Where to compete in AI

A product announcement from an AI lab can be an odd thing to read when you're building with its models. You might find a way to offer customers something that used to be too expensive, alongside a feature you were planning to sell.

ChatGPT Sites supports creating and publishing websites and lightweight apps. That puts pressure on a company selling the same capability. It also gives a small team a way to deliver software it previously needed more people to build.

I don't want to build on the assumption that the labs will leave deep industry work to everyone else.

What I think the labs are building

OpenAI's charter describes an ambition spanning most economically valuable work. I expect pursuing it to keep pushing its products into more domains.

They are already entering industry-specific work. ChatGPT for Financial Services includes tools for producing models and client materials. Anthropic's finance agent templates include ledger reconciliation and month-end close. OpenAI also launched its Deployment Company and announced an agreement to acquire Tomoro, an applied AI consulting and engineering firm.

I would build on the assumption that the labs will keep trying to complete more of the customer's work. A specialist has to earn a place beside them without a reliable map of where they will stop.

What would someone keep paying for?

I would look for recurring work where a specific result is hard to get from a general agent. Better models may make that work affordable to deliver for the first time. Julien Bek's Services: The new software makes the case for selling the completed work, such as closing a customer's books. The provider has to gather authorized context, deliver the result reliably, and charge enough to cover the work that remains.

Suppose we build a language tutor for adults who can read English but struggle to hold a conversation. A general agent that can talk with them and remember their goals is already a serious alternative. Adding a lesson screen would give them little reason to pay us too.

We would need to find where they get stuck. Perhaps they can answer familiar questions but lose the thread when the other person interrupts. Our service could choose practice around that difficulty and check whether they need fewer prompts over time. The result I would care about is whether they can handle a conversation outside the lesson.

For a small retailer preparing a weekend promotion, the paid result is different. The owner has to select eligible products, check that the discount leaves enough margin, get her business partner's approval, and make sure the offer ends on Monday. Generating a page is one part of that job. She would need to spend less time preparing and checking the promotion, or avoid enough costly mistakes, to justify what we charge.

How much work does each customer add?

Shopify offers its own AI Store Builder and also supports stores built with Lovable, with Shopify handling payments and inventory. We could use an existing builder and concentrate on what the owner still needs help with.

Shopify already supports automatic discounts and lets owners ask Sidekick to create or troubleshoot them. Before building a promotion service, I would watch which records the owner still checks and what she has to repair.

Connecting the catalog may still leave us with work specific to that store. A discount could apply to one size rather than the whole range, and a bundle could have a different price from its individual items. We have to make the advertised offer agree with checkout. If every store represents those cases differently, each sale may create another setup project.

If several retailers share the same bundle problem, a tested solution could reduce the next store's setup work. For the tutor, I would ask how much we need to learn before giving a new student a useful first session. Personalization that takes a week to become useful may lose them first.

Custom work can support a profitable service if customers pay enough for it. A teacher's review, store configuration, compute and support all belong in the delivery cost.

Which data makes the next job better

The tutor could retain practice history with the learner's permission. If learners repeatedly struggle after an exercise, we could change its sequence and test the revision with new learners. I'd check whether they handle an unfamiliar interruption with fewer hints. More time spent chatting would tell me little about that.

Suppose store owners repeatedly correct how the software handles bundles. Those corrections could reveal a missing pricing rule or a lost connection between the bundle and its products. We can turn the failure into a check, then see whether owners spend less time correcting later promotions.

An improved exercise or tested pricing check may be reusable while the underlying customer records remain private. We need permission for the evidence we use and a way to tell whether the change helps the next customer.

What a competitor would have to reproduce

Suppose the learner can give a general agent their goals and recent mistakes. If it can then provide equally useful practice, remembering that history has given our tutor little protection. I would want to know whether our exercises, assessment, and teaching choices produce a difference the learner can notice. Their personal agent might remain where they keep their goals even if they pay us for lessons.

For the retailer, connecting the same catalog may be easy for a rival. Knowing which bundle rules fail and how to introduce the service without making every owner explain the business again could take longer. An incumbent may already have both that knowledge and the customer relationships. I'd compare the completed work, checking left to the owner and cost of delivery.

An easier experience or a lower price we can sustain can also win customers. We can build a business while testing how much of an advantage we have.

Where the customer starts

Salesforce's September AIforce announcement describes making its records, permissions and business rules available through other interfaces, including a beta inside Claude. A customer can reach Salesforce's work while spending less time in its interface.

Our specialist service could occupy a similar position inside a personal agent. That agent might bring us customers, but it could also compare us with another provider on the next request. I would want to know whether customers ask for our service by name, whether we can reach them directly and what happens if the platform stops recommending us.

When choosing where to compete, I would start with people whose next attempt I can observe. For the tutor, that means seeing a learner get stuck, trying a lesson, and finding out whether they want another one after using the alternatives. I would learn more from that than from counting how many conversations our product had stored.