Check four things before paying: the vendor's sub-processor list and privacy policy (a wrapper names OpenAI, Anthropic, Google or a cloud host of their models), whether its output matches ChatGPT's or Claude's for the same prompt, whether it refuses the same requests in the same words, and what it adds beyond the model itself.
Being a wrapper isn't a flaw in itself. Most AI features in business software call a model from one of a handful of providers; what matters is whether the tool's price reflects what it adds, such as your data, integrations, a workflow built around your job, review steps and support. The model is usually the cheap part: a typical user's month of drafting can cost the vendor a dollar or two. If the tool adds little beyond a prompt, you're paying a markup for something you could set up yourself.
Why almost every AI business tool runs on someone else's model
Training a leading language model costs sums no small software company can spend, so most AI products rent one. They send your text to a model provider's API (a programming interface that lets software send requests and get answers back), pay per token (roughly three-quarters of a word; a million tokens is about 750,000 English words) and wrap the result in their own interface. "Wrapper" is the unkind name for this. It covers a wide range:
| Type | What the vendor built | What you're really paying for |
|---|---|---|
| Thin wrapper | A prompt and a nicer screen around a rented model | Convenience; often little you couldn't do in a general assistant |
| Workflow product | Integrations, your documents as reference, approvals, templates, audit logs | Time saved in your actual process, plus support |
| Specialist model | A model trained or tuned for one domain, sometimes alongside a rented one | Accuracy on a narrow task a general model handles badly |
The five stages below tell you where on that range a tool sits. Allow about an hour and a half in total, spread over a trial period.
Stage 1: read the sub-processor list and privacy policy (15 minutes)
Vendors that handle business data usually publish a list of sub-processors, the other companies that process your data on their behalf. It's typically linked from the privacy policy, the data processing agreement or a trust or security page. Open it and search the page for these terms:
- OpenAI, Azure OpenAI, Anthropic, Claude, Google Cloud, Vertex AI, Gemini, Amazon Bedrock, Mistral
- "large language model", "LLM", "third-party AI provider", "model provider"
An illustrative extract from a listing tool's sub-processor page:
Sub-processor Purpose Data involved
Amazon Web Services Hosting and file storage All customer data
OpenAI Generation of AI text features Prompts, property details
Postmark Transactional email Email addressesThat second line answers the wrapper question on its own: the writing feature is built on OpenAI's models, and your property details leave the vendor to reach them. It also points you to the right follow-up. OpenAI states that data sent through its API isn't used for training by default. Anthropic, since June 2026, keeps prompts and outputs on its most capable models for 30 days even for customers with zero-retention agreements, unless they apply for an exemption. Which provider sits behind a tool therefore changes your answer on data handling. What to check in an AI vendor's data processing agreement covers the contract side.
If there's no sub-processor list at all, note it. For a tool that will see client data, that's a bigger concern than whether it's a wrapper.
Stage 2: run the same task through the tool and ChatGPT (30 minutes)
Take three real tasks the tool is meant to do and run each one through the tool and through a general assistant such as ChatGPT or Claude, giving the assistant the same inputs and a similar instruction. Don't expect identical text: models vary their wording from run to run and the vendor adds hidden instructions. Look for shared habits instead: the same structure, the same favourite phrases, the same factual gaps, the same mistakes.
An estate agency trialling a listing-description tool ran this test with a three-bedroom semi. The input to both:
3-bed semi-detached, 1930s, extended kitchen-diner, south-facing garden
around 60 ft, off-street parking for two cars, new boiler 2024,
EPC rating C, walking distance to primary school and station.Illustrative openings from each:
Listing tool: "Nestled in a sought-after setting, this charming 1930s
semi-detached home offers the perfect blend of period
character and modern living..."
General assistant, asked for a property listing:
"Nestled in a popular location, this charming 1930s
semi-detached home blends period character with modern
comforts..."The same opening word, the same "charming", the same period-meets-modern angle, and both invented "period character" that the input never mentioned. Across all three test properties the tool's copy followed the general assistant's shape almost paragraph for paragraph. That doesn't prove which model sits underneath, but it tells the agency the tool's writing adds little that a well-briefed assistant wouldn't. Both also needed the same fix: a rule to use only the features listed.
A tempting shortcut is to ask the tool what model it runs on. Treat the answer as a hint at best. Models can misreport their own identity, and a vendor may have told its tool not to say.
Stage 3: probe refusals, limits and error messages (20 minutes)
Wrappers show their seams where things go wrong. Try these during a trial:
- Refusals. Ask for something the model makers restrict, such as a fake five-star review. Compare the refusal wording with ChatGPT's and Claude's. Identical phrasing, or a message that mentions a provider's usage policies, is a strong clue.
- Oversized inputs. Paste in an extremely long document. Error messages that mention tokens, a context length or a context window are model-provider language that has leaked through.
- Error details. Some tools show raw error codes. Strings that start with a model family name, such as "gpt-" or "claude-", settle the question.
- Timing of outages. If the tool slows or fails when a model provider's public status page reports an incident, it depends on that provider. That matters less as proof and more as a risk: does the vendor have a fallback?
That last point matters in practice. Picture an illustrative courier firm whose "AI quote assistant" went dark for an afternoon; the vendor's status page said nothing, but the model provider's showed an incident at exactly that time. The firm hadn't known the tool depended on one provider, and the vendor had no fallback model. Quotes went out late and by hand. Nobody had asked the question before signing.
Stage 4: price what the wrapper adds on top of the model
Once you know or suspect the model, you can estimate what each month of your usage costs the vendor. Provider list prices per million tokens, input then output, as of September 2026:
- OpenAI gpt-6-luna: $0.10 and $0.50; gpt-6-sol: $2 and $10; gpt-6-astra: $10 and $50
- Anthropic Claude Haiku 4.5: $1 and $5; Claude Sonnet 5: $2 and $10
The estate agency's tool costs $99 a month (an illustrative price for this worked example). The agency lists about 60 properties a month and usually generates three versions of each description before one is approved, so 180 runs. Each run sends about 3,000 tokens (the spec, notes and the vendor's hidden instructions) and gets back about 800.
| If the tool runs on | Input: 540,000 tokens | Output: 144,000 tokens | Model cost per month |
|---|---|---|---|
| gpt-6-luna | $0.05 | $0.07 | about $0.13 |
| Claude Haiku 4.5 | $0.54 | $0.72 | about $1.26 |
| Claude Sonnet 5 | $1.08 | $1.44 | about $2.52 |
| gpt-6-astra | $5.40 | $7.20 | about $12.60 |
Photos add a little more if the tool reads them. Anthropic's documentation puts a one-megapixel image at about 1,300 input tokens, so 15 photos for each of 60 listings is about 1.2 million tokens, or roughly $2.30 a month on Sonnet 5. Even on the priciest model in the table, text generation costs under $13 of the $99; the other $86 or more pays for everything else: hosting, development, support, sales and profit.
That's normal for software, and it isn't a reason to walk away. It's a reason to ask what the rest buys. Compare the alternative: ChatGPT Business needs at least two seats, $40 to $50 a month for two people, and gives them a general assistant for everything else too, with a shared project holding the agency's style rules and a list of banned phrases. The project instruction can be as short as this:
You write property listings for our agency.
Use only the features in the details provided. Never add period
features, views, school catchments or distances unless stated.
Banned words: nestled, charming, stunning, sought-after, boasts.
Structure: headline under 70 characters; three paragraphs covering
layout, outside space and location; end with how to book a viewing.
Keep to 180 words.Run it on the same three test properties you used in Stage 2 and compare the drafts side by side with the tool's. If the listing tool only writes descriptions, the general assistant wins. If it pulls each property's details from the agency's CRM, formats them for the property portals, checks descriptions against the recorded features and routes drafts for approval, it's saving a negotiator maybe 20 minutes a listing, about 20 hours a month, and $99 is cheap. Wrapper or not, the test is the workflow.
Per-item pricing hides the same gap. A wholesaler was quoted an illustrative $0.40 per product description by a catalogue-writing tool, for 2,000 descriptions: $800. Each description needs about 1,500 tokens in (the product data plus instructions) and 300 out. On Claude Sonnet 5 that's 2,000 x (1,500 x $2 + 300 x $10) / 1,000,000 = $12 in model costs; on gpt-6-luna it's about 60 cents. The tool could still be worth $800 if it reads the wholesaler's product feed, applies the house style, flags missing attributes such as pack size or material, and exports straight back into the web shop. If staff would be pasting data in and copying text out by hand, the same result costs a couple of evenings with a general assistant and a good template.
Stage 5: put five questions to the vendor in writing
Ask by email so the answers are on record. Here is the estate agency's list, with the answers it received (illustrative) and how it read them:
| Question | Answer received | Reading |
|---|---|---|
| Which model providers process our data, and are they on your sub-processor list? | "We use a blend of leading models, including our proprietary AI." | Evasive. The sub-processor list names OpenAI and nothing proprietary. |
| Is our data used to train any model, yours or the provider's? | "No. We use the provider's API, which doesn't train on our data by default, and we don't train our own." | Clear and checkable. |
| What happens when your provider retires or changes a model? | "We switch automatically." | Weak. Ask whether they test your templates first and tell customers. |
| Can we bring our own API key, and does the price change? | "Not currently." | Fine, but you can't cut the markup later. |
| What does the tool do that a general assistant with our documents can't? | "CRM import, portal formatting, approval flow." Demonstrated live. | This is the real value; verify it with your own data. |
Bring-your-own-key options, where you supply your own model account and the vendor charges less, are worth asking about because they turn the model cost into something you can see. Zapier, for example, counts an AI step as a single task when you connect your own API key, rather than the three or five tasks its dearer model tiers use. More vendor questions, beyond the wrapper issue, are in questions to ask an AI vendor before you sign.
When a wrapper is exactly what you should buy
A storage facility shows the other side. Its website chat assistant runs on a rented model, which its sub-processor list says openly. But the assistant checks live unit availability in the facility's management software, quotes the current price for each unit size, books viewings into the site calendar and hands anything about arrears or access codes to staff. Rebuilding that in a general assistant would mean connecting the management software, the calendar and the website, then maintaining all of it. The markup on the model is the least interesting number in that decision; the build-or-buy question matters far more.
Signs you're looking at a wrapper worth paying for:
- It connects to systems you already use and keeps data in sync both ways.
- It has controls a general assistant lacks: approvals, audit logs, per-user permissions, usage limits.
- The vendor names its model providers, explains its data handling and has a plan for model changes and outages.
- Its price is in proportion to the time it saves you, whatever the model costs.
Signs of a thin wrapper priced as something more:
- Output that tracks a general assistant's almost word for word, with no integrations to justify the price.
- "Proprietary AI" in the marketing and a single model provider in the sub-processor list. There's more on that pattern in how to spot software that oversells its AI.
- No answers on training, retention or model changes.
- A monthly price many times the model cost for work your team could do in a shared project in an afternoon.
Put the five stages together and most tools land in one of four boxes:
| What you found | Price against value | What to do |
|---|---|---|
| Thin wrapper | Price well above the time it saves | Set up the same job in a shared project in your general assistant; revisit if volume grows |
| Thin wrapper | Cheap and saves real time | Buy monthly, and keep your own copy of the prompts and examples you feed it |
| Workflow product | Priced in line with the hours it saves | Buy after a trial on your own data and a read of its exit terms |
| Can't tell | Vendor won't answer model or data questions | Keep client data out of it; ask again in writing or walk away |
If the tool will run inside an automation rather than a chat window, you can price the direct route too; what the ChatGPT API costs for a business automation walks through the sums. Whichever way you go, write down the model provider behind each AI tool you pay for. When a provider has an outage, changes its terms or retires a model, you'll know within a minute which of your tools are affected.
Further reads
- How to Evaluate an AI Software Vendor: A Small Business Scorecard — A full scorecard for any AI vendor on your shortlist.
- How to Run an AI Software Demo So You See the Real Product — Run a demo that shows the real product, not the script.
- AI Vendor Lock-In: How to Keep Your Data and Prompts Portable — Keep your prompts and data portable if you switch.
- Questions to Ask Before Buying AI That Touches Client Data — Extra questions when the tool will see client data.
- SOC 2 and ISO 27001 Explained: Checking an AI Vendor's Security — What security certificates do and don't prove.
- How to Check an AI Software Vendor's Customer References — Talk to real customers before you sign.
- What to Do Before You Buy Any AI Tool: A 10-Point Checklist — Ten checks to run before paying for any AI tool, from pricing the job it replaces to reading the exit terms, with a kitchen-fitter worked example.
- Best AI Tools for Etsy Sellers, Ranked by Time Saved — Seven kinds of AI tool ranked by the hours they save a small Etsy shop each week, with prices, a worked example for each, and the tools not worth paying for.
- Is Your AI Consultant Independent? Checking for Tool Commissions — The referral terms vendors publish, a sum of what a recommendation is worth to a consultant, six questions that get straight answers, and a disclosure clause.
- Are AI Lifetime Deals Worth Buying for a Small Business? — What lifetime really means on a deal page, why per-use AI costs strain one-off prices, a payback test, and an exit plan to write on day one.
- AI Consultant Red Flags: 12 Warning Signs to Walk Away From — Twelve warning signs when hiring an AI consultant, each with an example, the question to ask, its innocent version and a scored two-proposal comparison.
- AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.
Sources: OpenAI and Anthropic API pricing (per million tokens) and data-use terms; Anthropic documentation on image token costs; Zapier help on AI steps with your own API key. Checked September 2026.