A weekend as our own first customer, and what our AI team told us to do next

On 27 and 28 September our founder stopped building Dhanur AI for a weekend and used it. He set up a workspace for his own website, nishgaba.com, connected his LinkedIn and Google Calendar, and put an assistant on the site that answers visitors and books calls with him. Posts it drafted went out on his LinkedIn profile, one with a photo, after he approved each one.

It was the first time anyone had used the whole product end to end, for their own business, on their real accounts. This post covers:

• what confused him;
• what we built and tested in response;
• what we got wrong;
• the full discussion with our AI employees about how to reach more businesses and help them understand the product.

What confused him

He wrote these down as he went. Most of them are fair.

1. "It's so disconnected." The app was organised by tool: Agents, Tools, Workforces, Knowledge, Tasks, Integrations. He thinks in outcomes: getting clients, posting on LinkedIn, answering enquiries. "Your AI agent is ready" told him nothing. Ready for what?
2. No guidance. He asked for the guided tours "a lot of the big sites have".
3. Home didn't show what he had. He couldn't see his assistants or what was connected.
4. "Publish" read as "publish a post". It means "make the agent live", which never posts anything.
5. No way to add a photo to a LinkedIn draft. When we added one, the draft card showed the wrong state while it was finding a photo. He called it out: we hadn't checked the states.
6. Lost after publishing. "Imagine I have contributed to the development. What about a new user?"
7. No view of the week. He wanted to see the weekly plan and what his AI team had booked and posted.
8. Our AI employees weren't at the top of our own company page, the page that's meant to show them.

What we built and tested in response

All of it went live the same weekend. The changelog lists each change.

• Start from a goal. Home now asks what you want help with:

  • get clients and book calls;
  • post on LinkedIn and Instagram;
  • answer every enquiry;
  • follow up leads.

  Picking one sets up what it needs, then shows a checklist. Each step says whether it's live, needs you, or isn't available yet, and why. How it works

• Guided tours. Every step that needs you opens the right page and points at the button. Published agents get Show me around. We used driver.js, an open-source library that runs in your browser and sends nothing anywhere.

• "Your setup" at the top of Home. Your website, connected apps and assistants, with Draft or Live.

• "Publish agent", not "Publish". Also Publish workforce and Publish tool.

• Post to your LinkedIn.

  • Each draft waits for your OK, and you can change the words first.
  • For a photo, choose Find a stock photo (free Pexels photos, with the photographer's credit) or Use my own photo.
  • While a draft is changing, it's locked and can't be approved. Only one version waits at a time. Post to LinkedIn

• "It's live" on every published agent. Where people reach it, where its work shows up, and a line explaining each tab.

• Calendar. A week at a time, in India time: scheduled runs, calls your agents booked, and posts that went out. See your AI team's week

• Google Meet links on bookings, when you turn them on.

• A booking assistant on nishgaba.com. A chat bubble and a chat link. Both are the same product anyone can set up.

How we tested it.

• Every change ships with automated tests. The engine's suite had over 1,700 passing at the last full run.
• The founder then did each thing for real on his own accounts: connecting LinkedIn and Google Calendar, posting text and a photo, and trying the booking assistant on his website.
• A person on our team has a 24-point test round to run on the live app, from goals to bookings. They'll also read every public page we changed, the way a visitor would.

What we got wrong

This matters most, because it's the kind of mistake an AI company is most tempted to make.

• An agent said something was done when it wasn't. A test agent told the founder a LinkedIn post was "waiting for your approval" when it had never drafted one. Another said a call was booked when it wasn't. Both are fixed at the root: before a reply goes out, we check it against what the agent actually did. If it claims a post is waiting or a call is booked without having done it, it goes back once to do it properly. Otherwise the reply says plainly that it didn't happen. What the agent says matches what it did

• We said "waiting on Meta" about Instagram. We had never applied. Nothing was waiting. Goal steps like that now say Not yet, and say what to do instead.

• Our public pages had drifted:

  • They still offered making images with AI after we paused it.
  • They listed booking as "planned" weeks after it went live.
  • They listed "paid plans" as coming, though we decided not to have monthly plans.
  • They said WhatsApp was "next" when it's already in a pilot.
  • They said "free during early access" when new workspaces actually start with ₹25 of free credits and then pay as you go.

  We corrected all of these on 28 September.

Asking our AI team what's next

Then the founder asked the team: what should we do to expand the knowledge and reach of the product? By knowledge he meant how well small-business owners understand what Dhanur AI does and how to use it. By reach, how many of the right people find us.

How this discussion ran.

1. We gave six of our AI employees the same brief:
   • the founder's notes;
   • a fresh walk through the product;
   • what we'd learned from reading the public pages of other tools;
   • which of our own plans are built and which aren't.
2. Each answered from its own role and instructions, the same ones it runs on in production.
3. Three more answered the decisions that came out of it, and Meera, our chief of staff, turned the whole thing into a memo for the founder.

This round ran in our development environment, not on the live system, so it isn't in the activity feed. The predictions below are the employees' own forecasts, not results. Each one wrote down what would prove it wrong.

What we learned from other tools

We read the public pages of Relevance AI, Zapier, Pipedream, Predis, Submagic, and the Indian WhatsApp tools Wati, Interakt, AiSensy and Gallabox. We also read Meta's plans for businesses on WhatsApp. We didn't copy anything. Here are the patterns that held up across several of them:

• The best first screen asks what you want done, then offers starter kits for your trade.
• In India, partners bring customers, not ads. The WhatsApp tools grow through agencies and consultants on recurring commission.
• A page per task brings people from search. Free tools only help when they sit close to the product.
• Billing surprises cause the harshest reviews in this category. Show the balance all the time, and never charge before people expect it.
• Owners want to approve work where they already are. In India that's WhatsApp.
• Channels go stale. Pick a few and keep them alive.

What each employee suggested

Tara, product. Make goals the spine of the app:

• Goals first in the sidebar, with the builder's tools in their own section.
• One checklist instead of two, and one clear next step on Home.
• One word for the thing you set up.

Later, fill in your business details from where you already keep them, such as your Google Business Profile. And send a Monday note for each goal: what your assistant did, what it couldn't answer, and what you corrected.

Raghav, support. Stop the questions before they come:

• Put Ask us on WhatsApp in the Help menu and on every step that needs you.
• Show your credit balance all the time.
• Add troubleshooting for running out of credits and for customers not being able to reach your agent yet.

Diya, community. Organise the docs by goal, not by tool. Every week, answer the most-asked question in public, in English and simple Hindi. None of our docs pages is in Hindi yet, and they should be.

Vikram, sales. Start with the people our product already fits end to end: consultants, coaches and small agencies who win work on LinkedIn and close it on a call. Small online brands come second. Clinics and coaching centres come later, when WhatsApp is self-serve, because that's the first thing they ask for. Offer five businesses a hands-on setup as design partners. Use the founder's own site as the live demo, rupee cost included.

Ishaan, marketing.

• A weekly post about what confused our first real user and what we changed.
• A free check: paste your website and see what an assistant could answer from it.
• Three honest comparison pages that say where the other option is better.
• Drop X: one live LinkedIn channel beats two half-alive ones.

Neel, partnerships. Make the partner programme real. It has terms (20% of what a referred business pays, for 12 months) but no page and no codes, so nobody can earn yet. Local digital agencies and web developers come first. Give partners a kit that teaches only what's true.

Where they disagreed

• How many search pages? Vikram wants none until design partners are live. Ishaan and Tara want a small set now. The plan: three comparison pages and pages for the first segment now; the rest later, written from design partners' real questions.
• A course or academy now? Diya says yes. Tara and Neel say later, and for partners first. The plan: goal walkthroughs and a partner kit now, the course after.
• Should Ishaan's posts keep going out on their own? Raghav and Vikram said no, not while our pages were out of date. Ishaan agreed, and asked for a check that holds any draft naming something that isn't live yet. We think that's right.

What the founder decided, and what's still open

Decided: no more free access. Our 17 September posts said early access would be free, with paid plans to come later. Ananya and Raghav pointed out that this no longer matched what we do, so the founder settled it. Every new workspace starts with ₹25 of free credits and no card. After that you pay as you go for what your AI team uses, in rupees plus GST, with a tax invoice for every top-up. There are no monthly plans. Both 17 September posts now carry a dated note saying so.

Still open, each with the team's recommendation and what happens if the founder says nothing:

• Partner setup fees: Neel recommends allowing them if the client is told first, partners never resell our credits, and a hidden fee pauses the partner's code.
• Design partners: Ananya, our finance employee, costed the offer. She estimates the worst case for five design partners at about ₹6,400.

What happens next

Update, later on 28 September: goals are in the sidebar, your balance is always in view, you can ask us on WhatsApp from the Help menu, and Ishaan's drafts wait for the founder when they mention anything that isn't live. See the changelog.

Next up:

• the goals-first changes to the app;
• your credit balance always in view;
• WhatsApp in the Help menu;
• the first of Ishaan's weekly posts;
• a design-partner offer, once the founder decides its terms.

After that:

• the partner programme;
• docs organised by goal;
• the Monday note;
• the first comparison pages.

We'll write up what worked and what didn't, with numbers, the same way.

If you run a small business and want to try this for yourself, talk to us.