Can AI Make Work Worth Doing That Was Never Worth Doing Before?

A financial research firm used Claude to test a question it could never justify before. Here is what that means for a small business in Northeast Indiana.

Can AI Make Work Worth Doing That Was Never Worth Doing Before?

Yes, when a test costs less than the answer is worth, AI lets a small business try work it could never justify before. The clearest recent example involves neither a restaurant nor a clinic, but a financial research firm that spent under a couple hundred thousand dollars testing a question about drug discovery, as described in the October 21 episode of The AI Daily Brief.

The lesson for owners is less about biology and more about how cheap experiments change what a business decides to attempt in the first place.

Key Takeaways

  • Citrini Research, a financial research firm, used Claude to design scorpion anti-venom protein candidates, then had an independent lab test them.
  • Of 318,000 candidates, 17 bound four different scorpion toxins, but none has yet been shown to neutralize venom.
  • Verification and physical testing set the pace, since design took 64 hours while results took 39 days and release could take four to seven years.
  • AI can make an experiment cheap enough that a business tests a question it could never justify before, provided someone checks the answer.
  • An AI teammate can draft the first pass of this kind of work, but nothing goes out without your approval.

What did Citrini Research actually test?

Citrini Research wanted to know whether AI could speed up drug discovery, and it could not settle that question by reading expert opinions alone. The firm is a thematic equity and global macro shop that publishes a Substack on investing. Its analysts had concluded that life sciences looked undervalued, a sector sitting at cyclical lows with a possible AI-driven upside.

Yet surveying biology experts produced views that ranged from the claim that AI has never made an important drug to the claim that it will cure all diseases within five years.

So the firm ran the experiment itself. It used Claude to design proteins that bind scorpion neurotoxins, contracted a lab to synthesize and validate them, and then used the model again to interpret the results. The starting point was a thirty-five-page prompt that Anthropic had already tested in an earlier study, which Citrini modified to suit its own purpose. According to Anthropic's account, the firm approached the company in July with a question about whether Claude could help motivated non-experts tackle a genuine problem in human health.

The target was chosen for a reason that matters to anyone deciding where to spend limited effort. Scorpion stings affect about 1.2 million people a year and cause roughly 3,200 deaths, mostly among children with little access to medical care. Anthropic argues that treatments for neglected conditions like this are expensive to develop and unlikely to earn back their cost, which is precisely the gap that a cheaper discovery process could narrow. That argument comes from Anthropic itself, not from an independent study.

Why did the experiment make economic sense for a firm?

The experiment made sense because it was the cheapest way for a firm to answer an investment question, so work that could never be justified on the core business alone suddenly became reasonable to attempt. Citrini did not need to become a biotech company in order to benefit from the result, and it has said as much.

The firm remained, in every practical sense, exactly what it had been before the experiment. Citrini remains a financial research firm, and the experiment served its own investment thesis rather than a new business line. Before AI, a team that wanted a real answer about biology would have needed to hire scientists, lease laboratory space, and wait months for results. Those costs would have made the question impossible to justify for a shop of its size, no matter how interesting the answer might be.

AI changed that arithmetic. When design work takes hours instead of months, a small team can test a hypothesis for a fraction of the old cost. The host describes the outcome as orthogonal work, meaning effort that does not sit on a firm's main product line but still makes its core work better. The host treats this as the opportunity side of an efficiency-versus-opportunity divide, where AI both saves time on existing tasks and makes new tasks worth attempting.

Citrini made the same point in its own write-up, saying its digital work was now set to "create work for human scientists."

How much did the experiment cost, and what did it produce?

The whole effort cost under a couple hundred thousand dollars and produced 17 candidate binders, but none has yet shown that it can stop venom from working. Those two facts belong together, because the cost looks low only in relation to a result that still needs considerably more testing before anyone could call it useful.

Taken together, the numbers describe a funnel with a very wide top and a narrow, expensive bottom. Citrini ran 300 parallel Claude sessions, which generated 318,000 protein candidates, and the team shortlisted 211 of them for further testing. An independent lab then confirmed that 17 of those candidates bind four different scorpion toxins. A very large generation step narrowed down to a small set of physical results, and the physical results are the part that counts when a decision depends on evidence.

The funnel also shows where the limit sits. Binding is a long way from neutralization. None of the 17 candidates has yet been shown to stop a neurotoxin from working, and the team says it still intends to run neutralization tests. The claim of progress rests on the firm's own write-up plus binding results from one outside lab, so readers should wait for neutralization data before treating this as a treatment.

That same caution applies to any claim a vendor makes about a process you have not yet measured inside your own shop.

Does binding a toxin mean a treatment works?

No. Binding means a protein sticks to a target, while neutralizing means it stops that target from causing harm, and the second step has not been shown in this case. Confusing the two is a common mistake when people read a promising early result, and it is worth avoiding before any budget decision follows.

Critics raised this objection almost immediately after the write-up appeared. One critic on X noted that proteins binding scorpion toxins had appeared in the scientific literature more than fifteen years earlier, and that the missing piece is neutralization. Citrini replied that the team expected skepticism, did not want to wait another three months before publishing, and still plans to pursue neutralization testing.

Alex Emoss, director of AGI economics at Google DeepMind, offered a useful framing. He said the work showed large speed-ups in generating hypotheses and running computation, while also showing that verification and testing remain the bottleneck. Owners can draw the same practical lesson from this story, which applies to nearly every operational decision they will face over the coming year. Where a process is slow, find the step that truly limits it before spending money on the step that feels most exciting.

How does an AI teammate use this kind of thinking in your business?

An AI teammate drafts the work and returns it to you, so the slow step, your review, stays in your hands and nothing goes out without your approval. That split matches the lesson about bottlenecks, because the fast part gets faster while the judgment part stays with a person who knows the customer.

Everyday small-business work has the same shape. Quoting a job, answering a customer, or sorting incoming requests all contain a quick step, drafting, and a slower step, checking that the draft is accurate and appropriate.

An illustrative example makes the split concrete. An inquiry such as a request for a spring cleanup quote gets organized, the missing details such as the address and photos are identified, and a reply asking for them is drafted for your approval. Our post on who checks AI work before customers see it covers that review step in more depth.

The question of access levels matters considerably here as well. Hey Button sets up read-only, draft-and-approve, or full access, and go-live means clear approval rules for what the teammate may read, draft, and do on its own. Customer messages, spending, and sensitive changes stay behind the rules you choose. Because AI can make mistakes, review is part of the setup rather than an afterthought. Owners who settle these limits before launch avoid giving a tool more reach than it has earned through real work.

When does this idea not apply to your business?

The idea applies least when your work is routine and already fast, because an experiment only pays off when the question is valuable and the answer will actually change a decision. Most small businesses do not need a wet laboratory or a research budget to benefit from AI, and pretending otherwise wastes time that could go to customers.

A lawn care company that quotes jobs in an afternoon gains little from a new experiment on quoting speed, since the current process already works. The idea matters more when a real question has sat unanswered because testing it felt too expensive, too slow, or too specialized for the team's limited hours. A contractor who suspects that a new estimate format wins more jobs, but has never had time to compare formats side by side, sits squarely in that position.

There are also clear limits. An experiment is only useful if someone checks the result before acting on it. A business that tests an idea with AI and skips verification can end up with confident but wrong conclusions, and acting on those conclusions can cost more than the test saved in the first place. Start small, decide in advance which result would change your mind, and keep one accountable person responsible for the final judgment.

How do you pick a first experiment without overspending?

Pick one question tied to a real decision, set a cap on time and money, and define in advance the result that would change what you do. Those three steps keep a test from becoming an open-ended project that consumes staff hours, attention, and budget without producing a usable, decision-grade answer.

Start with a question you already need answered. For a Fort Wayne contractor, that might be whether a new estimate format wins more jobs, or whether a draft follow-up reduces the number of unanswered quotes. For an Auburn clinic front office, it might be which scheduling requests take the longest to sort each week. Write the question down, then write the answer that would lead you to change a process, so the result has a clear purpose before the work begins.

Next, cap the test. Give it a fixed number of hours and a fixed budget, and decide the stopping point before you start. Track what the test really costs, including staff time, not only software fees. Our post on AI cost per task for a small business budget explains how to think about spending on one piece of work. Keep the people who will live with the result involved from the start, because they will notice problems that a spreadsheet alone would miss.

What else was in the October 21 episode?

The episode also covered Meta's Muse app, a security problem in its share links, Anthropic's audited financials, and OpenAI's acquisition of a biology tools startup. Each story offers a separate lesson about how AI companies compete for users and capital, and each one connects back to whether a claim can be trusted before anyone verifies it.

Muse reached 10 million US downloads in 40 days, a pace that only about a dozen apps have matched. According to Sensor Tower, Muse took 39 percent of Meta's October US ad spend, and its seven-day retention of 30 percent beat ChatGPT's roughly 25 percent in the same comparison. A security firm then enumerated 6,800 shared Muse links, found private business records among them, and Meta has temporarily removed the feature while it hardens the links.

On the money side, the Wall Street Journal reported that Anthropic's first-half net loss was about $80 billion. Roughly $72 billion of that came from an on-paper stock-grant accounting change, leaving an operating loss of $8.7 billion. Audited first-half revenue was $16.3 billion. Reuters also reported that OpenAI is trying to close a $30 billion raise at a $1.4 trillion valuation within weeks, and that OpenAI acquired Orchestra Bio, a four-year-old startup that builds tools for lab workflows such as compound manufacturing monitoring and trial tracking.

Reuters further reported that Anthropic set up its own biology lab, and Anthropic's head of life sciences confirmed it. The host read the two moves as different bets. OpenAI is building tools for existing pharmaceutical companies, while Anthropic is starting to do the physical work itself. Both bets show AI companies moving closer to the work their customers already do, which is a question every software vendor should be ready to answer plainly.

Start with one question

The episode's central lesson, stated plainly, is that an experiment deserves the money and the time only when its answer will change a decision that a business is already facing. If your team has postponed a question because testing it felt too costly, bring it to us. Hey Button starts with a short questionnaire about the work you would hand off first, and the teammate is configured around your real workflows before anything touches a customer. Start the Hey Button questionnaire and tell us which question to test first.

Sources

  • The AI Daily Brief, "How AI Is Changing What Work Means," October 21, 2026 (episode, show notes, and transcript): https://aidailybrief.ai/e/2026-10-21
  • Citrini Research, write-up of its AI-assisted protein design experiment, as summarized in the episode.
  • Anthropic, account of its collaboration with Citrini Research, as quoted in the episode.
  • Wall Street Journal, coverage of Anthropic's updated prospectus draft and audited first-half financials, as summarized in the episode.
  • Reuters, reports on Anthropic's biology lab, OpenAI's fundraising, and investor valuation comments, as summarized in the episode.
  • Sensor Tower data on Meta's Muse rollout, as cited in the episode.
  • OpenAI, announcement of the Orchestra Bio acquisition, as summarized in the episode.

Frequently Asked Questions

Usually yes, if the question is specific and the answer will change a decision. Keep the test small, set a time and money limit, and assign one person to judge the result. You do not need scientists on staff to run a simple test of a workflow, a message, or an estimate format.
Not yet, and the difference matters. A neutralizing molecule stops a toxin from doing harm, while a binding molecule only sticks to it. Citrini's 17 candidates have bound four scorpion toxins, but none has yet been shown to neutralize venom.
Start by listing what the test would cost in staff hours, software, and usage charges, then compare that with what a good answer would save. Ask any vendor which subscriptions or usage charges are identified before you commit. If the total is smaller than the decision's value, the test is worth running.
Any share link that is easy to guess can expose private files to strangers. Before you let a tool create shareable links, ask how those links are built and who can open them. Treat guessable addresses as public, and ask vendors plainly how access is controlled.
The story shows how cheaper AI changes which questions are worth asking, even outside biology. A trades business, clinic, or retailer can face the same pattern, where a question that was too costly to test becomes affordable. The principle carries across industries.
Is it worth running a small AI experiment without a research team?
Usually yes, if the question is specific and the answer will change a decision. Keep the test small, set a time and money limit, and assign one person to judge the result. You do not need scientists on staff to run a simple test of a workflow, a message, or an estimate format.
Does a binding molecule mean a treatment works?
Not yet, and the difference matters. A neutralizing molecule stops a toxin from doing harm, while a binding molecule only sticks to it. Citrini's 17 candidates have bound four scorpion toxins, but none has yet been shown to neutralize venom.
How do I know if an AI test is cheap enough to run?
Start by listing what the test would cost in staff hours, software, and usage charges, then compare that with what a good answer would save. Ask any vendor which subscriptions or usage charges are identified before you commit. If the total is smaller than the decision's value, the test is worth running.
What is the lesson from Meta's Muse share-link problem for my business?
Any share link that is easy to guess can expose private files to strangers. Before you let a tool create shareable links, ask how those links are built and who can open them. Treat guessable addresses as public, and ask vendors plainly how access is controlled.
Why would a small business care about a biology story?
The story shows how cheaper AI changes which questions are worth asking, even outside biology. A trades business, clinic, or retailer can face the same pattern, where a question that was too costly to test becomes affordable. The principle carries across industries.

Hey Button

Want a teammate that handles this for you?

Answer a few questions about your business and Hey Button will show you what an AI teammate would take off your plate, with you approving anything that goes out.

Start the questionnaireSee what teammates do
Lucas M. Button

Written by Lucas M. Button

Founder, Hey Button

Lucas builds AI teammates for small businesses across Northeast Indiana and writes about what works, what doesn't, and what to hand off first. More about Lucas