A Cheaper Model Won’t Fix a Messy Workflow

For smaller firms, the most useful AI increasingly finishes a narrow workflow instead of producing another draft. The gain shows up as hours returned, fewer handoffs, lower operating cost and less routine work for the person who still owns the outcome.

This week produced plenty of impressive numbers, from cheaper inference to 100,000 farmers using GeoAI. The more useful pattern sat underneath them: software is getting better at carrying work from one system into the next. For a small company, that matters because the expensive part of a task is often the repeated copying, checking, chasing and moving that happens after the first answer appears.

The model is the cheap bit

DeepSeek's V4-Flash was reported at $0.14 per million input tokens and $0.28 per million output tokens, with benchmark inference costing about three cents per test.¹ Those figures make sophisticated model access increasingly affordable, including for companies that could never justify a large internal AI budget. The barrier to trying a capable model is becoming almost trivial. But the price of inference says almost nothing about whether a business will actually save money after the model is placed inside a real workflow.

A restaurant owner does not buy tokens. They buy some version of fewer late nights writing posts, fewer missed enquiries, quicker menu updates or less time spent jumping between software. A local retailer does not care whether a model is cheaper if someone still has to copy the output into a spreadsheet, check stock in another system, rewrite it in the brand's tone and then remember to publish it. The cost that matters is the cost of finishing the job.

That distinction is becoming more important as the surrounding obligations grow. From 2 August, European AI transparency requirements began applying to certain AI interactions and generated or altered content, including disclosure and machine-readable marking requirements in relevant cases.² Google, meanwhile, has stopped allowing developers to create new Smart Campaigns through the Google Ads API, another reminder that a workflow can change because a platform changes an interface or removes a capability.³ The model may get cheaper while the work required around it becomes more demanding.

Cheaper models therefore create an odd effect for small firms. They make experimentation easier while making the remaining operational work more visible. Once generation costs pennies, the expensive parts are business context, permissions, review, integration, compliance, publishing and support. A five-minute task that still needs six manual handoffs can remain a five-minute task with a cheaper model underneath it.

This is why "AI for small business" is becoming a less useful category than it first appeared. The interesting question is which recurring job has been shortened from beginning to end. A model that drafts quickly but leaves the owner to finish the process has delivered partial automation. A narrower tool that completes one routine reliably may create more value with much less intelligence on display.

Handoffs are where time leaks

Outernet offers a simple example. It can turn a saved social post about a restaurant, place or event into something operational, including a date, location, map and calendar entry.⁴ The important step is not that software understood the post. It is that the saved item moved into the calendar where the user can act on it.

The same logic appeared in meeting tools, retail software and local-business platforms throughout the week. Thryv launched an AI-native growth platform while announcing planned connections with Wix and Ooma, bringing websites, communications, customer records and commercial activity closer together.⁵ The company reported that SaaS represented 76% of revenue and that monthly average revenue per user had risen to $394. Those numbers suggest businesses are willing to pay when software takes responsibility for a broader slice of the workflow rather than another isolated feature.

There is a reason this matters more to a twelve-person company than to a large enterprise. Big organisations can build integration teams, write middleware, hire analysts and tolerate a surprising amount of process friction. Small firms usually make one person absorb that friction. The owner, manager or marketer becomes the human API between the website, inbox, calendar, social account, payments system and customer database.

That is also why a saved hour can be worth more than a clever answer. SCORE's recent small-business workshop focused on using low-cost AI tools for competitor research, customer targeting and routine production work, with the wider promise of giving owners several hours back from tasks such as emails, social content, research and prospecting.⁶Returning three useful hours is a clearer product claim than adding another model option. An owner can decide what three hours are worth.

This is the same design test we use when thinking about AI tools for small-business Instagram work. A caption appearing in a chat box is only the first step. Instagram for small business works better when the workflow starts with the business's own photos and brand cues, prepares content for review, lets a person change it and moves approved posts towards the publishing calendar. The point is to remove repetitive production work without removing the owner's voice from what customers see.

That distinction also changes how small businesses should evaluate software. Ask where the task starts, where it ends and how many times a person has to move the work between those points. A product may save ten minutes in the middle while adding five minutes of checking at the start and another five minutes of administration at the end. A completed workflow is measurable; a promising feature is not.

Narrow systems earn trust

Several of the week's strongest examples were specialised. Delightree, which works with franchise and multi-unit operators, said it is used across more than 6,000 locations and has grown revenue nearly twentyfold before raising $25 million.⁷ Its pitch is tied to training, compliance, audits, communications and opening new locations, all jobs with clear inputs, repeated steps and visible consequences when something goes wrong. That gives buyers something concrete to judge before they ever ask which model sits underneath.

Focal Systems offered an even more concrete case. Its Top Stock product uses existing store cameras to identify inventory sitting above shelves and turn those observations into replenishment tasks. In a one-month pilot with Village Supermarkets, the company reported finding $24,000 in previously unaccounted inventory, cutting weekly audit time from four hours to one and saving the night crew ten hours each week.⁸ Those are operational numbers a store manager can test against their own costs.

That result is useful because it does not require anyone to believe in a general theory of automation. A retailer can measure the missing inventory, the staff time and the effect on replenishment. The system takes a narrow, unpleasant piece of work and makes it cheaper to perform. The employee is still needed to decide what to do with the information and to deal with the exceptions the camera cannot resolve.

This is a better model for adoption than asking a busy owner to invent an AI strategy. Start with a repeated task where better context changes an outcome, then keep responsibility with the person who understands the business. That could be a buyer reviewing a forecast, a restaurant manager approving a week's content, a franchise operator responding to a compliance issue or a consultant turning meeting notes into named actions.

The approach also protects against a problem that becomes more obvious as AI touches customer-facing work. Generic automation can save minutes while making every business sound as if it hired the same invisible copywriter. For Instagram AI content, the source material matters: the actual products, photographs, menu items, tone, offers and customer questions are what keep the output connected to the business. Speed is useful only when recognisability survives it.

For restaurants, that is especially visible. A system can help with Instagram content planning for restaurants, but the finished posts still need to look and sound like that restaurant rather than a generic food account. The chef's dishes, the room, the staff, the humour and the particular reasons regulars return are not inefficiencies to automate away. They are the raw material the software should help the business use more consistently.

Agents reward legible businesses

The week's commerce stories extend the argument beyond productivity. Shopify said AI-driven traffic and orders to stores had tripled year on year, with half of AI-driven sessions landing directly on a product description page and 75% of AI-attributed purchases falling outside the platform's top 100 categories.⁹ For smaller sellers, that suggests a route to discovery based less on fame and more on whether a product page answers a specific request clearly. A niche product can have an advantage when the request itself is niche.

Google Maps is pushing in the same direction for local businesses. Ask Maps can now move from finding a suitable restaurant to initiating an order through services such as Square, Toast or Uber Eats, while also supporting hotel bookings and event discovery.¹⁰ A customer can travel further through the purchase journey without visiting the business's own website or speaking to a member of staff. The assistant increasingly decides which information deserves to become an action.

That gives small firms access to powerful distribution, but it also raises the price of vague or stale information. An agent needs dependable product names, opening hours, menus, dietary details, prices, availability, locations and policies. When those details conflict across a website, listing and social profile, the business becomes harder for software to recommend with confidence. A neglected description can now affect more than search ranking because it can influence whether an agent completes the next step at all.

Hark's browser agent points to an even broader version of this behaviour. The product aims to navigate shopping and booking sites without relying on official APIs, handling tasks such as purchases, restaurant reservations and returns.¹¹ Small businesses may soon receive more visitors that are not people browsing pages but software trying to determine whether an action can be completed safely and correctly. That visitor has less patience for ambiguity than a loyal customer who already knows how the business works.

This creates a practical job that many owners will initially mistake for "SEO". It is really business legibility. Prices need to match. Policies need to be findable. Product descriptions need to answer real questions. Stock and opening hours need to be current. The website needs to make it obvious what can be bought, booked or returned, and under which conditions.

A second issue matters equally. If an assistant finds the restaurant, chooses the meal and hands the customer to a payment service, the transaction may succeed while the business relationship weakens. The customer remembers the convenience of the assistant, and the restaurant risks becoming a fulfilment layer behind someone else's interface. Being chosen by the agent and being remembered by the customer are different wins.

That is where a small business Instagram strategy still matters even as agents become better at completing transactions. Social content gives the business a place to show the human texture that product feeds and structured listings struggle to carry: the baker pulling a tray from the oven, the stylist explaining a cut, the owner introducing a new line, the regular customer everyone knows by name. Machine-readable accuracy gets the business into the shortlist. A recognisable identity gives the customer a reason to come back directly.

Count the work that disappears

The common lesson from cheap inference, connected platforms, specialist tools and agentic commerce is surprisingly practical. Small businesses should stop measuring AI adoption by the number of tools they have tried or the sophistication of the model behind them. Measure the work that no longer has to be moved, copied, chased or reconstructed by a person.

That measurement can stay simple. Pick one recurring task and record how long it takes today, including the hidden steps around it. Then test whether the tool reduces the full cycle without creating new checking, compliance or clean-up work somewhere else. If the owner gets three hours back, a stock audit drops from four hours to one, or an enquiry moves into the right system without manual copying, there is a result worth keeping.

The same test also protects human judgement. The person responsible for the customer, the brand or the operational decision should still have a clear point of review where mistakes matter. AI can prepare, sort, summarise, detect and move work forward. Accountability should remain visible rather than disappearing into the workflow with the admin.

The best systems for small firms may therefore look less impressive in a demo than the broadest assistants. They will know one job, use the business's real context, connect to the next step and stop at the point where a person should decide. That is not a smaller ambition. It is how software earns a permanent place in a week that was already full before AI arrived.

Sources