When AI Promises a Feature You Deprecated Last Quarter

The prospect's evaluation call went well. Your rep followed up the same day — thorough email, answered every question, mentioned the Zapier integration as a key fit for their workflow.

You deprecated the Zapier integration eight months ago.

Nobody caught it until the prospect replied asking how to set it up. Now your rep has to explain that the feature doesn't exist, the email was wrong, and the workflow they'd built their evaluation around isn't available. The deal didn't die — but it didn't recover cleanly either.

The changelog problem

Your product changes constantly. Features get deprecated. Integrations get sunset. Functionality gets renamed, rebuilt, or replaced. Pricing tiers get restructured. What was true about your product 12 months ago might be partially or entirely wrong today.

Generic AI has no changelog. It doesn't know what you shipped last quarter or what you removed. It knows what your product did — in aggregate, across everything it was trained on — and it generates descriptions based on that general understanding.

For a stable, mature product with minimal changes, that's a low-risk gap. For a growing SaaS company that ships and iterates constantly, it's a live accuracy problem in every customer-facing email your team sends.

Where it shows up

The feature commitment problem surfaces most often in three situations:

Sales follow-ups after demos. Reps are moving fast. They use AI to draft comprehensive follow-up emails covering everything discussed. The AI fills in product details from its general knowledge — including features the rep didn't explicitly mention, some of which may no longer exist.

Support emails about integrations. A customer asks if you support a specific integration. The AI knows you supported it at some point — or that tools like yours commonly do — and answers yes. Your integration was deprecated, or was never available on the customer's plan.

Competitive comparison emails. A prospect asks how you compare to a competitor on a specific capability. The AI generates a confident comparison that includes capabilities you once had, roadmap items that didn't ship, or feature parity claims that aren't current.

In each case, the commitment lives in the email thread. The customer acts on it. The correction costs more than the original answer.

The contrast

A prospect asks: "Does your platform integrate with Zapier? That's a hard requirement for us."

Generic AI draft:

"Yes, we integrate with Zapier — you can set up automated workflows to connect your inbox data with thousands of other apps. It's one of the most popular integrations our customers use for productivity automation."

Confident. Detailed. Wrong — if Zapier was deprecated.

KB-grounded draft:

"We don't currently have a native Zapier integration — that was sunset in Q2 2025. For workflow automation, we connect directly with Make (formerly Integromat) and support webhooks for custom setups. Happy to walk through what that looks like for your use case." Source: integrations-current.md, updated October 2025

The grounded draft pulled from your current integrations doc, surfaced the deprecation, and redirected to what actually exists. The prospect gets an accurate answer. The rep doesn't make a commitment they can't keep.

Why "just review the draft" isn't enough

The obvious response is: reps should check before sending. That's true — but it assumes reps know what they don't know.

A rep who joined six months ago doesn't have a mental inventory of deprecated features. They weren't there when the Zapier integration was removed. When the AI confidently describes it, they have no reason to flag it. The draft looks thorough, the tone is right, and they hit send.

This isn't a training problem. It's a grounding problem. The system the rep relies on doesn't know your current product state — so the rep can't know what to verify.

Source citations change the dynamic. When a draft claims an integration exists and cites integrations-current.md, updated October 2025, the rep has a specific thing to check. They open the doc, confirm the integration is listed, and send. If it's not listed, the discrepancy is visible before anything goes out.

What to do about it

Keep a current integrations and features doc. A single source of truth for what your product does today — not what it did last year. This is the document your AI should pull from, not its general knowledge.

Update it when things change. Deprecations, renames, new additions — any change that affects what you can promise in a customer email should be reflected in the doc, with a date.

Connect AI to that doc, not to its training data. A grounded system retrieves your current feature state before drafting. A rep using it can't promise a deprecated integration because the doc doesn't list it.

Inbox SuperPilot connects to your product docs, help center, and internal knowledge sources to create drafts for your review with inline citations on every factual claim. When a prospect asks about a specific integration or feature, the draft pulls from your current documentation — not from what your product used to do.

Try it free inside Gmail — no card required.

Further Reading & References

From the Inbox SuperPilot Blog

• What We Learned Building a Citation Engine for Email AI
• Gmail Gemini vs. Knowledge-Grounded AI: An Honest Comparison
• 5 Email Mistakes AI Catches That Humans Miss
• Superhuman Added a Knowledge Base. Here's What It Actually Changes.

External References

• Lost in the Middle: How Language Models Use Long Contexts — Liu et al., 2023. Why evidence position in retrieved context affects whether models actually use it — relevant to surfacing changelog and deprecation info.
• Factuality Challenges in the Era of Large Language Models — Augenstein et al., 2023. Covers how parametric knowledge limitations cause models to generate outdated or incorrect product information.
• Air Canada Held Liable for Its Chatbot's Misinformation — CX Today, 2024. Real case: a tribunal ruled Air Canada liable after its chatbot promised a policy that didn't exist. The precedent: companies own what their AI tells customers, regardless of accuracy.