AI Chatbot Platform for a Competitive-Intelligence Consultancy

The problem

The consultancy sells competitive intelligence. Their job is to research a software market — say, "the PRM (Partner Relationship Management) tool market" or "ESG reporting software" — and produce a detailed report comparing all the players, the features, the pricing, the gaps, the trends. SaaS founders, product managers, and enterprise sales teams pay for these reports because the information they contain genuinely changes business decisions.

The reports are 50–100+ pages long. They're packed with data. They're rigorously researched. They're extremely valuable.

And they have a problem.

Nobody reads them.

Or rather — nobody reads them all the way through. A prospect lands on the consultancy's website, sees a research report on a topic they care about, downloads the PDF, opens it, scrolls through the table of contents, reads the executive summary, maybe skims one or two sections, and closes it. The report does its job — they got the high-level picture — but the engagement is over. They have no reason to come back. They have no reason to talk to the consultancy. They've taken what they wanted and disappeared.

For the consultancy, this was the central problem of their business model:

• The reports were too long for the way prospects actually consume content in 2026. People want to ask a specific question and get a specific answer. Reading 80 pages to find one paragraph isn't how anyone works anymore.
• The download form captured an email but nothing else. Was this a real prospect? Were they researching a specific competitor? Were they evaluating a tool or building one? The form didn't know. The team didn't know. Following up was a cold reach-out with no context.
• The PDF was a dead end for the relationship. Once the prospect had the file, the conversation was over. There was no engagement signal, no interest signal, no qualification.
• The same questions came up over and over. Different prospects asked the same things ("Who are the main players?" "What's the pricing range?" "Which one is best for X?") — but every conversation started from scratch because there was no system holding the answers.

The founder saw the obvious answer. AI. Specifically: turn each research report into a chatbot that prospects could talk to. Let them ask their actual question. Let the bot answer from the actual report. Capture the lead in the middle of the conversation, when the visitor is most engaged. Use the conversation itself as the qualification signal.

But there were two specific things the founder wanted that no off-the-shelf chatbot tool could give him:

1. One platform, many bots. The consultancy produces multiple reports a year. Each one needs its own chatbot, trained on its own files, with its own welcome message and its own suggested questions. He didn't want to set up a separate ChatGPT subscription for every report. He wanted one admin panel that managed all the bots.

2. Lead capture as a first-class feature. Not a bolt-on form. The lead capture had to be part of the chat experience, configurable per bot — sometimes you want to ask the visitor for their details immediately, sometimes after a few questions, sometimes optionally at the end. And every captured lead had to be linked to the full chat history so the sales team could follow up with full context.

That second point is the part that turned this from "an AI chatbot" into "an AI sales tool."

---

The approach

There were three big decisions on this project, and they all came from one principle: build the system around the consultancy's actual sales motion, not around a generic AI chatbot template.

Decision #1 — Each research report = its own chatbot, but the platform is shared.

The lazy version of this product is "build a single ChatGPT-like interface and train it on everything the consultancy has produced." That breaks immediately. A prospect researching the PRM market shouldn't get answers contaminated by the ESG reporting research. Each bot needs to have its own knowledge, its own personality, its own welcome message, its own embed link.

So we built a multi-bot platform: one admin panel, one set of users, one CRM — but each bot is its own isolated entity with its own files, its own configuration, its own conversation history, its own leads. Adding a new bot for a new report takes minutes. Deleting a bot doesn't affect any other bot. Each bot can be embedded independently on a different page.

Decision #2 — Lead capture is built into the chat, not bolted on.

A "contact us" form on a website page captures 0.5% of visitors. A lead capture form that pops up after a visitor has already had 3 great answers from a chatbot converts much higher — because the visitor is now engaged, they trust the tool, and they want to continue.

We built lead capture as a configurable, in-chat moment. Each bot can be set to ask for the visitor's details:

• At start — before they can ask anything
• After 2 queries — give them a taste, then ask
• After 4 queries — let them really engage first
• At the end — just before they leave
• Optional — never required, always available

The consultancy team picks the right setting per bot based on the audience. A high-intent report ("PRM Vendor Comparison") might require details up front. A high-funnel report ("State of the SaaS Industry") might let visitors explore freely first.

When a lead is captured, it's linked to the full chat history. The sales team doesn't just see "John from Acme Corp signed up" — they see exactly what John asked, what the bot answered, how many questions he had, and what he was specifically trying to find out. That's a 10x warmer lead than a contact form.

Decision #3 — Use the best AI tool for the job, don't reinvent it.

We didn't try to build a search engine or a vector database from scratch. We used OpenAI's Assistants API with vector stores as the AI engine — which means each bot gets state-of-the-art document understanding, citations from the source material, and conversation memory, all built-in. We focused our engineering on the parts that matter for the consultancy's business: multi-bot management, configuration, lead capture, analytics, embedding, and the admin experience.

This is the right tradeoff for any client building a Gen.AI product in 2026. Don't compete with the AI providers. Build the layer of business logic that turns their general-purpose AI into a specific business tool.

---

What we built

A complete multi-bot platform

• One admin dashboard that lists every chatbot the consultancy has ever created
• Each bot shows its name, file count, status (active/inactive), date created
• Click any bot to manage its files, edit its instructions, view its conversations, see its leads
• Create a new bot in under 5 minutes — name it, upload files, configure, activate, share
• Soft-delete + reactivate so old bots can come back
• Share link generation for embedding bots on any page

Per-bot configuration

Each bot has its own settings:

• Custom instruction — how this specific bot should behave
• Bot description — the welcome message visitors see when they open the chat
• Predefined questions — suggested first prompts to get visitors started ("Who are the competitors?", "Summarize this report", "What's the pricing landscape?")
• Query limit — how many free questions before requiring sign-up
• Lead capture timing — start / after 2 queries / after 4 queries / end / optional
• Active status — toggle on or off without deleting

A master instruction layer

On top of per-bot config, there's a platform-wide master instruction that applies to every bot. This is where the consultancy sets the global rules: "Answer only based on the documents. Don't make things up. If you don't know, say so." Each bot inherits this and can layer its own instructions on top.

This is the kind of safety net that matters for a business chatbot — it stops the bot from confidently making up answers it doesn't actually know, which is the failure mode that destroys trust.

File management per bot

• Upload PDFs, documents, datasets to any bot
• Each file is tracked with its filename, upload time, and the bot it belongs to
• Files are processed and indexed automatically — visitors can ask questions and the bot answers from them
• Replace or delete files as the report gets updated

The conversation experience

When a visitor opens a bot, they see:

• The bot's name and welcome message
• Suggested first questions (the predefined ones the admin set up)
• A clean chat interface
• A "thinking" indicator while the AI processes their question
• Source citations (where in the document the answer came from)
• Smooth, fast responses

They can ask anything. The bot answers from the actual research files — not from general knowledge. If the answer isn't in the documents, the bot says so politely instead of making something up.

Lead capture done right

The lead capture form is clean and fits into the chat naturally:

• Full name
• Work email
• Product & competitors ("What are you working on, and who are your competitors?")

That third field is the key one. It's not a generic "tell me about yourself" — it's specifically "what are you researching?" — and the answer is exactly what the consultancy needs to know to qualify the lead. A SaaS founder building in the PRM space who's evaluating two specific competitors against each other is a much warmer lead than "interested in PRM software."

The CRM view

Open the CRM view and see every lead captured across every bot:

• Name + email
• Which bot the lead came from
• Their product and competitor context
• A unique lead ID
• The session it was captured in (with full chat history just one click away)

The consultancy sales team can sort, filter, and follow up — and every follow-up email starts with the actual context, not a cold "I see you downloaded our report."

Analytics

• Total conversations per bot
• Total leads per bot
• Active vs inactive bots
• Most-asked questions across each bot
• Session-level details (browser, OS, country, IP) so the team knows where prospects are coming from
• Aggregate analytics across all bots

Session tracking

Every chat session is stored with:

• Session ID
• The bot it belongs to
• The visitor's user details (if captured)
• IP address
• Browser, OS, country
• Start time and end time
• Full message history (every question and every answer, in order)

This is the data that powers both sales follow-up and product intelligence.

JWT-based admin authentication

• Secure login for the consultancy team
• Token-based with refresh
• Profile management
• Multiple admin users supported

Daily database backups via Telegram

Same production-grade pattern as the other live projects: every day the database is backed up and pushed to a private Telegram channel. Conversations, leads, files, configurations — all safe even if the server has a bad day.

---

The hard part — and how we solved it

There were two genuinely hard parts on this project, and they both came from the same source: AI is easy to make demos with and hard to make production-grade.

Hard part #1 — Making the bots actually trustworthy

The single thing that destroys a business chatbot is hallucinated answers. The bot makes up something that sounds confident but isn't actually in the source material. The prospect quotes it back in a sales call. The consultancy team has to walk it back. Trust gone.

We solved this with three layered safeguards:

1. A platform-wide master instruction that tells every bot to answer ONLY from the documents and to explicitly say "I don't know" if the answer isn't there.
2. Per-bot instructions that can reinforce the same point in the bot's specific context.
3. OpenAI's Assistants API with vector store grounding — meaning the bot actually retrieves passages from the document before answering, rather than answering from general knowledge.

The result: when a visitor asks "Who are the competitors covered in this report?", the bot looks up the actual document, finds the actual list, and returns it. When a visitor asks "What's the average team size at Series B SaaS companies?" — and the report doesn't cover that — the bot says so, instead of making up a number.

This sounds simple. It's the difference between a chatbot the consultancy can stand behind and one they can't.

Hard part #2 — Lead capture timing

The default behavior of every "chatbot for sales" tool is to either (a) require an email up front (which kills 90% of conversations) or (b) never ask for an email at all (which captures 0% of leads). Both are wrong.

The right answer is configurable timing per bot. A high-intent visitor researching a specific product comparison should be asked for their details immediately — they're motivated, they'll convert. A high-funnel visitor browsing a market overview should be allowed to explore freely first, then asked once they've shown engagement.

We built the timing as a per-bot setting with five options (start / after 2 queries / after 4 queries / end / optional). The consultancy team can experiment with each bot independently and figure out which timing converts best for which audience. The query counter is tracked per session so the prompt fires at the right moment.

The other piece of this is what happens after the visitor submits the form. A lot of chatbots end the conversation when the form is filled. We don't — the form is a checkpoint, not an exit. The visitor submits their details and continues chatting. The conversation is now linked to their identity, the lead is captured in the CRM, and the sales team has the full context. The visitor doesn't feel like they got blocked — they feel like they unlocked something.

---

The outcome

• Live in production at the consultancy running multiple real competitive intelligence reports as chatbots
• Real bots running today: ESG Reporting Software Market (for an ESG client), a client mega bot, Retail Sector, Partner Relationship Management, Contract Lifecycle Management, Low-Code Application Platform, and more
• One admin panel managing all bots — create, configure, monitor, deactivate, reactivate
• Lead capture as a first-class feature with configurable timing per bot (5 options)
• Every lead linked to full chat history so sales follow-up has actual context
• A CRM view for managing leads across all bots
• Per-session analytics with browser, OS, country, query count, conversation history
• Master instruction safety net preventing hallucinated answers
• Soft delete + reactivate so old bots can come back
• JWT admin auth with multi-user support
• Daily automated database backups to Telegram

The bigger outcome: a competitive intelligence consultancy stopped relying on PDF downloads and started having actual conversations with their prospects. Every conversation now produces qualified intent data. Every research report is now a sales motion, not a static deliverable. The "I downloaded it and disappeared" pattern is replaced with "I asked specific questions and the system knows exactly what I'm trying to figure out."

This is what AI looks like when it's used to solve a real business problem instead of being slapped on a wall as a feature.

---

Client testimonial

"While it's still early in terms of analytics and outcomes, the feedback was positive: the chatbot was easy to use, and outperformed static formats in terms of reception." — The founder of the consultancy