in the range of tens of thousands of rupees per month.
Build: when it is actually justified
The line is clear and it is about data, not sophistication.
Off the shelf tools answer from documents. They cannot answer questions that require looking at your systems:
- Where is my order, specifically, right now.
- Do you have this item in stock at the Lahore branch.
- What is my outstanding balance and when is the next instalment due.
- My warranty claim, reference 44812, what is happening with it.
These are the questions that make up most of the volume for retail, ecommerce, property and service businesses, and every one of them requires a live lookup into your own database.
That is the case for building: not a smarter model, but a model connected to your order system, your inventory and your customer records.
Cost: 400,000 to 1,500,000 PKR to build something that connects to real systems, plus running costs. See what an AI feature actually costs per month before committing.
The three things that decide success
1. Grounding, not general knowledge
The system must answer from your data and say it does not know when the answer is not there. A model that generates a plausible sounding delivery date is worse than useless, because the customer believes it.
Insist on this behaviour explicitly during evaluation: ask it something it should not know and see what it does.
2. A visible escape to a human
The most common failure across every deployment. Every conversation needs an obvious route to a person at every step, and frustration should trigger handover automatically on the first signal.
An annoyed customer who cannot reach a human is a complaint, a review and sometimes a lost account. The cost of that far exceeds the saving from deflecting their ticket.
3. Reading real conversations first
Before configuring anything, read two hundred actual conversations from last month. Not what you imagine customers ask. What they asked.
This single hour reshapes most projects: the top five questions are usually not the ones the team expected, and automating exactly those five captures most of the available benefit.
Start assisted, not autonomous
For the first two months, let AI draft and a human send. Every draft is reviewed, and the corrections tell you exactly where it is wrong.
This is slower to show savings and dramatically safer. Teams that switch to fully automatic replies on day one discover their errors in public, and the trust cost outlasts the efficiency gain.
Move to autonomous only for question types where you have seen it be reliably correct, and keep everything involving money, commitments or complaints with a human.
What to measure
- Deflection rate, meaning conversations fully resolved without a human. The headline number.
- Escalation rate after an AI reply. If customers frequently ask again, the AI answered without resolving.
- Customer satisfaction on AI handled conversations, compared with human handled.
- Cost per conversation, including the AI usage cost, compared with the agent cost.
Deflection alone is a misleading metric. A bot that ends conversations by exhausting the customer deflects beautifully.
The Pakistan specific issue
A large share of customers write in Roman Urdu or mixed script. Model quality varies substantially here, and generic platforms often perform noticeably worse than their English benchmarks suggest.
Test with real customer messages from your own inbox before committing to any platform. This is also an argument for menu driven flows for the highest volume questions, which sidestep language handling entirely. Our note on WhatsApp chatbots covers that pattern in detail.
Frequently asked questions
When is AI customer support worth it?
Above roughly 500 repetitive conversations a week. Below that, better documentation and saved replies deliver more for less.
Should we build or buy?
Buy if questions are answerable from documents. Build when answers require live lookups into your own order, stock or account systems.
What deflection rate is realistic?
Thirty to fifty percent is a good outcome for a well grounded system on repetitive volume. Seventy percent claims usually involve counting conversations the customer abandoned.
Will it work in Urdu?
Variable, and worth testing with your own customer messages before buying. Menu driven flows handle language differences without any model involved.
Ezitech builds AI support systems connected to real business data, and will tell you when an off the shelf tool or a better FAQ would do the job. See our AI solutions.