Delivery Hero’s 15% Lift Came From Knowing the Job

Delivery Hero says restaurants and shops using its agentic assistant increased customer orders by 15%. The useful lesson is not the agent label. The software is attached to familiar jobs, supplied with business context and judged by an outcome an owner already cares about.¹

This week, that same pattern turned up in legal software, restaurants, beauty, retail and business finance. General AI is already easy to try. The stronger products are beginning with a narrower question: what recurring piece of work can this system understand well enough to remove, shorten or improve?

The 92% number misleads

A Small Firms Association survey reported that 92% of respondents use generic tools such as ChatGPT or Copilot, while only 30% use customised models tailored to their business needs. Only 18% said AI had been incorporated into most or some business practices.² Those figures describe three very different behaviours that often get bundled together under the word adoption. A business can therefore appear AI-active while still having almost none of its routine work changed by AI.

Opening ChatGPT to rewrite an email is usage. Giving a system access to approved business information, fitting it into a repeatable process and deciding who reviews the output is operational adoption. The distance between those two states is where most of the useful work now sits.

The UK numbers tell a similar story. Novuna Business Finance found that 31% of small firms were already using AI, with London at 41%, while 27% nationally said they did not plan to use it and 21% did not see it as relevant.³ That does not look like a market waiting for more demonstrations of what a model can do. It looks like a market waiting for products that make the relevance obvious.

For a small firm, the cost of experimentation is not mainly the subscription fee. It is the owner’s attention, the staff time needed to correct outputs, the risk of feeding a tool the wrong information and the extra step created when the software sits outside the work. AI for small business earns trust when it reduces that burden rather than asking the business to become better at prompting.

That distinction matters because smaller teams feel process friction immediately. A ten-person company has fewer layers between a repetitive task and the person paying for the time it consumes. When a tool removes twenty minutes from a job that happens five times a week, the value is visible without a strategy deck.

The job comes first

Actionstep’s new intelligence layer for midsize law firms is built inside the practice-management software firms already use. It works with approved firm data and workflows, can recommend or carry out approved operational and financial tasks, and keeps human oversight and audit trails in the process.⁴ That is a very different product decision from adding a chatbot window beside the existing software. The user stays in the system where the matter, permissions and history already live.

MoxiWorks and Cloze are attacking a similarly ordinary source of waste in estate agency: duplicate contact and activity data spread across systems. Their integration creates a shared source for client information so agents do not keep re-entering the same details before follow-up and marketing can happen.⁵ Nobody needs to be persuaded that duplicate entry is undesirable. The product starts with a nuisance the user already recognises.

Hospitality is following the same route. me&u says venue operators commonly work across five to seven technology partners around guest data, and its newer platform brings reservations, voice AI for phone bookings, functions and events into one workflow.⁶ Hungry Jack’s, across close to 500 restaurants, is also moving away from spreadsheet-heavy reconciliation for roughly 300 million transactions a year so software can match routine items and people can deal with exceptions.⁷ In both cases, the gain comes from reducing handoffs between systems rather than making each individual system more conversational.

These examples make a useful point for smaller operators. A well-defined job is more valuable than a long capability list. The law firm cares about a matter progressing, the estate agent cares about a client record being current, and the restaurant cares about a booking or transaction being handled correctly.

That is also why vertical software has an advantage over a blank prompt box. It already has an opinion about which fields matter, what sequence the work follows and where a mistake becomes expensive. Domain context turns a model from a clever generalist into something closer to a trained assistant with a specific remit.

Fifteen per cent is the proof

Delivery Hero gives this argument a number. More than 40,000 restaurant and shop partners were using its agentic assistant, and the company said those partners increased customer orders by 15%.¹ The assistant handles jobs such as responding to reviews, improving menus and product descriptions, setting up promotions and helping manage advertising. Those are activities that owners already connect to demand, which makes the reported lift easier to interpret than a generic productivity score.

Each task sits close to revenue. A better menu description can affect conversion. A promotion can change demand. A review response can influence how a prospective customer interprets the business. The useful measure is commercial work with an outcome. Impressive answers are secondary.

That matters for Instagram marketing for restaurants too. An owner thinking about Instagram for small business rarely needs a generic explanation of how captions work. They need Instagram content planning that starts from their actual dishes, opening hours, offers, photos and tone, then produces something accurate enough to review quickly and publish.

This is the logic behind tools such as Asteris for restaurants, where the useful unit is not a prompt but a repeatable content job grounded in the business’s own material. The same principle applies to Instagram AI content more broadly. If the tool knows nothing about the brand until someone explains it again each time, the business is doing part of the software’s job.

The more specific the workflow becomes, the easier it is to decide whether the tool is helping. Orders rose or they did not. Time spent preparing posts fell or it did not. Fewer enquiries were lost, fewer records were duplicated, fewer transactions needed manual matching. Smaller businesses benefit from that clarity because they cannot afford indefinite experimentation with vague productivity claims.

Keep the person in the work

Zenoti’s new predictive agents show another side of the same shift. The company says its systems can identify beauty and wellness clients at risk of not returning, forecast staffing needs and predict inventory demand across more than 30,000 businesses.⁸ Those predictions can save time, but the final action still carries human context. A risk score becomes useful only when somebody who understands the client decides what response is appropriate.

A salon manager knows why a long-standing client cancelled twice. A stylist may know that a customer’s spending fell because she moved farther away, not because she disliked the service. Software can surface a pattern, while the person closest to the customer decides whether a discount, a phone call or no action at all makes sense.

Tulip is explicit about that division of labour in retail. Its AI products give store associates customer summaries, preferences, recommendations and suggested next actions while leaving the employee in charge of the interaction.⁹ For smaller retailers, that is an important design choice because personal service is often part of the reason customers choose them over a larger chain. Better information can strengthen that service without asking software to impersonate the relationship.

Replacing that interaction with automation can destroy the thing the business sells indirectly: recognition, flexibility and judgement. Removing the repetitive work around the interaction has a different effect. The employee has more context and more time to use it.

This is where the founding argument for useful automation becomes practical. AI should increase the amount of judgement a person can apply, not reduce every task to a machine decision. A prediction can narrow attention. A draft can remove the blank page. A reconciled set of transactions can leave a finance team with the exceptions that actually require thought.

Small firms buy relief

The finance stories this week make the economics even clearer. Bluevine surveyed more than 700 US small-business decision-makers and found that, among owners who added or switched to a fintech platform, 74% said slow or outdated traditional banking had caused at least one major financial problem in the previous year.¹⁰ Fifteen per cent reported delaying staff payments and 17% delayed vendors or contractors. In a small company, a slow process can therefore become a cash problem rather than remaining an administrative annoyance.

Those are not abstract inefficiencies. They hit payroll, supplier relationships and cash confidence. A small business with limited buffer feels a payment delay in a way a large company with dedicated treasury staff may not.

Dream Payments and J.P. Morgan Payments launched Dream Payouts with real-time payment capabilities for eligible transactions and a design that could support authorised software agents initiating, approving and reconciling payments.¹¹ 9Spokes launched Pulse to combine consented banking, accounting, merchant, payroll and marketing data so businesses can ask practical questions about bills, location performance and cash position.¹² Both products move automation closer to a decision the owner already makes. That proximity raises the value of the software and also raises the standard for permissions, review and traceability.

The useful standard here is relief. Does the software remove waiting, duplicate checking or uncertainty from a recurring part of the week? Does it compress the path from information to action without obscuring who approved the action?

For a small business, that standard is more demanding than "can the model do this?" Many models can draft a payment reminder, write a caption or summarise a spreadsheet. The product earns its place when the owner no longer has to assemble the context from scratch every time. Repetition is where convenience turns into genuine operating value.

What should the first job be?

The best first workflow has four characteristics. It happens often, its inputs can be clearly defined, a good outcome is easy to recognise, and there is an obvious point where a person can review or intervene. That can be invoice follow-up, appointment reminders, review responses, content preparation, stock checks or client intake.

A small business Instagram strategy fits this pattern when there is already real material to work from. Photos, products, services, offers and brand language provide the inputs. A tool can prepare AI captions for Instagram business posts or help automate Instagram content creation, while the owner still decides what deserves to be published.

The same is true of AI tools for small business content when they are grounded in the business rather than used as a generic copy machine. The aim should be recognisable output with less repetitive effort. That is how to stay on brand with AI content without turning every post into a fresh prompt-engineering exercise.

This gives a practical answer to the question, "What is the best AI tool for Instagram marketing for small businesses?" It is the one that already understands enough of the business to reduce work safely, keeps the source material visible and gives the owner a fast review point before anything reaches a customer. Feature count matters less than how little context the user has to reconstruct. A tool that saves ten minutes every day can be more useful than one with fifty impressive functions that never settle into a routine.

The same test answers, "How can a small business save time using AI for Instagram?" Start with one recurring production step, such as turning approved photos into a week of draft posts, and measure whether review time genuinely falls. If staff are correcting invented details, rebuilding the brand voice or moving copy manually between several tools, the workflow has not earned its keep. Time saved after review is a better measure than drafts generated before review.

A blank prompt is a tax

The week’s stories point towards a stricter definition of useful software. Delivery Hero attaches its assistant to reviews, menus, promotions and advertising. Actionstep starts from legal workflows. Zenoti begins with retention, staffing and stock. Bluevine’s survey shows why owners care about the operational consequences when systems are slow or fragmented.

There is a cost when every use begins with an empty box. Someone has to restate the business, locate the right file, explain the customer, define the tone, check the facts and decide what happens next. That effort is easy to ignore during a demo and impossible to ignore when it repeats every Tuesday.

The stronger products are absorbing more of that preparation. They know the nouns of the business, the data they are allowed to use and the boundaries around the action. The person remains responsible for the decisions that touch reputation, money or relationships, but spends less time rebuilding context for the machine.

For small firms, that may be the most useful way to think about adoption now. Do not count tools. Count recurring jobs that became easier, faster or more dependable without making the business sound less like itself or leaving nobody clearly accountable for the result.

Sources