Data & AI Hot Takes for 2026

Four non-consensus predictions about agentic AI, jobs, RTO, and pricing that I’m willing to be wrong about

Why writing your forecasts down matters

I spent eight years as a senior sell-side analyst¹ before changing careers. I got enough calls right to stay employed. I also got plenty wrong, loudly and publicly.

I’ve made enough forecasts in public to know this: perfectly predicting the future is impossible, but not writing your views down is worse. In hindsight, everything feels inevitable. In real time, it never does. That is confirmation bias and hindsight bias at work. Forward-looking uncertainty is the only place where alpha² lives.

Why even make predictions? On the sell side, you can technically avoid them. I once worked with someone who never really made a prediction. They mostly wrote about how everyone else was wrong, usually via a strawman version of consensus³ that no real institutional investor actually held. They included lots of humor in their writing, and it was entertaining. It worked for their career.

I tried to do the opposite. I went for non-consensus calls and outlined the specific fundamental reasons why. The goal wasn’t just to be right; it was to advance the debate. On the buy-side, the logic is the same. The best investors set clear forecasts to establish a baseline. You cannot test a view you never articulated.

Good forecasters write things down. They establish a base case, think in probabilities, look for disconfirming evidence, and adjust. That process matters more than being right. In establishing a base scenario, they draw from first principles and structurally similar past cases. They explore the drivers and constraints of the world that narrow the range of possibilities. They try to avoid the tendency to only think linearly. They think in probabilities, not certainties. They consider what information would change their minds and seek that info out aggressively.

Not only do good forecasters write things down, but they are also specific. The forecasts should pass the clairvoyance test. A forecast passes the clairvoyance test if, once future data is available, it is unambiguous whether the prediction was right or wrong.

This article is not investment research. It is my forecast for how the practice of using data and AI will evolve as an industry. These views are my own and do not reflect the views or strategies of any past or present employer.

These are my most controversial takes. If you agree with all of them, I probably failed. The detailed measurable forecasts associated with each of these are included in the article.

1. Agentic AI will work. It just won’t help companies without a data strategy

2. Usage-based pricing is coming whether we like it or not.

3. Gen AI doesn’t level the playing field. It exposes judgment gaps.

4. Generative AI will accelerate the return to the office for knowledge workers.

(Note: These are listed in rising order of likely “pitchforks” from the audience at my upcoming BattleFin Miami panel with Dan Entrup and Michael Watson).

---

My friends at BattleFin are looking for new data providers to participate in their upcoming “New Data Provider Showcase” at BattleFin Miami. Apply by January 9th.

https://web.battlefin.com/miami-2026-new-data-provider-showcase-data-chorus

💡If you provide data or analytics that help investors see financial performance before it shows up in earnings, we’d like to hear from you.

If you, the firm, are new to BattleFin i.e. never participated at any one of our Discovery Day Conferences.

Please apply by Jan 9th, 2026 with a short description of your dataset(s), coverage, delivery methods, and any additional information that describes potential insights from the data.

Dan Entrup of AggKnowledge and “It’s Pronounced Data,” Michael Watson of Hedgineer, and I will help BattleFin decide which three providers will come to BattleFin Miami to pitch their product, with the winner receiving a complimentary data table at the next BattleFin conference.

Use this link to apply. https://web.battlefin.com/miami-2026-new-data-provider-showcase-data-chorus

---

Welcome to the Data Score newsletter, composed by DataChorus LLC. This newsletter is your source for insights into data-driven decision-making. Whether you’re an insight seeker, a unique data company, a software-as-a-service provider, or an investor, this newsletter is for you. I’m Jason DeRise, a seasoned expert in the field of data-driven insights. I was at the forefront of pioneering new ways to generate actionable insights from alternative data. Before that, I successfully built a sell-side equity research franchise based on proprietary data and non-consensus insights. I remain active in the intersection of data, technology, and financial insights. Through my extensive experience as a purchaser and creator of data, I have a unique perspective, which I am sharing through the newsletter.

---

#1 Agentic AI will work. It just won’t help companies without a data strategy

We hear nonstop about agentic AI⁴ taking over workflows. In theory, it makes sense. Call the right information, route it through the right processes, and generate the right output. We have been doing deterministic versions of this for decades. It is not that different from an orchestration tool that transforms raw data into insight-ready data.

The problem is that generative AI (GenAI)⁵ is non-deterministic⁶. That makes data quality non-negotiable.

Here is what will not work: Agentic workflows applied to poorly governed data where the AI is expected to magically infer a missing data dictionary, nonexistent semantic layer⁷, and uncleansed inputs. It is unlikely that a foundational model⁸ coming in 2026 that can reinterpret “Bad Data” into “Good Data” on the fly.

Here are the baseline stats to explain the current situation:

• In early 2025, Simcorp published a report highlighting its survey of 200 buyside operations leaders in Q4 2024. https://www.simcorp.com/resources/insights/white-papers-and-reports/2025-global-investops-report

• Only 9% of asset managers feel very prepared (actively exploring AI tools and feeling confident about their ability to integrate them).

• In a separate Simcorp report referencing a McKinsey study, 78% of asset managers said they have piloted agentic AI, but only 27% see business impact. https://www.simcorp.com/resources/insights/industry-articles/2025/scaling-ai-agents-in-investment-management

• Citisoft published a report on transformation for asset managers based on a survey of nearly 70 asset managers: https://www.citisoft.com/hubfs/2025 Transformation Survey/2025 Citisoft Transformation Survey - Survey Report.pdf - 90% of asset managers plan to overhaul their data operations and governance⁹ by 2027.

I would argue that there is a high overlap between the asset managers struggling with their data strategy and the asset managers struggling to leverage GenAI.

The good news is that investments in AI-ready data will start to pay off for teams that made the initial investment and commitment to the outcome. You cannot have an AI strategy without a data strategy.

Prediction: By Q4 2026, the percentage of asset managers reporting “Substantial Business Impact” from Agentic AI (currently 27% per McKinsey/Simcorp) will not exceed 40%, with the upside driven by companies that have executed a successful data strategy. 

Agentic AI will work on both the buy side and sell side, but only at firms with real data strategies. Everyone else will stay stuck in proof-of-concept purgatory. The outcome gap between firms will widen.

Where this forecast could be wrong: You could argue that generative AI foundational models will continue to advance reasoning skills and be able to infer what is meant from the poorly tagged and governed data, and in the meantime, human-in-the-loop¹⁰ can step in to make sure it is working properly. I would welcome the advancements in computer inference and deduction, but in my opinion, it’s not likely to be Gen AI or Large Language Models (LLM)¹¹ models that achieve it. In the meantime, even with humans in the loop, if GenAI doesn’t save time for the human who has to recheck everything manually, adoption stalls.

#2 Usage-based pricing is coming whether we like it or not

As AI embeds variable compute costs into data products, flat-price free trials become economically unstable.

A few years ago, on the “I See Data People” podcast, Mark Fleming-Williams joked that anyone asking for a paid data trial should be put in jail. For pure data feeds and for quant funds, he’s right about how paid trials undermine how quants are able to assess the data and are done because vendors misunderstand how the data is used to generate alpha going forward. Free trials matter for diligence (quote at ~10 minutes into the conversation).

AI breaks the free trial model. Once insight generation is embedded into a data product, costs become variable. I’m not talking about chatbot use cases for support or tutorials to use an interface. I’m talking about embedded insight generation based on the data product content where the AI is intertwined into the product. Agentic workflows, reasoning-heavy queries, and deep research all burn real compute power and tokens. A free trial with unlimited AI usage can bankrupt a startup.

There are clear price anchors when setting per-user pricing: On one end, $200-per-month “pro package” AI model subscriptions with capped usage. On the other, the Bloomberg terminal with bundled services and a high subscription price per user.

The solution is price transparency and separation.

• Pure data access should reflect its fixed-cost nature and support time-boxed free trials. We need, as an industry, to separate the pricing structure for pure data products that are delivered through programmatic access (e.g., API, database shares, etc.) without any GenAI enhancements. The pricing model should reflect the heavily fixed cost nature of the product, which lets the data vendors enable time-boxed free trials and also enable competitive pricing based on the value of the data at scale.

• For the Gen AI portions of the data product, there needs to be usage-based pricing (UBP)¹², where the vendor can protect the quality of their product and their cash flow.

  • AI utilities should be usage-based and priced close to pass-through, not marked up on foundational model costs.

  • Not only does this utility pass-through protect the vendor’s cash flow, but it also lets the user decide how they want to use their token budget¹³. And it should help avoid an unsustainable business model where data companies try to charge a markup on the Gen AI base model utility cost.

• Vendors should monetize their proprietary data and workflows and overall value add through subscriptions separate from the GenAI variable utility costs.

  • For example, raw transaction data delivered via API should be priced as a subscription, while agent-driven research queries that burn tokens should draw down usage credits.

As we move toward a post-dashboard world where insights are consumed rather than discovered, usage-based pricing for both data and compute becomes unavoidable. This may actually unlock the long tail of the alternative data market. This can unlock the ability to pay for insights by more companies, while the largest data buyers continue to leverage programmatic data access and their own models to generate proprietary insights.

The blocker to this increasingly transparent pricing structure is that the data companies must accurately assess where they are adding value and also avoid the temptation to maximize margins through blurred pricing that actually is just a markup on that GenAI utility.

Baseline data? There isn’t any on the data market’s use of UBP.

• The SaaS world has already crossed over to usage-based pricing (UBP), with 70% of contracts now featuring usage-based components according to Maxio (https://www2.maxio.com/l/699023/2025-01-31/23ndsk/699023/1738348735i2fnTL26/Saas_Growth_Strategies_for_2025_Maxio_250131.pdf),

• Or 85% of software companies have adopted the model, per Metronome https://metronome.com/state-of-usage-based-pricing-2025#:~:text=By%20the%20numbers,software%20companies%20have%20adopted%20UBP* ).

• Meanwhile, the alternative data market remains a black box. We know of a few vendors offering usage-based pricing, but the vast majority are enterprise or seat license “all you can eat.” Currently, usage-based pricing (UBP) in alt-data is too small to even be measured in industry surveys.

Prediction: We will see the alt-data market move from ‘too small to measure’ to 10% of all new contracts containing some form of usage-based or credit-drawdown pricing by the end of 2026 as measured by various industry data reports (assuming it’s actually measured by the data industry). 

Inclusion of Gen AI in the product will move the industry toward usage-based pricing, affecting the potential for fully free trials. Transparent pricing separation between proprietary data access, true value-add workflows, and AI utility will become table stakes. Free trials for data access and proprietary content can continue, but the AI utility cost will be passed along to the user, even in the trial period.

Where this forecast could be wrong: I get the argument that the vendors should eat the cost of the token use during the free trial period to enable high repeat usage. Perhaps it’s an option for the largest, most financially stable data vendors, but this is really about incentives. “Free tokens” potentially masks what the real valuable outcomes are for the users and may lead to misuse of expensive reasoning workflows. Another angle could be that the annual license fee could include estimates for high usage costs embedded in the single seat price, but this may push prices toward that upper bound of a per-seat license, which would limit adoption.

#3 Gen AI doesn’t level the playing field. It exposes judgment gaps.

There is a common view that AI brings the bottom performers up to an average level. However, I believe it doesn’t actually bring up the bottom performers. Gen AI will not lift everyone by default. Used well, it allows teams to do more and do better with the people they already have. Used poorly, it exposes gaps in judgment rather than closing them.

This is not a claim about headcount reduction or company policy. It’s a claim about how judgment scales when non-deterministic tools enter workflows.

What I see instead is widening dispersion driven by how AI is used, not by the technology itself. Strong performers use AI to accelerate thinking, pressure-test ideas, and eliminate low-value work. Weaker use looks productive but isn’t. The output sounds polished while missing the point.

Can we measure the AI Slop effect today? A company called Better Up is trying to answer it in partnership with the Stanford Social Media Lab (and labeled it “workslop” instead of “AI slop”). https://www.betterup.com/blog/hidden-costs-workslop

Our September 2025 survey of 1,004 full-time U.S. desk workers revealed that:

• 40% of employees believe they’ve received workslop in the last month.

• On average, workers estimate that 15.4% of the work they receive is AI workslop.

• Managers are more exposed: 54% report receiving workslop, compared with 38.5% of individual contributors.

• More than half (53%) admit that at least some of the work they themselves send may be workslop. In other words, workslop is circulating in every direction—sideways between peers (40%), up the chain to managers (19%), and down from leaders to their teams (16%).

I would note that these are perception-based measures, not an objective quality score, but perception is what drives trust, rework, and management overhead.

AI as a Salesperson?

Sales content is the most common offender I see day to day. It is verbose, generic, and disconnected from real problems. Granted, a human could generate useless content and distribute it and get the same effect. However, what’s different now is that useless content can be mass-produced without constraint.

The metrics make this worse. Open rates and content volume optimize for activity, not outcomes. The only sales metric that matters is contracts signed. Flooding inboxes with AI content digs a deeper hole until senders get flagged as spam. Open rates may be high without conversion until they plummet (sender is marked as spam).

As Dan Entrup often reminds people, data is not bought; it is sold. AI isn’t going to sell the data. If anything, it’s going to do the opposite. Building relationships to understand the real problems to be solved and then creating a scalable solution that the potential client is.

Dead internet content?

The same pattern shows up across the internet. AI content may drive traffic, but traffic without value does not convert to repeat usage. No repeat usage, no durable business. Perhaps AI-generated content will attract the less discerning audience. As low-quality content fills the internet, it’s hard to see past “the dead internet” theory¹⁴. The antidote is content created with human judgement and good taste.

AI as a copilot?

We are rapidly approaching text-to-app solutions and text-to-financial-model solutions (not just snippets of code). However, in the current text-to-code and text-to-analysis gen AI, it’s usually user error that leads to the poor results. Often it’s a poorly worded prompt or an ill-conceived idea that leads to the poor outcome.

There’s no shame in trying, and iteratively sparring with the AI can get to a great outcome, or deciding to roll up sleeves and just do it the old-fashioned way. The shame comes from missing that the AI did the wrong thing but the user didn’t have the judgement or process to catch it.

How can we help the most exposed workers to these risks?

Junior employees are especially exposed without clear review processes and feedback. AI output can sound polished without being right. Without reps, it is hard to spot the flaws. That leads to confident presentations that collapse under scrutiny.

AI can help as a checker. Asking it to tear apart an idea is genuinely useful. It can be a great BS catcher if prompted to call out all the flaws in the content. That can help protect the non-discerning who miss the nuance of the content that invalidates the idea.

For the content creators who don’t notice the work is “AI Slop,” AI could be used as a checker to catch the errors where the user initially used AI to generate bad output. The initial user can do this to catch their own errors in judgment before sharing it with the world.

The risk isn’t using AI and getting a bad result. The risk is failing to recognize when the result is wrong.

Prediction: By the end of 2026, we will see workforce-wide observations that 30% of the work received is recognized as being “workslop” or “AI slop,” up from 15%. 

AI will generate enormous activity. The advantage will accrue to teams and individuals who invest in judgment, training, and strong review processes around AI-generated work. Used well, AI raises the bar. Used poorly, it widens the gap.

Where this forecast could be wrong: There is a common counterview that generative AI is just another wave of automation. Historically, automation removes deterministic tasks, lifts the bottom, and narrows performance dispersion. I agree with that logic for automation. This is where generative AI is different. Automation removes deterministic tasks. Generative AI produces nondeterministic output that still requires human judgment to validate, contextualize, and decide what matters.

#4 Generative AI will accelerate the return to the office for knowledge workers

AI is going to bring people back to the office. Not despite productivity gains, but because of them.

The prevailing narrative says AI boosts productivity, reduces headcount, and enables more remote work. I think that is wrong. AI will be one of the forces that pushes knowledge workers back into offices.

This sounds counterintuitive until you follow the logic.

• Gen AI repackages existing information. GenAI can only produce what it’s been given in the prompt, training materials, and information it has access to. It cannot create something new on its own.

• AI removes rote work. AI can remove mundane, repetitive processes, which should be used to free time for deeper, thoughtful, and differentiated work.

• It does not take half-baked ideas and make them credible. Sure, AI can turn half-baked ideas into confidently written 20-page memos in just a few moments. Sure, some of the AI-generated memos include good ideas. But now there are many long documents to read and not enough time. So they get distilled back into a 200-character sound bite by the reader’s first prompt.

• Junior workers do not yet have the reps, wisdom, and judgment to separate signal from noise efficiently.

• New ideas still require organizational buy-in. That buy-in has always been relationship-driven. Even investment recommendations need network effect buy-in across investors for the narrative to be validated in the security’s price.

• The remaining human value is idea generation, judgment, and consensus building. Those are social skills. They improve with proximity.

If AI increases the volume of plausible-sounding but flawed ideas, human supervision needs go up, not down. Until models can generate truly new ideas and decide which ones matter, humans have to lean harder into what machines cannot do.

AI handles the rote work; the value left for humans is thoughtful idea generation, which requires human-led, relationship-driven consensus building, which is inherently an exercise of “in-person collaboration.” Because AI can be poorly used to generate “AI slop,” junior employees need more supervision, not less, so they can steer decisions in the right direction.

When idea volume rises but attention does not, alignment becomes the scarce resource. Alignment scales faster in high-bandwidth, in-person settings than in asynchronous channels.

How has return to office evolved over the last 5 years? When we look across all industries, we see a rising trend in return to office as measured by Kastle’s 10-City Barometer. https://www.kastle.com/safety-wellness/getting-america-back-to-work/ Currently there is 55% average occupancy across the 10-city barometer as measured by Kastle, with 63% peak day occupancy in those cities.

The time series shows a gradual increase in the 10-city average (it’s the dark red line in the 10 lines, now at 55%).

Prediction: By Q4 2026, the Kastle 10-City Barometer weekly average will break 60% for the first time since the data was tracked, and Class A+ building occupancy will remain above 90% for at least three days a week (Tue-Thu). 

This is not a statement about any specific firm’s policy, but about observed industry dynamics. The realization of AI removing mundane tasks is the increased focus on value-add and new ideas, which will require relationships and direct communication to break through the noise of countless half-baked AI-generated ideas. The most progress will be made with face-to-face collaboration, which brings people back to the office more frequently.

Where this forecast could be wrong: Of course great teams can do this virtually. It’s true for top-performing teams, but still we should expect more time in meetings, 1-on-1s, and synchronous discussion to have more iteration on the “so what.” For every 20-page memo generated by AI, there’s another prompt collapsing it into a tweet. So collaboration and communication need to take the place of that dance, even if virtual. When teams come together in person, that pace of understanding is accelerated.

Concluding thoughts

Writing forecasts down and sharing them is uncomfortable. That’s the point. It exposes you to being wrong. But it is the best way to learn and ultimately get to better outcomes proactively. Being specific so that the forecasts can pass the clairvoyance test when the future date of the forecast arrives.

These are my views on where data and AI are headed as an industry (These views are my own and do not reflect the views or strategies of any past or present employer). These views will age. Some will be wrong. That is fine. The real mistake is not having a written view at all and falling trap to hindsight bias.

Detailed Predictions:

1. Agentic AI will work. It just won’t help companies without a data strategy Prediction: By Q4 2026, the percentage of asset managers reporting “Substantial Business Impact” from Agentic AI (currently 27% per McKinsey/Simcorp) will not exceed 40%, with the upside driven by companies that have executed a successful data strategy.

2. Usage-based pricing is coming whether we like it or not. Prediction: We will see the alt-data market move from ‘too small to measure’ to 10% of all new contracts containing some form of usage-based or credit-drawdown pricing by the end of 2026 as measured by various industry data reports (assuming it’s actually measured by the data industry).

3. Gen AI doesn’t level the playing field. It exposes judgment gaps. Prediction: By the end of 2026, we will see workforce-wide observations that 30% of the work received is recognized as being “workslop” or “AI slop,” up from 15%.

4. Generative AI will accelerate the return to the office for knowledge workers. Prediction: By Q4 2026, the Kastle 10-City Barometer weekly average will break 60% for the first time since the data was tracked, and Class A+ building occupancy will remain above 90% for at least three days a week (Tue-Thu).

If you don’t have at least one take in here you disagree with, I probably didn’t push hard enough.

- Jason DeRise, CFA

1. Sell-side & Buy-side: Buy-side typically refers to institutional investors (Hedge funds, mutual funds, etc.) who invest large amounts of capital, and Sell-side typically refers to investment banking and research firms that provide execution and advisory services (research reports, investment recommendations, and financial analyses) to institutional investors.

2. Alpha: A term used in finance to describe an investment strategy’s ability to beat the market or generate excess returns. A simple way to think about alpha is that it’s a measure of the outperformance of a portfolio compared to a pre-defined benchmark for performance. Investopedia has a lot more detail https://www.investopedia.com/terms/a/alpha.asp

3. Consensus: “The consensus” is the average view of the sell-side for a specific financial measure. Typically, it refers to revenue or earnings per share (EPS), but it can be any financial measure. It is used as a benchmark for what is currently factored into the share price and for assessing if new results or news are better or worse than expected. However, it is important to know that sometimes there’s an unstated buyside consensus that would be the better benchmark for expectations.

4. Agentic AI Frameworks: A type of AI system that autonomously makes decisions and executes tasks with minimal human intervention, often used in data-driven workflows to enhance efficiency and automation. An example could be a large language model chat model used as an interface that calls other AI models depending on the prompt and, in turn, leverages other specific AI agents that handle specific tasks to enable the outcome.

5. Generative AI: AI models that can generate data like text, images, etc. For example, a generative AI model can write an article, paint a picture, or even compose music.

6. Non-deterministic: Systems whose outputs are probabilistic rather than fixed, meaning the same input can produce different outputs across runs.

7. Semantic Layer: A translation layer that converts raw database structures into human-readable, business-relevant terms, making data easier to analyze.

8. Foundational Model: Large, general-purpose AI models trained on broad datasets that can be adapted to many downstream tasks (e.g., large language models).

9. Data Governance: The overall management of the availability, usability, integrity, and security of the data employed in an organization. “Data governance is everything you do to ensure data is secure, private, accurate, available, and usable. It includes the actions people must take, the processes they must follow, and the technology that supports them throughout the data life cycle.” - Google Cloud’s definition: https://cloud.google.com/learn/what-is-data-governance#:~:text=Get the whitepaper-,Data governance defined,throughout the data life cycle.)

10. Human-in-the-loop: This is an approach to AI and machine learning where a human collaborates with the AI model during its operation, guiding its learning and correcting its output.

11. Large Language Models (LLMs): These are machine learning models trained on a large volume of text data. LLMs, such as GPT-5 or ChatGPT, are designed to understand context, generate human-like text, and respond to prompts based on the input they’re given. It is designed to simulate human-like conversation and can be used in a range of applications, from drafting emails to writing Python code and more. It analyzes the input it receives and then generates an appropriate response, all based on the vast amount of text data it was trained on.

12. Usage-based pricing (UBP): Usage-based pricing refers to pricing models where customers pay based on actual consumption of a resource, such as queries, tokens, compute time, or data volume, rather than a fixed subscription fee.

13. Token Budget: A token budget is the limit on how much AI model usage a user or team can consume, typically tied directly to compute cost.

14. Dead internet theory: The idea that much online content is increasingly generated by bots rather than humans, reducing authenticity and value.