Why an assistant should finish the work.
Most assistants stop at advice. The useful part is finishing the task, and that is where an assistant should spend its effort.
Ask most AI assistants to renew your vehicle insurance and you will get a tidy answer. It will list a few insurers, explain what a no-claim bonus is, and suggest you compare quotes before the policy lapses. All of it is correct, and none of it renews the insurance.
The work is still yours. You still open three websites, type the registration number three times, wait for an OTP, discover that one site wants the previous policy number, find that policy in your email, pay, download the PDF and put it somewhere you will remember. The assistant handled the part that took two minutes. You handled the part that took forty.
Where the time actually goes
When we looked at the tasks people sent to AskGogo, the pattern was consistent. Deciding what to do was rarely the hard part. People knew they needed a service appointment, a refund, a visa slot or a follow-up email. What they lacked was the time and patience for the steps in between: forms, logins, waiting, checking back.
An answer does not remove those steps. It describes them. So we set a different bar: an assistant should be measured by tasks completed, not questions answered.
A good answer describes the work. A good assistant does it, then shows you the receipt.
Finishing takes three things
Completing real tasks is harder than answering questions, and it rests on three capabilities that each have to work.
- Memory. To book your car service without twenty questions, the assistant needs to know which car, which service centre and that you prefer mornings. Without memory, every task starts with an interview.
- Autonomy. It has to operate the same websites and inboxes you do: sign in, fill the form, handle the OTP prompt by asking you for the code, and carry on.
- Judgement about when to stop. Finishing does not mean doing everything alone. Paying, booking and messaging someone else are steps where the assistant should pause and ask, with the details in front of you.
Take away any one of these and you are back to advice. Memory without autonomy produces a well-informed list. Autonomy without memory produces a stream of questions. Either without judgement produces something you cannot trust.
What done means
We define a task as done only when there is evidence the world changed: a booking reference, a calendar entry, a sent email in the thread, a saved receipt. If the assistant cannot observe that result, the task is not done, and it should say so plainly.
That standard is stricter than it sounds. Websites time out. Payment pages fail silently. A form accepts input and then quietly drops it. An assistant that reports success because it clicked the final button will be wrong often enough that you stop relying on it. One that checks the outcome earns the right to be handed the next task.
Why this matters more in India
Much of daily life here runs through portals that were not built for convenience: logins by mobile number and OTP, government appointment systems, utility sites with their own logic. WhatsApp is where people already manage their lives. An assistant that lives there and can handle the portal on your behalf, asking you only for the OTP and the final yes, removes a real and recurring cost.
That is the assistant we are building. It does not try to be impressive in conversation. It tries to leave you with fewer open loops at the end of the day than you had at the start.