# Where AI Should Actually Earn Its Keep in Biopharma - with Mukhtar Ahmed of Greenstone Biosciences

- **Show:** The AI in Business Podcast
- **Published:** 29 Sep 2026
- **Source:** https://podcast.emerj.com/where-ai-should-actually-earn-its-keep-in-biopharma-with-mukhtar-ahmed-of-greenstone-biosciences
- **Transcript:** https://podcasts.opennash.com/shows/ai-in-business/where-ai-should-actually-earn-its-keep-in-biopharma-with-mukhtar-ahmed-of-greens-82ce24/
- **Provider:** OpenNash Podcasts (machine-generated)

## In brief

Mukhtar Ahmed argues that biopharma has pointed AI at the wrong problem. Most money goes to generating molecules and targets, yet the industry has rarely run short of candidates. The costly mistakes come later, when companies decide which programs deserve more capital. He says AI should help with those portfolio decisions by pulling together scattered evidence, while people keep ownership of the final call.

The main tension is between speed and judgment. AI can compress discovery work, as in the TSP-1181 example, but a faster process does not make a program a better bet. Mukhtar says the benefits of decision-focused AI are hard to see because they are mostly things that did not happen: bad programs never funded and studies skipped. That is why he wants it measured across a whole portfolio over time.

## Key insights

1. **Faster discovery doesn't fix weak program selection** More candidates only help if a company can tell strong ones from weak ones. Otherwise AI just lets it chase weak hypotheses more efficiently.
2. **Biopharma is integration poor, not information poor** Biology, regulatory, competitive and IP knowledge sit in different teams, so decision makers get disconnected briefings. AI's value is building one coherent view of an asset across those teams.
3. **Accountability for go/no-go calls cannot be delegated** A recommendation nobody will defend to an investment committee or board is only an input. Mukhtar wants AI to augment decisions by exposing assumptions and contradictions, not to automate them.
4. **Define the decision before buying an AI platform** He says to start with decisions that repeatedly cause delay, disagreement or expensive reversals, and make the evidence criteria explicit. AstraZeneca's five R framework is his example of an explicit framework that technology can then support.
5. **Measure decision AI across the portfolio, not per asset** Its payoff is often counterfactual, such as a program not funded for another two years. He proposes measuring how much uncertainty is removed by committing the next dollar, rather than how much content the AI produces.

## Questions this episode answers

### Why does Mukhtar Ahmed say biopharma is using AI on the wrong problem?

He says most AI spending goes to generating molecules and targets, but the industry's real weakness is knowing early enough which programs deserve more money. Moving fast on the wrong drug candidate is the most expensive mistake.

### What does the TSP-1181 example show about AI in drug discovery?

Mukhtar says Exscientia and Sumitomo reportedly took the molecule from project start to a phase one trial in about twelve months, yet the program was later discontinued. For him, AI did what it was built to do by speeding discovery, but speed did not prove the asset was worth backing.

### What does it mean that companies are 'integration poor'?

Mukhtar means that the evidence needed for a go/no-go decision, such as biology, regulatory precedent, competition and patents, is held by different teams. Decision makers have to combine it in their heads, and he sees AI's value in joining it up.

### Should AI make the final decision on whether to continue a drug program?

No. Mukhtar gives three reasons: models rely on a biased published record, they are weak at predicting what regulators, payers and partners will accept, and someone must be accountable for the call. He favors 'decision augmentation'.

### How should a company decide where to spend its AI budget and judge the results?

Mukhtar advises starting with a specific, costly decision rather than a broad platform, and being clear about what good evidence looks like. He says to judge results over a portfolio and several investment cycles, by how much uncertainty is removed before committing the next dollar.

## Facts

- **Host** [0:00:12]: "Today's guest is Mukhtar Ahmed, director of project and portfolio strategy at Greenstone Biosciences."
- **Host** [0:00:50]: "the views and opinions expressed by Mukhtar on today's program are his own and do not reflect those of Greenstone Biosciences or its leadership."
- **Mukhtar Ahmed** [0:02:54]: "Most A I investments in biopharma is still concentrated at the molecule level, like you said, generative chemistry, structure prediction, virtual screening, target identification."
- **Mukhtar Ahmed** [0:02:54]: "You can show a molecule, you can show a docking score, or show that screening time fell by 50%."
- **Mukhtar Ahmed** [0:04:22]: "Exscientia and Sumitomo reportedly moved from molecule that molecule from project initiation to phase one trial in roughly twelve months."
- **Mukhtar Ahmed** [0:04:22]: "That was an extraordinary demonstration of what A I could do to compress part of the discovery process, but the program was subsequently discontinued."
- **Mukhtar Ahmed** [0:03:42]: "if we can generate more candidates and make more shots on goal, we'll improve our probability of success. But that only works if we are equally good at distinguishing the strong shots from the weak ones."
- **Mukhtar Ahmed** [0:07:35]: "Biology sit with discovery, regulatory precedent sits with another team, competitive intelligence may sit with strategy, intellectual property may sit with outside counsel."
- **Mukhtar Ahmed** [0:09:12]: "Organizations can end up allocating attention around according to volume, recency, and internal influence rather than evidentiary value."
- **Mukhtar Ahmed** [0:09:55]: "My own scientific background was in G P C R signal transduction, and there is a useful analogy there."
- **Mukhtar Ahmed** [0:09:55]: "A cell does not achieve specificity by receiving less information, it achieves specificity by controlling which signals are amplified and which are attenuated."
- **Mukhtar Ahmed** [0:10:54]: "Models are trained largely on the documented record, and that record contain publication bias and incomplete biology."
- **Mukhtar Ahmed** [0:11:36]: "A recommendation that nobody is prepared to defend in an investment committee or board meeting is not a decision, it's just an input."
- **Mukhtar Ahmed** [0:13:04]: "whether to go forward with that and spend the next ten to $20 million for early development."
- **Mukhtar Ahmed** [0:13:33]: "While A I helps, so far, I've seen majority of the calls are made by humans."
- **Mukhtar Ahmed** [0:15:54]: "AstraZeneca's five R means right patient, right target, right tissue, right safety, and right commercial potential."
- **Mukhtar Ahmed** [0:16:46]: "The strongest applications are usually focused, synthesizing a competitive landscape, mapping regulatory precedent, interpreting biomarker data, and or surfacing contradictions in the evidence."
- **Mukhtar Ahmed** [0:19:25]: "With decision layer A I, the benefit is often counterfactual. It is the program that did not get funded for another two years."
- **Mukhtar Ahmed** [0:20:07]: "the right K P I is not how much content the A I generates, it is how much uncertainty the organization removes by committing the next dollar."

## Transcript

**[0:00:12] Host:** Welcome everyone to the Emerj AI in Business podcast. Today's guest is Mukhtar Ahmed, director of project and portfolio strategy at Greenstone Biosciences. Mukhtar discusses why AI investment in biopharma R and D has concentrated on generating more molecules and candidates, even though the industry's most expensive mistakes happen further downstream in deciding which programs deserve continued capital. He argues that AI's greatest value lies in synthesizing fragmented evidence across discovery, regulatory and commercial teams, while keeping final judgment and accountability with human decision makers rather than the models.

**[0:00:50] Host:** A quick note for our audience that the views and opinions expressed by Mukhtar on today's program are his own and do not reflect those of Greenstone Biosciences or its leadership. Emerj works with a select group of AI vendors to reach Fortune 500 decision makers through research, media and direct access. If you want to be considered, download our media kit at Emerj dot com. Slash. A D one, that's e m e r j dot com. Slash a D one.

**[0:01:14] Host:** Now, the conversation with Mukhtar. Mukhtar, welcome to the Emerj AI in Business podcast studio.

**[0:01:41] Mukhtar Ahmed:** Thank you, pleasure to be here.

**[0:01:43] Host:** I want us to start this conversation somewhere that might be a bit uncomfortable, because I think it's probably the honest place to start. When we look at how AI is being framed in biopharma right now, it's almost as if every story is about generating more right, like more molecules, more candidates, more targets. And yet, the thing the industry has never really been short of is compounds. So, from your seat, when you look at where AI is being pointed inside R and D today, what stands out to you about where the spotlight is landing and and where it isn't?

**[0:02:19] Mukhtar Ahmed:** That's a great question. I'll start with saying that I think we are focusing A I on the wrong part of the problem. Drug development is not only a scientific problem, it is also a capital allocation problem conducted under extreme uncertainty. Biopharma does not primarily suffer from an inability to generate molecules, it suffers from an inability to know early enough which hypothesis deserves more capital. And that is where I think A I can create considerably more value.

**[0:02:54] Mukhtar Ahmed:** So, to your question, what stands out to me is how lopsided the investment has been. Most A I investments in biopharma is still concentrated at the molecule level, like you said, generative chemistry, structure prediction, virtual screening, target identification. That makes sense. These applications produce something tangible. You can show a molecule, you can show a docking score, or show that screening time fell by 50%. What is much harder to show is a bad program that was never started, or an asset that was killed two years earlier because the evidence simply wasn't strong enough.

**[0:03:42] Mukhtar Ahmed:** And yet economically, those may be some of the most valuable decisions A I can help us make. And I think, uh, there is an assumption underneath much of the current A I investment that if we can generate more candidates and make more shots on goal, we'll improve our probability of success. But that only works if we are equally good at distinguishing the strong shots from the weak ones. Otherwise, A I can make the discovery engine faster while simply allowing us to pursue weak hypotheses more efficiently.

**[0:04:22] Mukhtar Ahmed:** And in biopharma, the most expensive mistake isn't moving slowly, it's moving very quickly around the wrong asset. I think an example, an interesting example is the T S P eleven eighty one that Exscientia and Sumitomo reportedly moved from molecule that molecule from project initiation to phase one trial in roughly twelve months. That was an extraordinary demonstration of what A I could do to compress part of the discovery process, but the program was subsequently discontinued. And I don't think the lesson is that A I failed, it did what it was designed to do.

**[0:05:05] Mukhtar Ahmed:** It accelerated part of the discovery. The larger question is whether we are applying the same sophistication to deciding which programs deserve another $10 million, another two years and another piece of organizational attention. And you see, one week portfolio decision doesn't just consume capital, it can crowd out three better opportunities that never receive the same resource. So, when I look at A I in biopharma today, I see tremendous sophistication being applied to generating and optimizing candidates.

**[0:05:47] Mukhtar Ahmed:** The next frontier is applying that sophistication to judgment, which hypotheses deserve more programs, which programs deserve more capital, and which ones should have the discipline. We have the discipline to stop.

**[0:06:05] Host:** It sounds like the bigger problem is, if we had to put it in the plainest of terms, that we're pointing A I at the wrong bottlenecks. And I mean, you've been in this space for, I think, more years than what you would want our audience members to know. We don't want to give away your age, but is this pattern of pointing A I at the wrong bottleneck new, or is it just the latest version of something that you've seen before?

**[0:06:30] Mukhtar Ahmed:** I think these sorts of things usually tend to follow a similar trend. It could be some other somewhere else where we would see people get run with the hype of something new and they want to optimize based on that. For something that is that could be used to streamline their activities, but I don't think what we are seeing today is any different.

**[0:06:52] Host:** So I want to push this conversation or the next piece of the conversation into a direction that I think the audience really needs to hear you on. Because you are quite specific on this in a call that we had um a while ago. So if if molecule generation is where the industry has been pointing A I, and portfolio decision making is where it's actually been lagging for decades, then the question I want to sit inside for a while is what has to be true for A I to move that decision quality?

**[0:07:21] Host:** What's the difference between an R and D organization where A I improves the quality of the go, no go, the pick and progress and all of that, and one where it just adds noise on top of the noise that was already there?

**[0:07:35] Mukhtar Ahmed:** That's a very insightful question. A meaningful go no go decision is almost never based on one data set. It draws on the biological rationale, translational evidence, clinical feasibility, safety, competitive positioning, intellectual property, regulatory precedent, and commercial potential. The difficulty is that those inputs usually sit in different parts of the organization. Biology sit with discovery, regulatory precedent sits with another team, competitive intelligence may sit with strategy, intellectual property may sit with outside counsel. The decision maker then receives several disconnected briefings and is expected to integrate them mentally.

**[0:08:26] Mukhtar Ahmed:** That is where A I can create real leverage. Its value is not simply finding more information. It is constructing a coherent view of the asset across domains. And not naturally speaking, that is not naturally speaking to one another. Most biopharma companies are not information poor, they are integration poor, but synthesis alone is not enough. The system also has to distinguish between what is connected and what is actually decision relevant. And portfolio committee are exposed to a great deal of information that is allowed, but not necessarily important.

**[0:09:12] Mukhtar Ahmed:** It may be a competitor's press release, enthusiasm from an internal champion, a striking result from one underpowered experiment, or simply the weight of money already invested. Organizations can end up allocating attention around according to volume, recency, and internal influence rather than evidentiary value. That is the harder part of the problem. And A I can help identify relationships across large bodies of information, but relevant relevance is contextual. So let's frame the question this way. Does this finding matter at the current stage of development?

**[0:09:55] Mukhtar Ahmed:** Does it materially change the probability of success? Does it alter the capital at risk? Does it change what a regulator, physician, or payer, or potential partner is likely to accept? That weighting still requires judgment. My own scientific background was in G P C R signal transduction, and there is a useful analogy there. A cell does not achieve specificity by receiving less information, it achieves specificity by controlling which signals are amplified and which are attenuated. Portfolio decision making has the same challenge.

**[0:10:36] Mukhtar Ahmed:** The answer is not necessarily more data, the it is better selection machinery around the data.

**[0:10:43] Host:** That makes a lot of sense. And I am wondering, hearing you say all of this, should A I ever make the go or no go decision, or does there always need to be human judgment in the loop?

**[0:10:54] Mukhtar Ahmed:** Human judgment remains important for many reasons. I can point to three. First is mechanistic plausibility. Models are trained largely on the documented record, and that record contain publication bias and incomplete biology. It rarely captures all the tacit reasons experienced scientists stop believing in a mechanism. Second is institutional behavior. AI can help, inter precedent, but it is much weaker at anticipating what a regulator will accept, what a payer will reimburse, or what a partner may reject during diligence.

**[0:11:36] Mukhtar Ahmed:** Frankly, the third one is accountability. Someone still has to own the call. A recommendation that nobody is prepared to defend in an investment committee or board meeting is not a decision, it's just an input. So for me, the right model is not automated decision making, it is rather decision augmentation. And A I should help leaders see the entire asset more clearly, identify contradictions earlier, and understand which uncertainties matter most. But the judgment and the accountability remains human.

**[0:12:13] Mukhtar Ahmed:** So, in other words, I would just say that A I should make the evidence and trade-offs much harder to ignore. It should challenge inconsistency and expose assumptions, but the accountability cannot be delegated.

**[0:12:29] Host:** I like that. That's the way that you phrased it, saying that the accountability cannot be delegated. That even if you have all these tools, and I'm gonna ask a question that I think I already know exactly what you're going to answer to it, but I think it's there's reason for us to put a bit of emphasis on this for our audience. Obviously, you have lived experience and you've seen this in the real world. When you've seen a portfolio decision improved by A I in the real world, whose fingerprints were on it at the moment of making that call?

**[0:12:59] Host:** Was it the A I model, or was it the team that was working with it, or was it the combination of the two?

**[0:13:04] Mukhtar Ahmed:** I would say it was the combination of two, but it's the human decision that follows that is most important. You can ask a lot of questions to a I, design the molecule, try and figure out if it is gonna be safe when you put it in humans. But then there are so many other information that you have to still integrate and then make a decision whether to go forward with that and spend the next ten to $20 million for early development.

**[0:13:33] Mukhtar Ahmed:** And then it is still unknown at some point whether your investors are gonna buy in for you to raise more capital so you can take it to clinical trials. While A I helps, so far, I've seen majority of the calls are made by humans.

**[0:13:50] Host:** I expected you to to answer that based on what the conversation has been so far, and it makes a lot of sense. And I think we we basic. We touched on this in an earlier call as well, where we said, you know, basically, we let's use A I to to make us superhumans and help us make more informed decisions, but the decision stays with the human. You also hit on the investment topic, and I think this is what I want to close the conversation with, is thinking about the investment in this space.

**[0:14:15] Host:** So let's bring it down to to the person actually listening to this, whether that is a senior R and D leader or a V P of research or a chief medical officer. They obviously don't have an infinite A I budget. It's it's capped, they've got a certain amount of money available, and everybody in the organization has some kind of use case that they feel is more important than anyone else's. So everybody's demo looks impressive. And what I think I'd like to to hear from you is not necessarily giving our listeners a framework, but I'd like for you to counsel them in a way, you know, one on one to one, um, on where to actually put the money.

**[0:14:53] Host:** So that a year from now, they can point at a decision that was made in a better way, not just a workflow that got faster. Cause I think we get trapped in that a lot. Like, oh, let's just let's automate and make things faster, but that is not necessarily going to help us make better decisions. So where do you think, how? or how should they be thinking about where to focus A I investment from here?

**[0:15:15] Mukhtar Ahmed:** I'm glad you bring up this question. I would point investment up a layer from generation to decision. The first wave of A I in biopharma focused on generating targets, molecules, predictions, and experimental options. The next wave should focus on helping organizations decide which of those options deserve resources that we have been discussing so far. But the sequence matters. I wouldn't start by buying a broad A I platform and then searching for a problem to attach it to.

**[0:15:54] Mukhtar Ahmed:** I would start by defining the decision. What decision repeatedly consumed time, creates disagreement, or produce expensive reversals? I think is the kind of question you want to ask. What evidence should determine that decision, and which criteria are explicit versus existing only in the heads of a few experienced people? Only after answering those questions should the organization decide where A I belongs. For instance, I I find AstraZeneca's five R framework as a useful example. By the way, AstraZeneca's five R means right patient, right target, right tissue, right safety, and right commercial potential.

**[0:16:46] Mukhtar Ahmed:** The important lesson isn't that everyone, every organization, should copy those five criteria. It's that technology creates far more value when it's applied to an explicit decision framework. Otherwise, you risk automating inconsistency. So before you build an A I decision tool, you need to know what a good decision looks like. Technology cannot compensate for ambiguous criteria, right? The second principle is to stay fit for purpose. The strongest applications are usually focused, synthesizing a competitive landscape, mapping regulatory precedent, interpreting biomarker data, and or surfacing contradictions in the evidence.

**[0:17:41] Mukhtar Ahmed:** In my own portfolio work, A I has been most useful when it brings fragmented information together, challenging assumptions and flagging risks earlier. It does not replace the decision process that we just discussed a little bit earlier. It helps the people from making the decision, ask better questions and have a better informed conversation. So to cut to the chase, I would say, start with the decision, not the platform.

**[0:18:15] Host:** I I like that, and I think that it ties into the whole conversation. If I had to nutshell our conversation today, I would say that, you know, first of all, uh, the bottleneck in biopharma R and D has never been generating more molecules. It's it's been deciding which ones to take forward. And if your A I investment is pointed at generation and not at decision quality, then you're basically speeding up the part of the pipeline that has never even been a problem, so you won't see R O I there.

**[0:18:45] Host:** Uh, we can see A I improving R and D decisions when it can pull scattered signal, like you mentioned in that second question, looking at more than just one thing and getting the human judgment involved there and not leaving it all to A I. But then ultimately, like you just said, if you really want to see R O I here, it's not going to be. What are we short of in terms of molecules? It is what are we?

**[0:19:09] Host:** what are we short on in terms of decisions? And where can we deploy A I faster is not necessarily the question. It is at which decision point in our portfolio does better information change what we choose to progress? Is that the message that we want to send today?

**[0:19:25] Mukhtar Ahmed:** Yeah, I mean, I think leaders need to measure A I differently. Decision layer A I will rarely justify itself through one dramatic win on a single asset. Its value appear cumulatively across the portfolio and over several investment cycles. With decision layer A I, the benefit is often counterfactual. It is the program that did not get funded for another two years. The study you avoided because it would not change the decision or the capital you redirected before some cost logic took hold.

**[0:20:07] Mukhtar Ahmed:** So leaders should not judge these tools on one quarter, one program, or a simple adoption metric. Decision layer A I should be measured on a portfolio horizon, not an asset horizon. I think the way I would put it is that the right K P I is not how much content the A I generates, it is how much uncertainty the organization removes by committing the next dollar.

**[0:20:35] Host:** Yeah, I think that is something that it sounds very quotable, and it sounds like something that we should put on a banner. Um, it's been a great conversation. Thank you so much for giving us so much of your time today, and I hope that we can continue this conversation in another episode in the future.

**[0:20:50] Mukhtar Ahmed:** Thank you.

**[0:21:05] Host:** Wrapping up today's episode, I think there are three key takeaways from our conversation with Mukhtar. First, the biggest bottleneck in biopharma R and D has never been generating molecules, it's deciding which ones deserve continued capital and attention. Second, A I creates the most value not by producing more candidates, but by synthesizing fragmented evidence across teams so decision makers can see the full picture of an asset. And finally, human judgment and accountability cannot be delegated to A I.

**[0:21:34] Host:** The strongest organizations use A I to sharpen decisions, not replace the people who own them. Emerj works with a select group of A I vendors to reach Fortune 500 decision makers through research, media, and direct access. If you want to be considered, download our media kit at Emerj dot com. Slash a D one, that's e m e r j dot com. Slash a D one. On behalf of the team at Emerj, we'll see you on the next episode.