Why This AI Company Deleted Its Chat Interface

Every AI product seems to ship with a chat interface. Ask your data a question. Talk to your documents. Prompt your way to insights. Tabs, the $91M AI billing platform, made the opposite bet — and it paid off.

Deepak Bapat, CTO and co-founder of Tabs — which automates contract-to-cash for 200+ customers including Cursor and Statsig — decided early on that his finance users wouldn’t get a chat box. Not because the technology wasn’t ready. Because prompting was actively making things worse.

The Prompting Paradox

Tabs processes contracts through carefully constructed “context objects” — structured representations built during onboarding that carry each customer’s specific terminology, pricing structures, and business rules. These context objects are what push accuracy from the mid-80s into the high 90s.

The problem emerges when you let users prompt on top of that context. “We have found that when people try to prompt on top of the context object, they oftentimes cause more issues because now you’re giving differentiated advice from what the context object has,” Bapat explains. “And so these systems get confused, and you can actually generate more hallucinations.”

The user’s prompt introduces information that conflicts with the carefully calibrated context. The model tries to reconcile both inputs. The result is worse than if the user had done nothing at all.

What Finance Teams Actually Want

The decision to skip the chat interface came from observation, not ideology. Tabs tried it. Finance teams didn’t use it the way product teams expected.

“We made a bet that we didn’t want to build a chat interface in the platform. The reason for that is because we found that the background agent model worked better for what our finance teams want,” Bapat says. “Actually finance people don’t really want that.”

What they want is to click three buttons and fix a mistake. If a contract was processed wrong, the right experience is a correction interface — not a conversation. “When you prompt it, it gets it wrong. Now you’re already starting to erode real trust. This is why I think the UI experience is still incredibly important.”

The trust dynamic is different in finance than in, say, content generation. When a chat interface hallucinates a blog post, you rewrite it. When a chat interface hallucinates an invoice, the payment cycle resets and you lose months.

Background Agents Instead of Chat

Tabs replaced the chat paradigm with what Bapat calls “background agents” — ambient processes that run pipelines with built-in determinism. Instead of waiting for a user to ask a question, the system processes contracts, generates invoices, runs anomaly detection, and flags the 10% that need human judgment.

For users who do want conversational interaction with Tabs, the company built an MCP server. “We’ve kind of put all of our eggs into that basket rather than trying to build an internal chat interface, which we have found to be kind of a nice gimmick,” Bapat explains.

The MCP approach means users who want to talk to Tabs can do it through Claude Desktop or Claude Code — tools already designed for conversation. Tabs handles the backend pipeline. The conversation layer belongs to the tools that specialize in it.

The Correction Loop That Replaces Prompting

When Tabs processes a contract incorrectly, the fix isn’t a prompt — it’s a structured correction. The user adjusts fields in the platform, and those adjustments auto-calibrate the context objects for next time.

“Once they make changes, we will learn from that. We’re going to learn from that, and we’re going to improve the next time around. But they don’t want to go in and prompt the changes. Just go in and click three buttons and fix it.”

This creates a feedback loop that actually improves accuracy over time, unlike a chat interface where each prompt is a one-off interaction that doesn’t compound. The structured correction tells the system exactly what was wrong and what the right answer should be. A prompt gives the system a vague instruction filtered through natural language ambiguity.

FAQ

Why do chat interfaces cause AI hallucinations in enterprise software?

When users prompt on top of carefully calibrated context objects, they introduce conflicting information. The model tries to reconcile the user’s ad-hoc instructions with structured context built during onboarding. This conflict produces more hallucinations than running the pipeline without user prompts — a finding Tabs confirmed through production testing with finance teams.

What are background agents in AI products?

Background agents are ambient AI processes that run pipelines automatically without waiting for user prompts. At Tabs, they process contracts, generate invoices, run anomaly detection, and flag exceptions — handling 90% of work without human interaction. Users only intervene when the system escalates decisions requiring judgment.

How does Tabs handle AI mistakes without a chat interface?

When Tabs processes a contract incorrectly, users fix it through structured corrections — clicking buttons and adjusting fields, not writing prompts. Those corrections auto-calibrate the context objects for future contracts. This creates a compounding feedback loop that improves accuracy over time, unlike one-off chat interactions.

What is MCP and how does Tabs use it?

MCP (Model Context Protocol) lets AI tools communicate with external systems. Tabs built an MCP server so users who want conversational interaction can use Claude Desktop or Claude Code to talk to Tabs. This separates the conversation layer (handled by purpose-built tools) from the processing pipeline (handled by Tabs’ background agents).

Why don’t finance teams want AI chat interfaces?

Finance teams have high liability — CFOs and controllers sign off on assertions going to auditors and boards. Chat introduces unpredictability at the exact point where predictability matters most. When a chat hallucination produces a wrong invoice, payment timelines reset. Finance teams prefer clicking three buttons to fix an error over typing a prompt that might create new ones.

How does Tabs’ context object system work for contract processing?

During customer onboarding, Tabs builds structured context objects from all existing contracts and cross-references patterns from 200+ merchants. These objects encode business rules, pricing structures, and terminology specific to each relationship. Context objects push contract extraction accuracy from the mid-80s to the high 90s.

What is the difference between prompting and structured corrections in AI?

Prompting gives the AI a natural-language instruction that introduces ambiguity and can conflict with existing context. Structured corrections tell the system exactly which field was wrong and what the right answer is. Corrections compound into better future predictions. Prompts are one-off interactions that don’t systematically improve the system.

Should AI startups build chat interfaces for enterprise customers?

Not necessarily. Tabs found that chat was a “nice gimmick” that finance teams didn’t actually want. The decision depends on the domain’s error tolerance. In fields where mistakes have high cost — finance, healthcare, legal — background agents with structured correction loops outperform chat interfaces that introduce prompting-driven hallucinations.

How does human-in-the-loop work at Tabs without a chat interface?

Tabs toggles human-in-the-loop approval on by default for new features and customers. Merchants click an approve button on generated invoices and contract extractions. As confidence builds, some merchants enable auto-approval. The system provides accuracy metrics in-platform so merchants can see exactly how the AI is performing before trusting it fully.