An AI sales assistant gives all your sales advisers and brokers the same answer if its knowledge base holds a single description of each offer. You settle this before roll-out, with three decisions: one description per offer, the terms to keep, and the document of record. This article is for marketing and sales leaders at insurers and financial institutions that sell through a network and are about to load their documents into an assistant.
Take an insurer that sells its policies through a network of brokers and has not made these decisions. It gives its sales advisers an AI assistant and asks each department for its documents on a home insurance policy. Marketing supplies the brochure for private customers, the sales department its sales pitch, and the partner team the sheet sent to brokers. A sales adviser asks the assistant who the policy is for. Depending on the passages it retrieves, the answer draws on the brochure, the pitch or the broker sheet.
The three decisions to make before filling the knowledge base
These decisions rest with senior management, marketing and the sales department. They do not rest with the IT team that installs the tool, or to the assistant.
- One description per offer. It says who the offer is for, what it changes for that client and what sets it apart from the neighbouring offer. Our article on client transformation gives three questions for writing this description.
- The terms to keep. For each concept, one word, and the list of the words it replaces. The franchise network described below, for example, had to choose between “dealer” and “franchisee”.
- The document of record. For each offer, a single reference document goes into the knowledge base. The others stay out, or cite it.
The three decisions applied to the home insurance policy in our example
Before loading its documents, the insurer in our example brings together senior management and the three departments that supplied the documents: marketing, sales and the partner team. Together, they make the three decisions, in order.
- The description. The three departments agree on the clients the policy is for, for example: “owners and tenants of their main residence”. The brochure, the pitch and the broker sheet use this wording exactly as written.
- The terms. Suppose the brochure says “your contact”, the pitch says “intermediary” and the sheet says “partner”. The three departments choose “broker” and note the three words it replaces.
- The document of record. A one-page offer sheet, approved by the three departments, is the only document that goes into the knowledge base. The brochure, the pitch and the broker sheet remain sales materials and stay out of it.
When a sales adviser asks the assistant who the policy is for, they now get the same answer, however they word the question: the one in the offer sheet.
Our article on creating a custom GPT then details what the assistant’s instructions and knowledge base should contain.
Try it
Before filling the knowledge base, put the same client question to a sales adviser, someone from customer service and someone from marketing. If the three answers differ, your assistant’s knowledge base will receive three descriptions of the same offer. To check the description itself, paste your offer page into the Clear Offer Framework GPT, free to use with a ChatGPT account.
Why the assistant repeats every version of the offer
An assistant fed with company documents works in two steps. For each question, it searches its knowledge base for the passages closest to the question, then writes its answer from those passages. This is called retrieval-augmented generation, or RAG.
The assistant does not check whether these passages contradict each other. If the brochure says the policy is for owners and the broker sheet also mentions tenants, the answer depends on the passage retrieved. Asked twice, in two different wordings, the same question can therefore get two answers.
Removing old versions does not solve the problem: the three documents in our example are up to date, and each department approved its own. None of them is the document of record, because nobody has decided what the policy promises, and to whom.
A flaw that predates the assistant
The assistant exposes a gap that already existed between teams. In insurance, the Belgian Financial Services and Markets Authority (FSMA) has observed a practice among the companies that design policies, and this practice makes the gap likely. On 4 July 2023, the FSMA published a guide on insurance product oversight (in French), based on its inspections. It found that the vast majority of these companies define their target market in general terms. They do not set out the needs, characteristics and objectives of the clients they have actually chosen to target. The same document notes that several of these companies take little or no account of their distribution model, that is, the intermediaries who will sell the policy.
When the target market stays general, each team narrows it down in its own way. Marketing writes for the private customer, the sales department for its sales advisers, the partner team for the broker. The question “who is this policy for?” then gets three answers, before any assistant brings them together.
A franchise network, 2022–2023
We came across this flaw in a network of about twenty franchised companies in building renovation. The network did not use an AI assistant.
At the end of 2022, we analysed the website the brand used to recruit future franchisees. Two pages answered the same question: why choose this brand. The website used two words for the same status: “dealer”, the official term, and “franchisee”, the word candidates search for on Google. Some pages addressed the reader directly, others in the third person. The promise came down to adjectives the website did not explain. We recommended defining the brand’s promise and selling points first, then giving each page a single message.
In 2023, the network took on new franchisees. For each one, we wrote four pages built on the same template: the same structure, the same four guarantees, the same vocabulary and the same product names. Only the wording changed from one website to the next, to avoid duplicate content. With this shared template, every website presented the offer in the same way.
Two of the three decisions appear in these pages. For the description, each website highlighted the same four guarantees; for the terms, each used the same vocabulary and the same product names.
Like these websites, an assistant’s knowledge base receives the description it is given. If nobody has chosen between the versions in circulation, the knowledge base receives them all.
Frequently asked questions about AI sales assistants
Can an AI assistant make a sales message consistent?
No. An assistant writes its answers from the documents in its knowledge base. If these documents describe the same offer in several ways, the assistant uses one version or another, depending on the question. The message becomes consistent once the company has decided how to describe its offer.
What should go into a sales assistant’s knowledge base?
For each offer, one reference document that says who the offer is for, what it changes for the client and what sets it apart from neighbouring offers, with the terms to use. Competing descriptions of the same offer stay out of the knowledge base.
How can we tell whether our teams describe the offer in the same way?
Put the same client question to three people from three departments, without letting them confer, then compare their answers. If the offer proposed or the reason given changes from one answer to the next, the description of the offer has not been decided.
In our Clarity engagement, we decide with you how to describe each offer, before it goes into your sales materials and your tools.