Ten Kinds of AI Apps. Eight Still Stop at Answering and Suggesting.
Most AI applications I’ve seen in pharma stop at answering questions and making suggestions. The kind where the system proposes the default decision doesn’t exist yet at most companies.
I made a list of the AI applications I’ve seen in pharma commercial over the past few years. It has ten kinds. Eight stop at answering questions or making suggestions, one carries out a small task for a person, and the kind where the system proposes the default decision doesn’t exist yet at most companies. I’d guess the list looks much the same in other industries.
The list
The list comes from projects I’ve seen; I won’t name the companies. “Role” means how much decision-making power the AI has. A Chatbot only answers questions, a Copilot makes suggestions, an Agent carries out a task it’s been given, and a Harness keeps making decisions within set boundaries. I go into the differences in Models Are Bought. Harnesses Are Built.
| Application | What it does | Role |
|---|---|---|
| Medical and product knowledge assistant | Reps and medical affairs staff look up product information, literature and approved messaging | Chatbot |
| Role-play visits | The AI plays a doctor so reps can practice their pitch and objection handling | Chatbot |
| Scientific Q&A bot for doctors | Answers doctors’ scientific questions about a product | Chatbot |
| Interactive cases | Educational content for doctors that works through a patient case step by step | Chatbot |
| Public AI monitoring | Tracks how general-purpose AI models answer questions about the company’s products and disease areas | Copilot |
| Executive data assistant | Plain-language questions on top of BI, with the answers interpreted automatically | Copilot |
| Pre-call briefing | Summarizes a doctor’s situation and talking points before a visit | Copilot |
| HCP360 + NBA | Builds a profile of each doctor and recommends the next best action | Copilot |
| Call report generation | Turns a rep’s spoken notes after a visit into a structured record | Agent |
| Allocation and governance system | Makes ongoing decisions on territories, quotas and coverage plans | Harness |
On the map
Now put the ten into the four-quadrant map from Many AI Projects Die Before Anyone Picks a Model. The horizontal axis is how deep into the company’s own data an application needs to go; the vertical axis is how much judgment the AI takes on, from generating and retrieving up through recommending, deciding and executing.
The bottom-right is the most crowded, with the knowledge assistant, role-play visits, the scientific Q&A bot and interactive cases all there. That’s to be expected. They run on documents and content the company already has, the judgment is light, mistakes are cheap, and they get through compliance review fairly easily.
The second most crowded is the top-left, the trap zone, where the judgment is heavy and the data can’t keep up. NBA is the classic example. It aims to recommend or even decide, but at most companies the doctor-level data isn’t good enough, for the reasons given in that post. It’s meant to sit in the top-right and ends up in the top-left. The executive data assistant partly slides into this corner too: when metrics are defined differently across teams, nobody dares rely on its interpretations.
The top-right is almost empty.
Three constraints
Behind this are three main constraints.
The first comes from legal: retrieval is allowed, summarizing isn’t. Drug information is strictly regulated and one wrong sentence can become a compliance incident, so many companies let AI return only approved source text, with no summarizing or rewording. The rule has its reasons, but it locks a large number of applications into the most basic Chatbot tier. A more workable approach is to write the compliance rules into the system itself: what content can be summarized, what checks a summary has to pass, what gets logged. Once the rules are in the system, legal has grounds to loosen the restriction.
The second is that data collection falls entirely on the front line. The most important data, such as a doctor’s attitudes and concerns or what’s going on in a department, exists only in reps’ heads. The traditional fix is more forms, but every extra field makes the data a little worse. So AI projects that rely on the front line entering more data to feed the model rarely get far.
The third is that headquarters controls models and vendors. At multinationals, which model, which cloud and which vendor the China team may use is usually decided by global headquarters, and getting a new model approved can take six months. What’s left for the local team is configuring applications on platforms headquarters has already approved, and most of those only support the Chatbot and Copilot roles.
These constraints are especially strong in pharma, but they aren’t unique to it. Any company that is heavily regulated, relies on its front line to gather information and has its headquarters overseas will run into them.
What seems worth doing first
Bottom-right applications are worth continuing, and the ones that build up data along the way deserve priority. If role-play visits record which kinds of objections reps generally find hard to handle, they’re laying groundwork for the top-right. Pure Q&A leaves less behind.
The best value right now is in tools that lighten the front line’s data-entry load, such as turning a rep’s spoken notes after a visit into a structured record automatically. Reps do less work, the data gets better, and every application that depends on that data has something to stand on.
A next step could be to pick one decision that comes up often and can be undone when it’s wrong, such as the weekly visit plan, and build the full loop: the system proposes a default, a person can veto it, and the results flow back. The scope can be small, but the loop has to be complete. I can’t think of a shortcut from the top-left to the top-right that skips these steps.
There are also a few kinds of project where the ones I’ve seen haven’t done well: NBA without the data to support it; an “AI assistant” button added to an existing system that doesn’t correspond to any specific decision, whose usage falls off within a few months; and projects whose KPI is how many people have used them.
Why the top-right is empty
Getting into the top-right takes three things at once: a solid data foundation, accountability that makes clear who answers for mistakes, and business processes redesigned to match. Without any one of them, a project struggles to stay there. None of the three is only a technology problem, so buying something won’t solve it. That’s also why whoever gets there first may stay ahead for quite a while.
