The Prompt Is Getting Shorter

The prompt is getting shorter because the software increasingly knows the business before the owner asks a question. The useful tools this week already had access to menus, company records, books, inboxes, invoices, customer data or catalogues, which meant less explaining before useful work could begin.

That is a more important shift than another jump in model capability. Small businesses do not have spare staff whose job is to brief an assistant, clean up its output and move the result into another system. The better products this week reduced that briefing work first, then automated a narrow task around information the business already had.

Context arrives first

LegalZoom's new AI Assistant can answer questions using a customer's actual business record, including business type, formation state, registered agent and filing status. When a question needs professional judgement, it can route the owner to an attorney or specialist.¹ The interesting design choice is not the chat interface. It is that the assistant starts with the company record already in view.

Menuthere is applying the same idea to restaurants. An owner can start with a restaurant name, upload or photograph a menu, then create an online ordering setup through conversation rather than learning a fresh dashboard.² The menu carries a surprising amount of the burden because it already contains products, prices, descriptions and the structure of the offer.

Collective makes the same argument from the back office. It raised $100 million for an AI-native financial platform aimed at solo businesses, pulling bookkeeping, payroll, tax, invoicing and related work into one system.³ A solo operator does not have a finance team available to connect those jobs manually, so the product becomes more useful when the books and business structure are already part of the working context.

This changes what prompting means. A blank assistant forces the owner to explain the business every time the task changes, while a context-rich product can move towards requests such as "show me what changed", "prepare the response" or "find the problem". The prompt becomes shorter because the setup has moved into the product.

For AI for small business, that is a useful buying test. Before asking how clever the model is, ask how much of your business it will understand before you type the first useful instruction. If the answer is very little, the owner may still be doing a large part of the work.

Most work happens before send

Hostinger's analysis of roughly 213 million Agentic Mail actions gives a good picture of what useful automation actually looks like. Only 2.5% of those actions were sends. Most involved searching, listing and reading messages, and Hostinger estimated roughly 18 information-gathering actions for every email sent.⁴

That matters because the visible action is often the smallest part of the job. Sending an email takes seconds, while finding the right thread, checking the attachment, locating the customer history, comparing dates and understanding what happened can take much longer. The software earns its keep when it shortens that preparation.

Intuit's Agentic Books Close follows the same pattern. It can run more of the monthly close using a firm's own procedures while accountants supervise the process, approve results and deal with exceptions.⁵ Black Ore's Tax Autopilot prepares returns and links figures back to source documents, with licensed professionals retaining final sign-off.⁶

These examples are less dramatic than a fully autonomous agent, but that is precisely why they fit smaller companies. The owner does not need software to perform autonomy for its own sake. They need fewer searches, fewer copied figures, fewer tabs and fewer pieces of unfinished work waiting for a person to stitch them together.

There is a broader product lesson here. The valuable part of automation often sits one step before the obvious action. If an AI system can gather the evidence, organise the records and bring a nearly finished piece of work to the person responsible, the time saving can be substantial without pretending the software should own the final decision.

The café should not become the fraud team

A Dubai café provided the uncomfortable counterexample this week. Taho Avenue said a customer used an AI-generated image of spilled drinks to support a delivery refund claim, and the café identified inconsistencies in cup sleeves, lids and ingredients.⁷ Generating a plausible complaint can now be quick, but disproving it can still cost the business time, attention and the value of the order.

That creates a new asymmetry for small firms. A platform may be able to compare order records, customer history, timestamps, image signals and patterns across thousands of claims. An individual café cannot. If the platform accepts synthetic evidence too easily, the verification cost lands on the party least equipped to absorb it.

This is where a generic call for "human review" is too easy. Someone still has to decide a disputed case, but the useful question is who should do the expensive checking before that decision reaches a person. The café owner should not become the fraud analyst simply because AI made the fake evidence cheaper to produce.

Supply'd offers a better example of where software can do more of the preparation itself. Its Expense Agent can investigate supplier-invoice discrepancies and recommend corrections, tax changes, price updates or credit requests inside the ERP.⁸ The system does the digging first, which means the business is reviewing a prepared recommendation rather than beginning the investigation from zero.

The same principle applies to customer service, payments and content moderation. Asking a human to check everything can preserve accountability while destroying the time saving. Good product design reduces the number of cases that need attention, gathers the relevant evidence when one does, and leaves a clear reason for the person who has to decide.

Existing software has the advantage

Several launches this week suggest that the strongest position in small-business AI may belong to software that already sits inside the workflow. Shopify launched Shopify Protect to bring fraud prevention and dispute tools into Shopify Payments. It says a $100 chargeback can cost a merchant about $238 once lost product, fees and staff time are counted, so preparing evidence inside the payments system addresses a concrete expense rather than creating another AI destination.⁹

Virid is approaching the same problem from a different direction. Its new MCP connector lets merchants use compatible assistants to search products, customers and orders or manage catalogue information and promotions, while some sensitive information remains read-only and proposed changes can be reviewed.¹⁰ The assistant can vary, but the useful context remains attached to the merchant's actual store data.

The appeal is straightforward: the business does not have to build a separate AI operating layer before seeing value. The intelligence arrives inside the system where the relevant records and recurring tasks already live. For a five-person company, avoiding that connection work can be the difference between trying a feature and abandoning it.

That distribution advantage is easy to underestimate. A restaurant already has reservation software. A retailer already has a catalogue. A small employer already has payroll. The provider that understands those records can add intelligence at the point where the work happens, while a standalone assistant must first ask the owner to recreate or connect the context.

The difference shows up in setup time. A general assistant may be able to perform a task once the owner has supplied the right files, copied in the relevant rules and explained what counts as a good result. Software sitting on the underlying business system can skip much of that ceremony. That makes small improvements more economical because the owner does not have to pay the setup cost again for every routine task.

It also changes where product quality shows up. Better answers still matter, but so do mundane details such as knowing which customer record is current, which product is out of stock, which invoice version is final and which photo is approved for use. A small business often experiences those details as the difference between software that saves ten minutes and software that creates ten minutes of checking.

There is still a trade-off. The more context a system holds, the more important it becomes that the business can understand permissions, export its data and see what changed. Deep integration is useful because it removes handoffs, but it becomes a problem if switching providers means rebuilding the business memory from scratch.

This matters for Instagram for small business too. A useful Instagram AI content tool should know the real products, approved photos, offers, opening details and brand context before it drafts a post. Instagram content planning becomes easier when the software works from material the business already owns, rather than asking for another generic prompt every time.

One job should disappear

Oracle's new AI SMB Bundle shows how much capability vendors are now willing to package for smaller companies, including cloud credits, ChatGPT and Codex benefits, NetSuite QuickStart and Oracle App Composer.¹¹ Bundling may lower the cost of access, but a broad package still has to prove that it removes work rather than adding configuration. The value will be judged by which recurring task disappears first.

The practical workshops running alongside these product launches make that point more clearly. A Boston event with Grow with Google and the Massachusetts AI Hub focused on prompting, idea generation and repetitive-task automation, with attendees working towards a personalised plan for their own business.¹² The emphasis on an owner's actual work is more useful than teaching a long catalogue of prompts.

That is a healthier adoption model than measuring progress by how many tools a business has tried. The useful unit is a recurring job: this invoice check, this inbox search, this product-description draft, this weekly content plan. If one of those jobs becomes materially shorter without producing a new queue of cleanup work, the software has earned its place.

For someone asking how to automate Instagram content creation, the same standard applies. The target might be turning one product shoot into a week of drafts, preparing AI captions for Instagram business posts from approved product details, or building a small business Instagram strategy around media already in the library. The goal is not to generate more content. It is to remove repeated setup while keeping the output recognisably tied to the business.

The most useful pattern this week was therefore not "AI can do more". It was that the owner had to explain less. The software knew the menu, the company record, the books, the inbox, the invoice or the catalogue before the real task began, and that made each subsequent interaction cheaper.

That also gives owners a simple way to evaluate new products. Ask what information the tool already has, what information you will have to keep supplying, and whether that context survives from one task to the next. A system that forgets the business after every interaction may still be capable, but the labour of repeatedly briefing it belongs in the cost calculation.

The same question helps separate useful automation from novelty. If a tool needs a long prompt to do a familiar job every week, the product may still be too far away from the workflow. If the instruction can shrink to a few words because the system already knows the relevant records and constraints, the software is starting to behave more like part of the business.

That is a demanding standard for product makers because it shifts attention away from the demo. A clever answer in a blank chat box is easy to show. A product that quietly knows enough about the business to save twenty minutes every day is harder to dramatise, but much easier for an owner to keep using.

Sources