Tokens Are the Cheap Part — Steve Jamison, Jellyfish

Hi, I'm Steve Hi, I'm Steve Jameson, Head of Jameson, Head of AI Research and Development at Jellyfish . I'm here to show . I'm here to show . I'm here to show you why tokens are the you why tokens are the you why tokens are the smallest part of the smallest part of the smallest part of the cost. The research cost. The research cost. The research I will present is I will present is I will present is based on data from the based on data from the based on data from the Jellyfish platform. This is the Jellyfish platform. This is the Jellyfish platform. This is the industry's largest industry's largest engineering engineering workflow data set combined with workflow data set combined with workflow data set combined with AI telemetry. It's no AI telemetry. It's no AI telemetry. It's no secret that secret that secret that AI spending is rising. They have AI spending is rising. They have AI spending is rising. They have increased 60 times increased 60 times increased 60 times compared to the same compared to the same compared to the same time last year. time last year. time last year. The areas of use are The areas of use are The areas of use are wide and wide and wide and varied. Models varied. Models varied. Models and tools and tools and tools are getting much are getting much are getting much better. The consequences of better. The consequences of better. The consequences of this are fascinating. But this are fascinating. But this are fascinating. But discussing budgets discussing budgets discussing budgets and expenses isn't that much and expenses isn't that much and expenses isn't that much fun. These conversations fun. These conversations fun. These conversations involve your involve your involve your colleagues, senior colleagues, senior colleagues, senior management, and the board of management, and the board of management, and the board of directors. This isn't directors. This isn't directors. This isn't always the easiest always the easiest always the easiest audience to target, especially audience to target, especially audience to target, especially when you don't have the when you don't have the when you don't have the right data. And right data. And right data. And today it is today it is today it is real money. real money. Productivity Productivity is also increasing. The number of is also increasing. The number of is also increasing. The number of consolidated consolidated consolidated pull requests (PRs) per pull requests (PRs) per pull requests (PRs) per engineer increased by 43% engineer increased by 43% over the same annual over the same annual over the same annual period. This is period. This is period. This is phenomenal. We have not phenomenal. We have not phenomenal. We have not seen such seen such performance indicators, performance indicators, especially in such a especially in such a especially in such a short time frame. There is short time frame. There is short time frame. There is enough research enough research enough research and evidence and evidence and evidence to support that to support that to support that AI and AI and AI and AI-enabled coding are AI-enabled coding are AI-enabled coding are driving this driving this driving this growth in growth in growth in throughput. After all, throughput. After all, throughput. After all, even our graphs even our graphs even our graphs show this connection show this connection . Pay attention to where the . Pay attention to where the break points are.

break points are. AI spending increased in early AI spending increased in early AI spending increased in early 2026, when 2026, when 2026, when productivity gains began productivity gains began to increase steadily. However, there is a certain there is a certain tension between these two numbers. How to tension between these two numbers. How to tension between these two numbers. How to reconcile them? Costs reconcile them? Costs reconcile them? Costs have increased 60 times. have increased 60 times. have increased 60 times. Productivity also Productivity also Productivity also increased by 43%. Is this good? increased by 43%. Is this good? increased by 43%. Is this good? Badly? 43% is a big Badly? 43% is a big Badly? 43% is a big number. But 60 times— number. But 60 times— too. How do you know if too. How do you know if too. How do you know if you are you are you are spending enough? Are there any other spending enough? Are there any other productivity gains that productivity gains that can be gained for can be gained for can be gained for money? Is it worth money? Is it worth money? Is it worth increasing costs by increasing costs by increasing costs by 50%? Or twice as much? 50%? Or twice as much? 50%? Or twice as much? Or how do you know if Or how do you know if Or how do you know if you've spent you've spent you've spent too much? I'm sure too much? I'm sure many of you many of you many of you have felt this tension. The have felt this tension. The have felt this tension. The CFO CFO CFO or board of directors or board of directors or board of directors requires reporting requires reporting requires reporting on expenses. Can you on expenses. Can you on expenses. Can you show them this? Will show them this? Will show them this? Will they be satisfied that they be satisfied that they be satisfied that the number of PRs the number of PRs the number of PRs has increased by 43%? Besides has increased by 43%? Besides , what happens if , what happens if , what happens if you limit you limit you limit usage and usage and usage and spending? Will you stop spending? Will you stop spending? Will you stop burning tokens? Will burning tokens? Will burning tokens? Will you stop further you stop further efficiency gains? You efficiency gains? You can ask can ask can ask colleagues from other colleagues from other colleagues from other companies how much companies how much companies how much they pay. For they pay. For they pay. For many, it is still a many, it is still a many, it is still a moving target. But how does moving target. But how does moving target. But how does this compare to your this compare to your this compare to your organization? Of course, organization? Of course, organization? Of course, simply comparing simply comparing simply comparing their score with their score with their score with yours is not enough. We've yours is not enough. We've yours is not enough. We've heard reports heard reports heard reports of companies of companies of companies spending $100,000, $ spending $100,000, $ spending $100,000, $ 500,000, or even 500,000, or even 500,000, or even millions on AI. They millions on AI. They millions on AI. They set and then set and then set and then revise their revise their revise their AI budgets as AI budgets as reality sets in. Ultimately, reality sets in. Ultimately, these two numbers

these two numbers these two numbers cannot tell the cannot tell the cannot tell the whole story. And even whole story. And even whole story. And even together they don't together they don't together they don't take into account the largest take into account the largest take into account the largest expense item. These two expense item. These two expense item. These two figures do not take into account the figures do not take into account the figures do not take into account the main expense— main expense— developer salaries developer salaries . AI is making . AI is making . AI is making developers developers developers more productive on a more productive on a more productive on a scale we have scale we have scale we have n't seen before. This is n't seen before. This is n't seen before. This is not just another not just another not just another tool, it is tool, it is tool, it is intelligence. This intelligence. This intelligence. This intelligence dramatically intelligence dramatically intelligence dramatically changes changes changes the productivity of the productivity of the productivity of every developer. every developer. every developer. Therefore, we need to Therefore, we need to Therefore, we need to change the approach to change the approach to change the approach to assessing its value. assessing its value. assessing its value. Not as an Not as an Not as an overhead, but as part of the overhead, but as part of the overhead, but as part of the cost of creating cost of creating cost of creating each PR. We will call each PR. We will call each PR. We will call this the this the this the intelligence cost per PR. intelligence cost per PR. Developer salaries Developer salaries plus AI costs— plus AI costs— combine these into one combine these into one combine these into one expense category. AI expense category. AI expense category. AI accounts for about 2.5% of accounts for about 2.5% of the cost of creating a the cost of creating a the cost of creating a unit of merged code. unit of merged code. unit of merged code. Salary is the remaining 97.5 Salary is the remaining 97.5 %. This is consistent with other %. This is consistent with other %. This is consistent with other reports, the data is reports, the data is reports, the data is current as of August current as of August current as of August 2026. Here, tokens 2026. Here, tokens are the smallest part of the are the smallest part of the are the smallest part of the costs. But these are costs. But these are costs. But these are current relative current relative current relative costs, not dollars.

costs, not dollars. costs, not dollars. Let's take a look at what happened Let's take a look at what happened Let's take a look at what happened to the value of merged PR to the value of merged PR to the value of merged PR over the last year. over the last year. The dollar value of The dollar value of merged PR has declined merged PR has declined merged PR has declined 26% over the past year. This refers 26% over the past year. This refers 26% over the past year. This refers to to to PRs that are not created, but rather PRs that are not created, but rather PRs that are not created, but rather merged. The time for merged. The time for merged. The time for verification was taken into account, verification was taken into account, verification was taken into account, and a certain level of and a certain level of and a certain level of quality was maintained. This quality was maintained. This quality was maintained. This decrease in PR costs decrease in PR costs decrease in PR costs includes a 60-fold includes a 60-fold includes a 60-fold increase in increase in increase in AI spending over the same AI spending over the same AI spending over the same period. This change is period. This change is period. This change is almost universal. 83% of the almost universal. 83% of the companies we studied companies we studied saw a saw a saw a decrease in PR costs. decrease in PR costs. decrease in PR costs. What does this What does this What does this tell us? Increasing tell us? Increasing tell us? Increasing bandwidth bandwidth bandwidth allows salaries to be distributed allows salaries to be distributed allows salaries to be distributed to more PRs, to more PRs, to more PRs, and this happens and this happens and this happens faster than faster than faster than AI costs increase. What is AI costs increase. What is AI costs increase. What is the range of the range of the range of token costs? We ranked token costs? We ranked token costs? We ranked companies by companies by token usage and divided token usage and divided them into quintiles.

them into quintiles. The range here is quite The range here is quite wide. In the group with wide. In the group with wide. In the group with the lowest costs, the lowest costs, the lowest costs, AI is about AI is about AI is about 0.3% of the cost of PR. That's 0.3% of the cost of PR. That's 0.3% of the cost of PR. That's approximately 60 approximately 60 approximately 60 million tokens per million tokens per million tokens per developer per month. The developer per month. The top 20% in top 20% in AI spending per developer is the AI spending per developer is the AI spending per developer is the group with the highest group with the highest group with the highest spending. There, AI spending. There, AI spending. There, AI accounts for 8.5% of accounts for 8.5% of accounts for 8.5% of PR costs. That's approximately 2.5 PR costs. That's approximately 2.5 PR costs. That's approximately 2.5 billion tokens per billion tokens per billion tokens per developer per month. developer per month. developer per month. Approximately two Approximately two Approximately two orders of magnitude more than in orders of magnitude more than in orders of magnitude more than in the group with the lowest the group with the lowest the group with the lowest costs. We will costs. We will costs. We will use these use these use these groups in the following groups in the following groups in the following slides to see the slides to see the slides to see the impact of token spending impact of token spending impact of token spending on other metrics. on other metrics. First, First, First, PR throughput, measured PR throughput, measured PR throughput, measured as the number of merged PRs as the number of merged PRs as the number of merged PRs per developer. per developer. per developer. Distributed by the same Distributed by the same Distributed by the same quintiles of quintiles of quintiles of AI spending from the previous AI spending from the previous AI spending from the previous slide. The bottom 4/5 by slide. The bottom 4/5 by slide. The bottom 4/5 by cost received a cost received a cost received a throughput increase of throughput increase of throughput increase of about 40%.

about 40%. But the high- But the high- spending group performed spending group performed spending group performed better—a better—a 67% increase. 67% increase. 67% increase. Higher spending on AI Higher spending on AI Higher spending on AI has led to greater has led to greater has led to greater growth in growth in growth in throughput. Now throughput. Now throughput. Now let's consider the cost of let's consider the cost of let's consider the cost of one PR. Positive one PR. Positive one PR. Positive results results results are again observed in are again observed in are again observed in all groups. The cost of all groups. The cost of all groups. The cost of PR is decreasing for all PR is decreasing for all PR is decreasing for all groups of AI users. groups of AI users. However, again, the However, again, the most active most active most active fifth—that is, fifth—that is, fifth—that is, the companies that the companies that the companies that spent the most— spent the most— spent the most— achieved the best achieved the best achieved the best results. Over the results. Over the results. Over the last year, they last year, they last year, they have seen a 32% decrease in have seen a 32% decrease in have seen a 32% decrease in the cost of one PR the cost of one PR the cost of one PR . The other four groups . The other four groups . The other four groups saw a decrease in saw a decrease in saw a decrease in the cost of one PR the cost of one PR the cost of one PR by about 25%. by about 25%. Investments in Investments in increased increased increased AI spending contributed to greater AI spending contributed to greater AI spending contributed to greater improvements in the cost improvements in the cost improvements in the cost per PR compared to per PR compared to per PR compared to those who those who those who invested less. So we invested less. So we invested less. So we saw the same thing as saw the same thing as saw the same thing as in the in the in the averages: changes in averages: changes in averages: changes in the cost of one PR are the cost of one PR are the cost of one PR are largely largely largely driven by driven by driven by PR performance, PR performance, PR performance, not AI spending.

not AI spending. not AI spending. Large payers Large payers saw a 67% increase in PR productivity compared to saw a 67% increase in PR productivity compared to approximately 40% for their approximately 40% for their approximately 40% for their counterparts. Spreading the counterparts. Spreading the counterparts. Spreading the same salary over a larger same salary over a larger same salary over a larger number of PRs per month number of PRs per month number of PRs per month has a greater effect has a greater effect has a greater effect than increasing than increasing than increasing AI costs, since AI costs, since AI costs, since tokens are the cheap tokens are the cheap tokens are the cheap part. This does not part. This does not part. This does not mean that mean that mean that organizations should organizations should organizations should waste tokens and waste tokens and waste tokens and spend money spend money spend money thoughtlessly. Here we thoughtlessly. Here we thoughtlessly. Here we look at look at look at tokens used tokens used tokens used on the horizontal on the horizontal on the horizontal axis and the cost of axis and the cost of axis and the cost of one PR on the one PR on the one PR on the vertical axis. vertical axis. There is a There is a sharp drop-off at the beginning, where sharp drop-off at the beginning, where sharp drop-off at the beginning, where relatively small relatively small relatively small AI investments yield AI investments yield AI investments yield significant initial significant initial significant initial results. As results. As token usage increases, the token usage increases, the cost improvement cost improvement cost improvement per PR decreases. per PR decreases. per PR decreases. This is the law of diminishing This is the law of diminishing This is the law of diminishing returns. However, I want returns. However, I want returns. However, I want to emphasize: diminishing to emphasize: diminishing to emphasize: diminishing returns are not the returns are not the returns are not the same as no same as no same as no returns. We see that the returns. We see that the returns. We see that the top 10% top 10% spend 8x spend 8x spend 8x more tokens more tokens more tokens compared to the median compared to the median compared to the median and achieve a 25% and achieve a 25% lower cost lower cost lower cost per PR. There will come a time per PR. There will come a time per PR. There will come a time when tokens when tokens when tokens will no longer be the will no longer be the will no longer be the cheap part, but cheap part, but cheap part, but overall for overall for overall for most companies most companies most companies we are still far from we are still far from we are still far from that threshold. More that threshold. More that threshold. More tokens lead tokens lead tokens lead to more to more to more PR, which leads to lower PR, which leads to lower PR, which leads to lower shipping costs.

shipping costs. So far we have talked So far we have talked about companies. Next, I about companies. Next, I about companies. Next, I would like to digress would like to digress would like to digress and talk and talk and talk about the research about the research about the research we conducted with the we conducted with the we conducted with the developers. Here, we developers. Here, we developers. Here, we looked at looked at looked at developer weeks with developer weeks with developer weeks with heavy heavy token usage compared to the token usage compared to the same weeks with same weeks with same weeks with less less less token usage. Here we token usage. Here we token usage. Here we saw the same saw the same saw the same phenomenon, but on an phenomenon, but on an phenomenon, but on an individual individual individual level. Compare level. Compare level. Compare weeks of active weeks of active weeks of active use of use of use of agent skills agent skills agent skills by developers with by developers with by developers with weeks without them. They weeks without them. They weeks without them. They spend about spend about spend about twice as much on twice as much on twice as much on tokens, aggregate tokens, aggregate tokens, aggregate 27% more PR, and each PR 27% more PR, and each PR 27% more PR, and each PR costs about 18 costs about 18 % less when % less when % less when considering salary considering salary considering salary and tokens together. There are two and tokens together. There are two and tokens together. There are two valid questions that valid questions that valid questions that I am asked when I I am asked when I I am asked when I talk about this talk about this talk about this work. First, does work. First, does working faster lead to poorer working faster lead to poorer quality? And secondly, is the quality? And secondly, is the quality? And secondly, is the higher higher higher amount of PR we amount of PR we amount of PR we see due to lower see due to lower see due to lower volumes of work?

volumes of work? The first question: is The first question: is cheaper and faster cheaper and faster cheaper and faster work of lower work of lower work of lower quality? If AI quality? If AI quality? If AI makes PR faster and makes PR faster and makes PR faster and cheaper, doesn't that cheaper, doesn't that cheaper, doesn't that mean we're mean we're mean we're just making just making just making more mistakes? One more mistakes? One more mistakes? One indicator for indicator for indicator for this is the cancellation ( this is the cancellation ( revert) of requests. revert) of requests. revert) of requests. Merged PRs that Merged PRs that Merged PRs that cancel the cancel the cancel the previous PR. The previous PR. The previous PR. The heaviest fifth heaviest fifth heaviest fifth started from the started from the started from the lowest level even lowest level even lowest level even before the main before the main before the main increase in the number of increase in the number of increase in the number of tokens, and the tokens, and the tokens, and the cancellation rates are slightly cancellation rates are slightly cancellation rates are slightly decreasing for decreasing for decreasing for each group. Although these each group. Although these each group. Although these deltas only slightly deltas only slightly deltas only slightly exceed the exceed the exceed the statistical error statistical error . This is not evidence of . This is not evidence of . This is not evidence of an improvement in the situation. More an improvement in the situation. More importantly, however, importantly, however, we see no we see no we see no worrying signs of worrying signs of worrying signs of significant significant significant quality issues with higher quality issues with higher quality issues with higher token costs and token costs and token costs and faster PR processing. faster PR processing. The second question The second question I'm interested in is: is I'm interested in is: is I'm interested in is: is throughput driven by throughput driven by throughput driven by lower PRs? lower PRs?

And the answer to that is definitely no. No definitely no. No definitely no. No less PR. In fact, less PR. In fact, less PR. In fact, AI leads to an AI leads to an AI leads to an increase in the number of increase in the number of increase in the number of lines of code lines of code lines of code generated within a generated within a generated within a single PR. For single PR. For single PR. For the developers in the room, this comes as the developers in the room, this comes as the developers in the room, this comes as no surprise. no surprise. Raise your hand if Raise your hand if you've rolled your eyes at you've rolled your eyes at you've rolled your eyes at wordy, wordy, wordy, unnecessary unnecessary unnecessary AI comments or AI comments or AI comments or files filled with files filled with files filled with mostly mostly mostly irrelevant irrelevant irrelevant tests. tests. Generative AI Generative AI loves to generate loves to generate loves to generate code. Additionally, the code. Additionally, the code. Additionally, the entry threshold for entry threshold for entry threshold for developers is low, developers is low, developers is low, and there is a temptation and there is a temptation and there is a temptation to do more in to do more in to do more in one PR instead of one PR instead of one PR instead of opening a opening a opening a new ticket. It's just new ticket. It's just new ticket. It's just one, maybe two one, maybe two one, maybe two queries, and I'll deal queries, and I'll deal queries, and I'll deal with that edge with that edge with that edge case, or I can case, or I can case, or I can combine this and that combine this and that combine this and that . Less PR to . Less PR to work with. However, this work with. However, this doesn't really align doesn't really align doesn't really align with the stress people with the stress people with the stress people have been feeling during have been feeling during have been feeling during PR checks lately PR checks lately PR checks lately . There is also AI . There is also AI . There is also AI simply becoming simply becoming simply becoming more powerful, so we more powerful, so we more powerful, so we are becoming more ambitious. In are becoming more ambitious. In are becoming more ambitious. In this report, we will not delve into this report, we will not delve into what exactly is what exactly is driving this driving this driving this level of change in level of change in level of change in the number of lines of the number of lines of the number of lines of code. However, the growth in code. However, the growth in PR throughput is not associated with PR throughput is not associated with smaller PRs.

smaller PRs. smaller PRs. This is intuitive for most This is intuitive for most , but it's still worth , but it's still worth checking. checking. Let's go back to the Let's go back to the original question. original question. original question. Is 60% worth the extra Is 60% worth the extra Is 60% worth the extra cost for 43% more cost for 43% more cost for 43% more productivity? productivity? productivity? This cannot be judged separately This cannot be judged separately This cannot be judged separately . You should . You should . You should track both track both track both metrics. These are important metrics. These are important metrics. These are important numbers, but they are numbers, but they are numbers, but they are supporting supporting supporting metrics. AI spend per metrics. AI spend per metrics. AI spend per PR is the PR is the PR is the intelligence you intelligence you intelligence you purchase as a resource purchase as a resource purchase as a resource to help your to help your to help your developers developers developers work faster. work faster. work faster. The amount of PR per The amount of PR per The amount of PR per engineer is what you engineer is what you engineer is what you get for get for get for this cost. However, this cost. However, this cost. However, looking at this looking at this looking at this data alone hides data alone hides data alone hides a compromise. But a compromise. But a compromise. But add in add in add in the salary, take the the salary, take the cost-to- cost-to- productivity ratio, and you productivity ratio, and you productivity ratio, and you have the answer. have the answer. The cost of one PR The cost of one PR decreased by 26% over the decreased by 26% over the decreased by 26% over the same period of time. same period of time. same period of time. It is clear that AI It is clear that AI It is clear that AI contributes to increased contributes to increased contributes to increased cost efficiency, cost efficiency, cost efficiency, despite the increase in despite the increase in despite the increase in bills. And the bills. And the bills. And the AI ​​costs required to AI ​​costs required to AI ​​costs required to achieve this achieve this achieve this efficiency are efficiency are efficiency are relatively small. AI is still a relatively small. AI is still a tiny tiny tiny fraction of the cost of PR. On fraction of the cost of PR. On fraction of the cost of PR. On average, 2.5%. Less than average, 2.5%. Less than average, 2.5%. Less than 10% for the 10% for the 10% for the high-spending group.

high-spending group. high-spending group. Tokens are the Tokens are the Tokens are the cheapest part. In the cheapest part. In the cheapest part. In the range we range we range we see, higher see, higher see, higher spending continued to spending continued to spending continued to deliver results deliver results . More active . More active . More active token users token users token users saw a saw a 67% increase in PR throughput compared to 40% for 67% increase in PR throughput compared to 40% for their less-spending their less-spending their less-spending counterparts. This reduced counterparts. This reduced counterparts. This reduced the cost of one PR by the cost of one PR by the cost of one PR by 32%, compared to 32%, compared to 32%, compared to approximately 25% for approximately 25% for approximately 25% for companies with lower companies with lower companies with lower costs. costs. Analyzing the PR to token value curve , the top 10% of monthly , the top 10% of monthly token usage companies yielded 8-10 token usage companies yielded 8-10 times more tokens, times more tokens, times more tokens, but 25% lower but 25% lower but 25% lower PR value compared PR value compared PR value compared to the median. AI is making to the median. AI is making to the median. AI is making developers more developers more developers more productive on a productive on a productive on a scale we have scale we have scale we have n't seen before. This is not n't seen before. This is not n't seen before. This is not just another just another just another tool, it is tool, it is tool, it is intelligence that changes the intelligence that changes the intelligence that changes the scope of productivity of scope of productivity of scope of productivity of every developer. every developer. That's why we need to That's why we need to think about its think about its think about its cost differently: not cost differently: not cost differently: not as an as an as an overhead, but as a overhead, but as a overhead, but as a part of what it part of what it part of what it takes to takes to takes to create each PR.

create each PR. You need to add You need to add the costs of AI to the costs of AI to the costs of AI to the costs of the costs of the costs of specialist salaries to specialist salaries to specialist salaries to calculate the real calculate the real calculate the real cost of one PR. This does cost of one PR. This does cost of one PR. This does not mean that you should not mean that you should not mean that you should spend thoughtlessly. spend thoughtlessly. spend thoughtlessly. Of course, it is worth monitoring Of course, it is worth monitoring Of course, it is worth monitoring uncontrolled uncontrolled uncontrolled processes. You should processes. You should processes. You should track your track your track your AI spending. I'm not AI spending. I'm not AI spending. I'm not saying that you shouldn't saying that you shouldn't saying that you shouldn't look at how much look at how much look at how much AI money AI money AI money is spent on is spent on is spent on each PR. I'm saying you each PR. I'm saying you each PR. I'm saying you shouldn't make this your shouldn't make this your shouldn't make this your primary metric. primary metric. The new approach may The new approach may change your view of change your view of change your view of other investments, other investments, other investments, such as integrating such as integrating model routers, retraining your model routers, retraining your own models, own models, own models, implementing implementing implementing token limits, and so on. Will you token limits, and so on. Will you token limits, and so on. Will you gain more gain more gain more efficiency from efficiency from efficiency from controlling AI costs controlling AI costs controlling AI costs or from investing in or from investing in or from investing in things that increase things that increase things that increase throughput, throughput, throughput, such as such as such as expanding expanding expanding AI usage? In AI usage? In AI usage? In any case, any case, any case, track track track the total the total the total intelligence spend, human intelligence spend, human intelligence spend, human plus artificial, for plus artificial, for plus artificial, for each PR on a monthly basis. Next to the each PR on a monthly basis. Next to the each PR on a monthly basis. Next to the AI ​​score and AI ​​score and AI ​​score and next to the next to the next to the PR bandwidth. What is your PR bandwidth. What is your PR bandwidth. What is your dynamics? Is dynamics? Is productivity continuing to increase? What productivity continuing to increase? What about about about cost efficiency, cost of cost efficiency, cost of cost efficiency, cost of implementation? Keep an eye on implementation? Keep an eye on implementation? Keep an eye on bounces and bounces and bounces and quality metrics quality metrics quality metrics as a counterbalance to all of this as a counterbalance to all of this as a counterbalance to all of this . We don't . We don't . We don't need need need AI garbage that slows down AI garbage that slows down AI garbage that slows down work. So, work. So, work. So, on the eve of on the eve of on the eve of budget season, budget season, budget season, bring up the cost of PR

bring up the cost of PR bring up the cost of PR for discussion. Not for discussion. Not for discussion. Not just the just the just the token bill, not just the token bill, not just the productivity gains. productivity gains. The bill shows what The bill shows what The bill shows what you spent. you spent. you spent. The cost of PR shows The cost of PR shows The cost of PR shows whether it worked. And whether it worked. And whether it worked. And right now, for most right now, for most right now, for most companies, the answer is companies, the answer is to keep to keep to keep investing. Tokens investing. Tokens are the cheapest are the cheapest are the cheapest way to gain way to gain way to gain efficiency efficiency . Well done. You have . Well done. You have . Well done. You have reached the end. reached the end. reached the end. Here's how you can Here's how you can Here's how you can follow our follow our follow our latest latest latest research, research, research, based on the based on the based on the industry's largest industry's largest industry's largest dataset of dataset of dataset of engineering engineering engineering workflows and workflows and workflows and AI telemetry. You AI telemetry. You AI telemetry. You can follow can follow can follow me on LinkedIn, me on LinkedIn, me on LinkedIn, my colleagues, or my colleagues, or my colleagues, or Jellyfish. Subscribe to Jellyfish. Subscribe to Jellyfish. Subscribe to our Substack or our Substack or our Substack or check out our check out our check out our Jellyfish blog on the website. Jellyfish blog on the website. If you need If you need help tracking help tracking engineering processes engineering processes and AI performance in and AI performance in and AI performance in your organization, your organization, your organization, including the cost of including the cost of including the cost of unified PR, unified PR, unified PR, go to our website go to our website go to our website and book a demo or and book a demo or and book a demo or write to me and I write to me and I write to me and I will connect you with the will connect you with the will connect you with the right specialists right specialists . Once again, I'm Steve . Once again, I'm Steve . Once again, I'm Steve Jameson, Head of AI R&D Jameson, Head of AI R&D at Jellyfish.

at Jellyfish. Thank you for watching. Thank you for watching. Thank you for watching. See you later.