The problem with "press one"
Phone menus were built to save businesses money, and callers pay the price. A look at how IVR fails people, and what a voice agent should do instead.
Most people can recite the opening of a customer-service call from memory. A greeting, a language choice, then a menu. Kwa Kiswahili, bonyeza moja. Press two for account information. Press three for complaints. Press nine to hear these options again.
Interactive voice response, or IVR, was a reasonable idea when it arrived. It let businesses route calls without a person answering every one. But it was built to suit the business, not the caller. Decades later, it still feels that way.
This post looks at where press-one menus go wrong, and what we think a good voice agent should do instead.
How menus fail callers
The caller has to translate their problem into a number
Nobody calls a bank thinking "this is a category four issue". They call because money left their account and did not arrive. A menu forces them to guess which option covers their situation. Guess wrong and they land in the wrong queue, or back at the start.
Language is a gate, not a conversation
Many menus ask the caller to choose Swahili or English up front, then lock them in. Real speech does not work like that. A caller might choose Swahili and then need to read out an English product name, a reference code, or a term they only know in English. The system has no room for that.
Context gets dropped
A caller enters their account number, selects three menu levels, and finally reaches a person, who asks for the account number again. Every handoff loses what came before. The caller ends up explaining the same problem two or three times.
Feature phones and noisy lines make it worse
Menus assume a clean keypad and a quiet room. In practice, people call from feature phones with small keys, from a daladala, from a busy market, or on a weak signal. Tones get missed. Speech recognition built for studio audio gives up. The menu times out and repeats itself.
Waiting costs real money
For many callers, every minute on hold is paid for in airtime. A long menu, a slow queue and a dropped call are not just frustrating. They are expensive, and the cost falls hardest on the people who can least afford it.
What a voice agent should do instead
Replacing a menu with a "smart" menu is not enough. The goal is to let people say what they need, in the way they would say it to a person, and get it resolved.
Understand natural speech
A caller should be able to say, "Nimetuma pesa jana lakini hazijafika" — I sent money yesterday but it has not arrived — and be understood. No menu. No keywords. Just the problem, in their own words.
Accept code-switching as normal
People mix Swahili and English all the time, and in Nairobi they add Sheng on top. "Nataka ku-reset PIN yangu" is a perfectly clear request. An agent that only works when the caller sticks to one language is not built for the people calling. We go deeper on this in code-switching is not an edge case.
Confirm the details that matter
Speech recognition will never be perfect on a noisy phone line. A good agent knows which details carry risk, such as amounts, phone numbers, account references and dates, and reads them back before acting. "Umesema shilingi elfu hamsini, sawa?" A short confirmation prevents a long complaint.
Hand off to a person, with context
Some problems need a human. The agent's job then is to make that handoff clean: route to the right team, pass a short summary of what the caller said and what has already been checked, and never make the caller start over.
Respond fast enough to feel like a conversation
Long pauses make people talk over the system or hang up. Our target for Waridi Agents is first audio in about 500 ms after the caller stops speaking, and always under 800 ms, so turn-taking feels natural. We explain how we think about that in latency budgets for real-time voice agents.
Design principles we are holding ourselves to
As we build Waridi Agents, these are the rules we keep coming back to:
- The caller's time is the scarcest resource. Every question the agent asks must earn its place.
- Say it once. Information given at any point in the call should follow the caller to the end, including to a human agent.
- Speak like a good receptionist, not a robot. Polite, clear, warm. Use the forms of address a Tanzanian caller expects, such as karibu and samahani, without being stiff.
- Be honest about limits. If the agent cannot help, it should say so quickly and offer a person, rather than looping.
- Never trap anyone. There must always be a clear way to reach a human.
- Work on the phones people actually have. That means telephony over SIP, not just smartphone apps, and models trained on real, noisy phone audio.
Where this is going
We do not think the answer is to remove people from customer service. The answer is to remove the menu. Let callers explain themselves, solve the simple things immediately, and give human agents the context to solve the hard things well.
Waridi Agents is in early access. If you run a contact centre or a support line in East Africa and want to try replacing a menu with a conversation, get in touch. Developers can start with the quickstart.