AI Assistant in Business Software: What It Does
By now, "AI" appears on the product page of almost every business software package. What sits behind that word varies enormously: from a chat window that knows nothing about your company, to an assistant that pulls up your own figures and respects the permissions of the logged-in user. This article explains what an AI assistant in business software technically is, what it delivers, and which seven questions separate the marketing from reality.
The definition: a language model with access to your data, through defined tools
An AI assistant in business software is a language model that converts plain-language questions into queries against your own records, and returns the answer in plain language.
The crucial nuance sits in that middle part. The model gets no access to your database. It gets a list of predefined functions, called "tools", that it may call, each with fixed parameters. Ask "how many hours did we log on the Noordwijk project last month?" and the model picks the hours_summary tool with a start date, an end date and a search term. That tool then runs an ordinary database query written by the vendor, with exactly the same permission checks as if you had opened the screen yourself. The model only sees the result.
That distinction determines nearly everything that follows: how safe it is, how reliable the figures are, and what goes wrong when things go wrong.
Three levels that are all called "AI"
Level 1: a text generator with no access to your data. A chat window, or a "write this quote paragraph" button. Useful, but it knows nothing about your projects. You paste in the context yourself. This is by far the most common, and it's also what most established packages mean when they put "AI" on their website.
Level 2: an assistant that may read your data. The model has read access through tools and can therefore answer factual questions about your records. This is where the practical gain sits: you no longer need to know which screen a number lives on.
Level 3: an assistant that also makes changes. Logging hours, creating invoices, changing statuses. It sounds like the logical next step, but the risk profile changes completely: a misinterpreted question no longer produces a wrong answer, it produces wrong data. An assistant that finalises an invoice on its own does so in a gapless number sequence, and there is no undo for that.
Horixa deliberately sits at level 2: the AI assistant is fully read-only and never makes changes. It can draft an email or a call preparation from your CRM data, but that draft sits in its answer. Saving and sending is up to you.
What it actually delivers
The gain is rarely in spectacular things. It's in the questions you don't ask today because looking them up takes three minutes.
Questions nobody has to look up for you anymore. "Which invoices have been open for more than 30 days?" "How much unbilled work is sitting there right now?" "Who didn't log any hours last week?" In practice these are questions a project lead asks the office manager, who then runs a little report. That trip, two people and half a day of waiting, becomes one question.
Signals you'd otherwise spot too late. A weekly digest that walks through budget overruns, missing hours and stalled deals every Monday catches exactly the things that only hurt once they've been sitting for a month. Work in progress creeping up is the classic example: everyone sees it too late, because nobody looks at it weekly.
A lower barrier for people who rarely use the system. A partner or director who logs in twice a month doesn't know where the reports live. Asking a question, they can do.
What it does not deliver: better time tracking. An assistant retrieves data that exists. If the hours aren't in there, the assistant doesn't know them either. At best it can tell you who's missing. For increasing billable utilisation, entry discipline still matters more than any tool.
Where the limits are
Hallucination hasn't gone away, but it is manageable. A language model left to free-associate will invent plausible figures. The answer is that the assistant must never calculate from memory: every number comes from a tool. During a demo, ask for a figure you can verify yourself, then ask for the same thing phrased differently. If you get two different amounts, you know enough.
Permissions aren't a detail, they're the whole story. If the assistant runs with admin rights under the hood, any employee can reach revenue figures or colleagues' salaries through a cleverly phrased question. Good implementations set permissions per tool. In Horixa there are three levels: your own hours; financial data (amounts, invoices, pipeline); and admin (team-wide overviews such as work in progress and missing hours). An employee without admin rights isn't even offered the admin tools. They aren't in their list, so there's nothing to work around.
Your data goes to an external model vendor. This is the question most often glossed over. Almost no vendor runs its own model; the fragments needed for the answer go to a model provider acting as a sub-processor. That's perfectly fine, provided it's in the data processing agreement, provided they don't train on it, and provided you know which party it is. For Horixa that's OpenAI, it's stated in the privacy statement (Dutch), and every request is logged.
Bring your own AI: the MCP standard
Since 2024 there has been the Model Context Protocol (MCP), an open standard that lets any AI client connect to a data source. The practical consequence: you're no longer dependent on whichever assistant your software vendor happened to build in.
Horixa has a built-in MCP server with eight read-only tools. You connect Claude, Cursor or any other MCP client using an API key or via OAuth, and ask your questions there, limited to the scopes of that key. For agencies already working with an AI assistant, that's often more comfortable than a second chat window: "how is project X doing" then lands in the same conversation as the rest of your work. Details are on the developer page.
A vendor that offers an MCP server also gives you a way out. If the built-in assistant disappoints, you connect a different one.
The 7 questions to ask a vendor
- Does the assistant have access to my own data, or does it only generate text? (Level 1, 2 or 3.)
- Does every number come from a query, or is the model allowed to calculate?
- Do my user permissions apply inside the assistant too, and per component?
- Can the assistant make changes, and can I switch that off?
- Which model vendor is the sub-processor, and do they train on my data?
- Is every request logged, and can I inspect that log?
- Is it included in my plan or is it an add-on? Do the maths: an assistant at €10 per user per month costs €2,400 a year across 20 users.
Frequently asked questions
What exactly is an AI assistant in business software? A language model that converts plain-language questions into queries against your own records. The model gets no database access, but a set of predefined tools that each run a controlled query with the same permissions as the logged-in user.
Can an AI assistant reach data belonging to the vendor's other clients? In a sound implementation, no: the tools run within the user's own environment (tenant), enforced in the database itself. Ask explicitly how that separation is arranged technically. "That can't happen" without an explanation is not an answer.
Is my company data used to train AI models? That depends on your vendor's contract with its model provider. In business agreements, training on customer data is excluded as standard, but it is by no means always written into the data processing agreement. Get it confirmed in writing.
Can I connect my own AI instead of the vendor's? If the vendor offers an MCP server, yes. You connect your own AI client with an API key, carrying the permissions held by that key.
Is an AI assistant worth it for a small agency? From roughly ten employees onwards it pays for itself on the questions that otherwise all land with one person. Below that it's mainly pleasant rather than necessary. The time saved scales with the number of people who want to know something without going looking. See also time tracking software for agencies.
Want to see what a read-only assistant does with your own figures? Try Horixa free for 14 days, AI assistant included, no credit card.