An AI that prepares quotes in the inside sales team of a medical technology supplier is manageable as long as prices and stock levels come from the ERP and product texts come from approved templates. It becomes a legal risk the moment it casually answers a question about product suitability or attributes an effect to a wound dressing. If you draw this line cleanly, you keep control over every word that goes to the customer.
Every head of inside sales knows the work in question: typing up item numbers from PDF enquiries, looking up terms, checking delivery times in the ERP, copying standard texts. And the moment when an enquiry from a group purchasing organisation sits untouched for three days because the only colleague who knows its terms is on holiday. An AI can prepare this work. Which enquiries it may touch and which never is decided by three questions.
What the BVMed autumn survey 2026 says about sales
When it comes to staffing, sales is far ahead: 78% of companies are looking for people there, and no other area reaches even a third. Of the 220 voting members of the BVMed (German Medical Technology Association), 120 took part in the autumn survey in August and September, manufacturers as well as specialist retailers and wholesalers. The response rate is 55% (summary of results of 6 October 2026).
More than 80% have problems filling vacancies, again most often in sales (42%). 52% cite price pressure from group purchasing organisations and hospital chains as a cost driver.
49% of the companies already use AI in sales; in marketing the figure is 69%. So almost every second company has AI in sales, and almost four out of five are still looking for salespeople. The survey does not break down what AI is used for in sales. My reading: as long as it helps with wording and research, it does not take a single quote line off inside sales’ hands.
Catalogue items yes, product suitability never: which enquiries are suitable
What matters is whether the draft is correct without interpretation. Three questions are enough to tell:
- Is the product clearly identifiable, by item number, manufacturer reference or UDI?
- Is the price in the system, in a price list, framework agreement or customer terms?
- Does the enquiry contain a clinical or application-related question that requires a specialist answer?
If the answers to the first two questions are yes and the answer to the third is no, the draft is automated and a person approves it.
Two made-up enquiries show the difference. The first: “40 packs of nitrile examination gloves, size M, item no. 4711-M, delivery to the central warehouse by the end of the month.” Product unambiguous, price in the terms, no specialist question: automatable. The second: “Is your catheter X also suitable for children under 10 kg, and what does the pack of 50 cost?” It looks like a price enquiry, but it contains a suitability question and therefore belongs in the last row of the table.
| Enquiry type | What the AI prepares | Data source | Approval |
|---|---|---|---|
| Catalogue item, price and availability | Draft quote with delivery time | ERP price list, stock, customer terms | Visual check |
| Reorder of consumables | Draft from order history, note on batches and shelf life | ERP order history, warehouse management | Visual check |
| Call-off from a framework agreement with a hospital or group purchasing organisation | Draft with contract prices, check of term and volume tiers | Contract data in the ERP or CRM | Specialist review |
| Configurable devices, accessories, custom-made products | Query to the customer, proposed bill of materials | Product configurator, master data | Specialist review and costing |
| Enquiry with a clinical question or about product suitability | Summary and forwarding to the product specialists only | Enquiry text (no system data) | Human answers in full |
The last row is the most important. Whether a product is suitable for a specific procedure, patient group or method is determined by its intended purpose. The manufacturer defines it, and it is for this purpose that the product has undergone conformity assessment. Anyone who answers such a question is giving advice. And under Section 83 MPDG (German Medical Devices Implementation Act), anyone who professionally provides specialist information on medical devices to healthcare professionals needs the necessary expertise and experience, explicitly including on the phone.
AI quotation in B2B stands or falls with ERP prices and stock levels
A language model that writes quotes “off the top of its head” is worthless to inside sales. It doesn’t know your customer terms, it doesn’t know the stock level, and it doesn’t know that the batch for the hospital expires next week. Everything that makes a quote right or wrong lives in the ERP.
Suppose a hospital requests a quote for three line items. For item one, the ERP supplies the price from the hospital group’s framework agreement. For item two, only 120 of 200 units are in stock; the rest arrives in two weeks. For item three, there is a batch that expires in five months. The draft shows every detail with its source, and partial delivery and expiry date appear as notes at the top.
The same applies to the enquiry that sat untouched during the holiday. If the group purchasing organisation’s terms are maintained in the ERP, the draft with the contract prices is ready and waiting, and the stand-in only has to check it.
Technically, the AI reads the enquiry, identifies items, quantities, delivery address and customer number and then queries the ERP: price for this customer, availability, delivery time, valid framework agreement. From the answers, it builds a readable draft again. The same basic rule applies as with any ERP integration, whether SAP, Microsoft Dynamics 365, Xentral or a custom-built system with an API: the ERP remains the master.
The connection between model and ERP can run via the Model Context Protocol, an open standard for AI models to access other systems. How an MCP server separates read and write tools determines how secure the setup is. The rule for quoting: open read access wide, keep write access tight. The AI may read prices; it does not create orders and does not change terms.
That leaves the question of who sends this draft.
Why a person always approves in medical technology
In other industries, you can argue about whether standard quotes may go out fully automatically. In medical technology, the law argues against it, and the reason lies in an ability language models are otherwise praised for: they write persuasively.
Ask a language model to make a quote line more persuasive, and it quickly turns it into something like this:
Master data: “Hydrocolloid dressing, sterile, 10 × 10 cm”
Draft: “The dressing reliably ensures faster wound healing.”
The addition promises success and attributes an effect to the product that nobody has approved.
This is exactly the kind of sentence the Medical Device Regulation targets. Art. 7 MDR prohibits misleading claims in labelling, instructions for use, making available on the market, putting into service and advertising. Among other things, it is misleading to ascribe functions and properties to a product that it does not have, or to recommend uses that are not part of the intended purpose covered by the conformity assessment. For in vitro diagnostics in laboratory supplies, Art. 7 IVDR provides the same.
Then there is the HWG (German Healthcare Advertising Act). Under Section 1(1) no. 1a, it also applies to medical devices and in vitro diagnostics, and Section 3 HWG prohibits misleading advertising, for example when the false impression is created that “success can be expected with certainty”. Whether an individual quote already counts as advertising within the meaning of the HWG is something to clarify with your legal advisers. This article does not replace legal advice.
A draft that your team checks and sends under its own name remains your quote. What the AI Act requires beyond that is covered in the article on the EU AI Act for SMEs.
So approval is what makes the process legally manageable. It can’t be optimised away, but it can be made faster: as in the hospital example, every line item carries its source. If the person has to look up every figure themselves, you have gained nothing. How to safeguard drafts like these with approval levels is described in the article on AI agents as digital employees. How fast the review then is, the pilot will measure.
The pilot: eight weeks, one enquiry class
A pilot that takes in all enquiry types at once won’t produce a clear result. That is why my proposal is deliberately narrow.
Picture a laboratory supplies retailer with five people in inside sales. For the pilot, it takes only the first row of the table, i.e. catalogue items with price and availability enquiries, and only existing customers with stored terms. Everything else carries on as before. In the first two weeks, the only thing that happens is counting: how many enquiries fall into this class, how long does a reply take today, how often does the customer follow up?
In weeks three to six, the AI runs in draft mode. Every draft lands in the responsible person’s inbox and is checked, corrected or discarded. Every correction is logged, because it shows where the mapping between enquiry and item master snags. Typical candidates: the manufacturer number instead of your item number, an old number from before the range changeover, packaging units where your system counts in pieces.
Separately, you count every suitability question that slipped through as a price enquiry and got a draft, like the catheter enquiry above. Here the threshold is zero. Every such case goes to the product specialists, and recognition is sharpened before the pilot continues.
In weeks seven and eight, you compare response time, correction rate and time per quote in inside sales. Decide in advance what a result means. My rule: if more than one in three drafts needs a correction in substance, check the item master first before tweaking the model.
For management, you work out the gain yourself: enquiries in this class per week times minutes per quote today, minus the minutes needed for review. That is the time your team gets back, and it comes from your own measurement.
There are two rules I would not negotiate for the pilot. The AI gets no write access to the ERP, not even for quote documents. Batches, expiry dates and framework agreements with group purchasing organisations are exactly the fields where an automatically created document triggers knock-on errors, and the draft in the inbox is enough for measuring. In addition, product texts come exclusively from approved templates; the model only writes the salutation and the framing sentences. That is the practical answer to the misleading-claims problem described above.
How a pilot like this fits together with the ERP, punchout catalogues and industry portals is shown on our page for medical technology and laboratory supplies. Before it starts, regulatory affairs and quality management need a seat at the table: they know the approved product texts and are the first to notice when a draft strays beyond the intended purpose.
What I recommend to sales managers in medical technology
You don’t have to replace the position you can’t fill. You have to make the work it would have done smaller.
AI quotation takes typing, looking things up and standard wording off inside sales’ hands; advice and responsibility for the content of the quote stay with the team. Because a wrong product sentence can fall under the MDR’s ban on misleading claims and under the HWG, this division of labour is the prerequisite for the project to hold up.
On Monday, go through one week of your inbox and count two things. First: how many enquiries fit the first row of the table and come from existing customers with stored terms? That is your most boring enquiry class, and it is where you start. Second: how many contain a suitability question that looks like a price enquiry? This number tells you how precisely your AI has to be able to tell the two apart.