The first AI service to sell: a one-page table for your manager
Narrowing a new AI business's first service to one buyer, input and output settles sign-off and pricing.
Key points
- Narrow to one buyer, input and output
- With AI, pick work buyers can sign off
- Set sign-off terms first from past cases
- Price per case, exceptions separately
Contents
Narrow the first service by buyer, input and output
A&A perspective
A new AI business picks its first service by fixing one buyer, one input and one output. The first service here means a unit of work that a set buyer receives one case at a time, checks and signs off, and then pays for.
A&A perspective
A pitch like "our AI can take on your sales, your inquiries and your back office" leaves you without a one-line answer when your manager or the board asks who you are selling what to. Once the buyer, the material you receive and the thing you hand back are each fixed, that one line can be written.
From the sources
Anthropic, a US developer of conversational AI, publishes case studies of companies that use its AI. One of them is Lindy, a North American service that lets companies build AI to hand work to.
From the sources
On the case page, Lindy's head of engineering says the vision has always been an AI employee that offers both maximum capability and maximum ease of use.
From the sources
The page gives three examples of use. A software startup uses it to find leads and write the first outreach. A productivity tool company uses it for first-line support on inquiries that arrive around the clock. One executive founder uses it to prepare daily briefings and sort out the calendar.
A&A perspective
Automate & Augment (A&A), which writes this article, reads this list as follows. The vision is broad, but every actual use is narrowed to one counterpart and one job.
A&A perspective
Lindy can claim a broad vision because many companies use it as a tool and the buyer decides what to hand over. A new business selling from scratch has no such mechanism. A broad pitch leaves the prospect to work out what to ask for, and that is not settled in a single sales meeting.
If AI does the work, pick work the buyer can sign off
A&A perspective
Pick the first service by whether the buyer can sign it off. Sign-off here means the buyer checks the delivered work and accepts it. Work that cannot be signed off is also hard to invoice.
From the sources
In a technical article dated December 19, 2024, Anthropic sets out how to build systems in which AI does the work. Its appendix names four traits of tasks where AI adds the most value: the task needs both conversation and action, has clear success criteria, allows feedback loops, and builds in meaningful human oversight.
From the sources
Its examples are customer support and coding. In support, success can be measured clearly by whether the user considers the issue resolved. Code can be checked with automated tests, but the article says human review remains crucial to make sure it fits the broader requirements.
From the sources
The article opens with a note that much of the tooling it describes has changed since December 2024.
A&A perspective
A&A also reads these four traits as the conditions for a deliverable you can invoice. The article covers how to build the systems, not which work to sell. Still, when the finish line cannot be written in the buyer's own words, the "is this right?" check drags on after delivery, and the fixes tend to go unpaid.
A&A perspective
There is an opposing view: sell the work with the biggest pain and the highest price first. It is true that work that is merely easy to sign off can be cheap and solve nobody's problem. So A&A recommends this order: choose the area where the pain is largest, then carve out the part the buyer can sign off as the first service.
Hypothetical example
For example, take a new business for food manufacturers that uses AI to fill in product specification sheets, whose format differs for every trading partner. This is a hypothetical example. The buyer is the quality assurance manager, who decides both the order and the sign-off. The input for each case is a specification sheet sent by a trading partner. The list of product specifications is received once, at the start.
Hypothetical example
The output is the completed sheet plus a list of the items that could not be settled. The quality assurance manager can check it with the approval steps already in place. The final call on allergen labeling and legal compliance stays out of scope and remains with the client.
A&A perspective
The one page to show your manager is the six-row table below. If any row still falls in the "Not sellable yet" column, narrow the work one more step on that row.
| What to check | Sellable shape | Not sellable yet |
|---|---|---|
| Who buys | The same role decides both the order and the sign-off | Stalls every time at "I'll check with my boss" |
| What you receive | Material in the same format is already on the client's side | Starts with you collecting the material |
| What you return | A deliverable with a name, one per case | Can only be called "efficiency support" |
| How completion is checked | The client's current approval steps check it as is | Quality is argued out verbally |
| Decisions left to people | Final approval and exceptions stay with the client | Assumes everything can be handed over |
| Pricing unit | Invoiced per case, exceptions quoted separately | Only hourly estimates are possible |
Set sign-off terms first by laying out past cases
A&A perspective
Set the sign-off terms before delivery by laying out past cases, not by describing a quality standard in words. If the standard is set afterwards, the talk after delivery turns into an exchange of impressions: "the quality is low" against "it is good enough".
From the sources
Newfront, a North American company that handles business insurance, was looking for AI to answer benefits questions and had gathered 20 customer questions that earlier AI models struggled with. According to its CTO, without any changes to the system, Anthropic's AI handled 12 of the 20 correctly.
A&A perspective
The figure of 12 cannot serve as a target accuracy rate. Newfront chose the 20 questions itself, and neither the timing nor the kind of errors is public. It is also an account published by the company that sells the AI. What A&A borrows is the order, not the number. Before expanding, Newfront laid out hard real cases and decided first what would count as correct.
A&A perspective
A seller should follow the same order. Ask the client for about 20 past cases of the same kind, and at the proposal stage write down together how many of them must pass without rework. If many do not pass, narrow the scope or leave more decisions to people.
Hypothetical example
In the specification-sheet example, ask for 20 sheets submitted in the past and split them into those the trading partner sent back and those it did not. The sign-off terms then say that the 16 not sent back pass without rework, and that for the 4 sent back, the items behind each send-back appear on the list of unsettled items. The 16 and 4 are hypothetical numbers.
Hypothetical example
These 20 cases are also your material when you explain the plan to your manager. If you can say before delivery how the 4 sent-back cases will be handled, you can show where a failure stops.
Price per case and quote exceptions separately
A&A perspective
Price in the same per-case unit as the sign-off, and quote exceptions separately. If the deliverable has a name, you can invoice per case. If you can only estimate in hours, the unit you sell is not settled yet.
A&A perspective
Decide the exceptions before you start selling. The first case in a new format, a case with missing material, and a case that waits a long time on the client's decision are not covered by the normal per-case price; quote them separately.
From the sources
According to the same Anthropic article, several companies have shown that handing customer support to AI can work even with usage-based pricing that charges only for successful resolutions.
A&A perspective
A&A thinks charging only for results works in businesses with many cases. With many cases, the losses on hard ones even out across the whole. A new business that has just started has too few cases to even them out. Promising to "charge only when it works" leaves the client room to decide what counts as failure, and unpaid fixes pile up.
A&A perspective
So A&A recommends a form between an hourly estimate and paying only for results. A normal case has a fixed price, exceptions are quoted separately, and sign-off is checked against the cases agreed in advance. The client can predict the cost, and the seller can put a price on exceptions too. You can explain it to your manager in one line: "this much per case, this much for exceptions".
A&A perspective
How to fit the time spent checking and fixing into the economics of each case is covered in a separate article, "Low-ticket service profit rests on human touches, not AI speed".
Summary
A&A perspective
Narrow the first service by fixing one buyer, one input and one output.
A&A perspective
Pick work the buyer can sign off with their current approval steps.
A&A perspective
Set the sign-off terms before delivery by laying out about 20 past cases.
A&A perspective
Price per case and quote exceptions separately.
A new AI business should pick its first service by fixing one buyer, one input and one output, and choose work the buyer can sign off with their current approval steps. Set the sign-off terms from past cases before delivery, price per case, and quote exceptions separately.
FAQ
When should we add a second service?
Add it once the first service brings a second order from the same client and the reasons cases fail sign-off have settled into one kind. Anthropic's article also recommends starting AI systems in the simplest form and adding complexity only when it demonstrably improves results. A&A applies the same idea to the number of services you sell.
Does this work for a job that comes in only a few times a year?
Not well. You cannot lay out about 20 past cases, and you will not see a second order. The same goes for work whose input takes a different shape every time. In that case, treat the work of turning differently shaped material into one shape as a separate service.
What should we show our manager or the board?
Show the six-row table and how you split the 20 past cases used for sign-off. If asked where a failure stops, answer with how the sent-back cases are handled and the rule that exceptions are quoted separately.
Sources & editorial note
Primary pages read for this article. Publication dates below belong to the sources; access dates record our research.
- Building effective agents
Anthropic · 2024-12-19
Accessed 2026-10-22 - Lindy empowers teams to scale with AI Agents powered by Claude
Anthropic · n.d. (no date shown on page)
Accessed 2026-10-22 - Newfront modernizes insurance experiences with Claude
Anthropic · n.d. (no date shown on page)
Accessed 2026-10-22