Identify the AI Wrapper

A $69-per-seat demo looks like software. It is often a wrapper. Buyers need one clear test. Can the vendor show where the work happens? If not, assume a wrapper. Prove it otherwise.

Look at what vendors can’t fake: pricing, contract terms, demo behavior, and integration depth. Ask for model disclosure. Data flow. The logs. If answers stay airy, walk away.

Wrappers Hide Four Things

For due diligence, focus on ownership, not polish.

Wrappers hide vague pricing. They offer a tidy seat plan. Real usage hides behind credits, custom tiers. Jasper pricing starts at $69/month per seat, adds credits for advanced features (Jasper pricing). Notion bills AI interactions through credits deducted on message send (Notion pricing FAQs). This provides a clue. Does the price track actual work? Or just access to a front end?

They hide workflow ownership. The vendor can’t say where a task begins. Who approves it. What system it writes back to. How the handoff works. When you buy a wrapper, you get ChatGPT with a nicer skin. The underlying model quality doesn’t matter. The product must own the job.

Regarding data handling, ask for model disclosure. Training data. Data flow explanation. One breath. If they can’t say what happens to prompts, files, outputs, logs — the answer is no. A thin wrapper sounds confident but provides no substance.

Demos work only on perfect inputs. Clean prompt. Clean output. No interruptions. No weirdness. Real business work is messy. Failures, too.

Demo: Buyer's View

Run the same prompt twice. Back-to-back. OpenAI notes completions vary. Reproducibility depends on the setup rather than the demo vibe (OpenAI Help Center). First run magic, second turns to mush? You learned something.

Disturb the input. Add a typo. Reorder bullets. Paste a messy customer note. The Instruction Hierarchy paper is blunt about how real systems handle contradictions, odd formatting, and injected instructions. More gracefully than scripted ones.

Ask for failure modes. What can’t the system do? What does it refuse? What happens with an impossible prompt? A wrapper dodges the hard case. It keeps smiling. A real product shows bounded behavior. It says why.

Watch integration behavior. They claim to write to CRM, email, ticketing, phone systems? Ask for request logs. They won’t show them? Use browser/network inspection during the live demo. Seek proof that a workflow runs, rather than a stage show.

Latency patterns matter. Instant is fine. Suspiciously identical instant responses across vastly different tasks indicate a problem. Real work has shape.

Contract and Pricing: The Tells

Commercial language tells on the vendor. First, data rights. OpenAI’s business terms state that the provider will not use Customer Content to develop or improve services unless the customer agrees (OpenAI business terms). A strong baseline. The red flag includes terms like “use and store … to develop and improve,” “train,” “fine-tune,” or any broad, perpetual license to customer data.

Check deletion and export. Terms describe deletion within 30 days after termination — unless law or agreement requires otherwise. Services agreement points to “return or delete Customer Data” after expiry or termination (OpenAI business terms, OpenAI Services Agreement). No export path? No portability? No clear return-or-delete commitment? Lock-in is the product. Contract makes you solely responsible for trade/export compliance? Risk pushed downhill.

Pricing maps to work. Wrapper products often sell per seat, with credit packs. Workflow products charge per task or outcome. Zapier bills successful actions — tasks (Zapier pricing). Intercom’s Fin AI Agent advertises $0.99 per outcome (Fin pricing). These models show what the vendor sells. Compare with seat-based public plans like Airtable. Or hybrid per-user/usage pricing like OpenAI’s business and API offerings (Airtable pricing, OpenAI pricing).

Weak deletion. Fuzzy export. Broad customer-data rights. Migration cost shows later as pain, rather than money, indicating SLA risk.

Real Product Depth. Operational Proof.

A real workflow product learns the business, it doesn't just answer. Look for proprietary data. Feedback loops. Integrations. Business-specific rules. The system scores calls. Builds a shortlist. Drafts a launch note. Answers live chat with consistent context. It runs a process. It does more than wrap a model.

Look for artifacts such as Request IDs, output history, reproducible decisions, and clear rules for escalation or stopping. The best signal is less manual work over time, not a fancy interface.

ROI measurement applies here. Ask for a before-and-after metric, such as fewer missed calls, faster responses, cleaner handoffs, shorter hold times, better CSAT, higher shortlist quality, or more publishable drafts. Vendor can’t tie the tool to a business result? The product is decorative.

A live desk answers in Arabic and English. A newsletter researched, written, published, sent in public. A phone line picks up. These are operational signals wrappers struggle to fake. The work and the product are visible.

Decision Rule: Operators and Owners

Use a blunt scorecard. Score it as a Marketing Wrapper, Thin Wrapper, Workflow Product, or Differentiated System.

Test	Keep it	Walk away
Workflow ownership	Vendor shows where work starts, ends, and gets logged	Vendor only shows a prompt box and a pretty response
Demo behavior	Same prompt twice, messy inputs, and failure cases still make sense	Perfect-input magic, then collapse
Contract language	Clear export, short retention, narrow data rights	Broad use/store rights, weak deletion, no portability
Pricing	Tied to tasks, outcomes, or a known business unit of work	Seats and credits that drift away from value
Integration	Real APIs, request logs, stable handoffs	Hand-wavy “integrates with everything” claims

Passes the scorecard? Keep going. Only reskins a model? Stop.

The right AI vendor feels like a live work system. The wrong one feels like a front end with a speech bubble.

Buy the work. Leave the shell.