Working with AI today is less about knowing a fixed list of services than about testing changing capabilities against a particular business. A useful practice keeps the business context and evaluation criteria in view, and adapts the workflow when a new tool actually improves it.
The real work begins after the first useful answer
A small company may already have several AI subscriptions. The owner discusses ideas with a model, a marketer prepares drafts and an administrator summarises correspondence. Sometimes this genuinely makes the day easier. Yet the question “What does the business gain?” can remain surprisingly difficult to answer.
A text appeared faster, but needed extensive editing. A polished report was produced, but the source of its figures is unclear. One employee found an effective approach, while everyone else struggles to reproduce the result. The benefit exists, but depends on individual habits and attention.
At this point, someone needs to connect the technology, the process and accountability. That person helps establish the intended outcome, which inputs are appropriate and how to distinguish a successful experiment from a sustainable improvement. The role can exist inside the company. An external partner is one option when internal time or experience is insufficient.
Editorial diagram · practical synthesis
- Business goal
- Workflow
- Data and tools
- Team
- Quality and outcome
Why a general-purpose tool needs specific context
With familiar software, the boundaries are usually clear: accounting software maintains records; a calendar stores appointments. Generative AI works differently. One interface supports strategy discussions, document analysis, writing and prototyping. That flexibility is attractive, but it does not explain which application will help a particular company.
Two salons may look similar to a model. For their owners, the differences determine almost everything: services, staff availability, deposit policies, customer languages, administrator coverage and booking arrangements. Generic marketing advice can be reasonable in principle and unhelpful in this particular business.
Implementation therefore starts with questions. Where does the team lose time? At which step do errors occur? What knowledge exists only in the owner's head? What counts as a good result? The answers become working context: approved facts, examples, constraints and evaluation criteria. That context remains useful even when the model changes.
What research shows and what it does not promise
Dell’Acqua and colleagues' experiment, published in March 2026, involved 758 BCG consultants. On suitable tasks, participants using GPT-4 worked faster and completed more tasks. On a deliberately selected task outside the model's capabilities, however, the share of correct solutions was around 19 percentage points lower on average. The researchers describe a jagged technological frontier. Organization Science, 2026.
Research · Dell’Acqua and colleagues, 2026
The experiment involved a specific model, participants and assignments, including one selected difficult task. It does not measure every product's capabilities today or establish a return on hiring a consultant. Published paper.
A cautious business takeaway is that AI's ability to answer persuasively and its ability to support correct decisions need separate evaluation.
A specialist can make mistakes too. Their value becomes visible when they can identify unsuccessful applications, explain limitations and change the approach in response to testing. An impressive demonstration without an examination of weak points provides too little basis for implementation.
One business and four different conversations about AI
Consider a fictional company: two hair salons, one owner, administrators and several stylists. This is an illustrative example, not a client case. No achieved results or revenue are attributed to it.
In marketing, the team might prepare alternative posts from an approved service description. Accurate promises, tone and actual appointment availability matter. In finance, AI might help explain an already verified report, while source amounts, formulas and final decisions remain under a responsible person's control.
In operations, the business could test preparation of an opening-shift checklist or a summary of recurring difficulties. In support, it could start with draft answers about service duration and rescheduling policies. Exceptions, refunds and discussions involving sensitive information need a separate human decision.
An implementation partner does not need to replace the marketer, accountant and manager personally. They need to organise a conversation with each: understand the process, choose a bounded task and agree who checks the output. Attempting to automate all four functions immediately risks producing several unfinished experiments rather than one understandable improvement.
What the person leading implementation owns
It helps to assign four kinds of work. This is a practical recommendation, not a scientifically established universal organisational model.
The first concerns selection. The owner of the initiative helps weigh task frequency, the cost of errors, data availability and evaluation difficulty. They allow for the possibility that a template, a revised instruction or an existing software feature will be the best solution.
The second concerns experimentation. The team needs real examples, a description of the current process, tests of different approaches and a record of failures. Evaluation covers the entire path to usable output, including clarification, corrections and approval.
The third concerns handover. The company receives understandable materials, access through its own accounts, documented limitations and the ability to continue independently. If only the creator can operate the system, its practical resilience is limited.
The fourth concerns review. Business offers, data sources and tools change. Someone needs to notice when an old solution no longer fits and test updates before they affect customers. This can be a small recurring responsibility rather than a new full-time position.
When you can comfortably start without an external specialist
Outside help is often unnecessary when a task is bounded, sufficiently frequent and well understood by its performer. A marketer might turn approved copy into several short advertising variations. The facts are available, there is no confidential data, a person checks publication and an unsuccessful version is easy to discard.
Good conditions for an independent start include time to test, an output that can be evaluated without complex expertise, a reversible action and no need to connect several systems. Decide beforehand how much effort the team can spend experimenting and what would justify stopping.
Sometimes a short internal review is enough: compare several examples, record useful instructions and list common mistakes. At low volumes, integration can cost more than copying information manually. Repetition alone does not make an operation worth automating.
Organisational readiness also matters. If nobody can make time for review, an external person will struggle to establish a dependable process. If policies repeatedly contradict each other, a management decision comes first. AI cannot decide on the owner's behalf which promises the company is prepared to honour.
When a partner helps reduce uncertainty
An external specialist becomes more useful when a task crosses functions with different definitions of success. Marketing wants more enquiries, the administrator wants fewer repeated questions and the owner wants a clear view of capacity. Without coordination, faster advertising production can increase support workload.
Another reason to seek help is that the team already experiments but cannot compare solutions. One service produces more attractive text, another connects more easily to a knowledge base and a third costs less. Someone needs to translate those differences into process requirements and test the meaningful alternatives.
Help may also be useful for organising data access, integrations, training or full-cost evaluation. However, a broad implementation assignment does not make every AI specialist an expert in security, law or accounting. Where qualified domain expertise is required, it should be involved separately.
Choose the engagement around the uncertainty. Sometimes a discussion and an experiment plan are sufficient. Sometimes a bounded pilot with a proper handover is needed. An ongoing partnership makes sense when there are several successive tasks and a genuine need to revisit the system regularly.
Why adaptability matters more than a list of services
New AI capabilities can make starting over feel tempting. But a business does not need to follow every release. It is more useful to maintain its own task set and a clear way to test whether something important has changed for those tasks.
For example, the salon needs accurate draft replies about rescheduling. While its tool performs adequately, switching providers for a more impressive demonstration may not make sense. If a new capability promises dependable use of changing policies across both locations, test it against existing examples, including conflicting conditions and incomplete requests.
In a resilient system, business rules and source materials live separately from an individual chat. The team can access test examples. A manual fallback exists. Replacing a tool then becomes a bounded change rather than a loss of accumulated knowledge.
A person who follows new capabilities is useful precisely because they are selective. They can distinguish a meaningful development from an interesting novelty and do not promise that today's choice will remain the best forever.
What to discuss before working together
Ask the specialist to explain the proposed first step using your process. What data will they need? What will evaluation involve? What will the team retain if the pilot does not deliver the expected benefit? Can work continue without their constant involvement?
Agree access and action boundaries beforehand. Preparing a message, sending it to a customer and changing a booking carry different responsibilities. Financial actions, sensitive information and consequential promises require explicit oversight. Having a person somewhere in the process is insufficient: they need time, sources and authority to stop an action.
A proposal should identify concrete outputs: a description of the selected task, test cases, a working process, review rules and knowledge transfer. Clarify maintenance costs, account ownership, data export options and how the engagement can end.
Warning signs include identical savings promises for every business, pressure to buy a bundle of subscriptions immediately or reluctance to discuss errors. A good conversation leaves room for the conclusion that existing software is sufficient for now.
Where a small-business owner can start
Choose one task that genuinely obstructs work. Write down how it is performed today, who accepts the result and what makes an error harmful. Then decide whether someone internally can run a small, verifiable experiment.
If that person exists, give them time and clear authority. If experience, process complexity or limited capacity stands in the way, discuss outside help with a bounded scope. A useful first meeting may produce a clear decision about what to test, what to leave unchanged for now and how to recognise improvement.
Work with Dima Dart can begin with a 90-minute consultation for $250. Implementation partnerships are agreed separately according to the task. The purpose of the first step is clarity about the next decision, without a promise of universal returns.
Further reading
The results of Dell’Acqua and colleagues apply to the tasks and conditions studied. The partner-selection criteria and working arrangements above are editorial recommendations, not research findings establishing a need for consultants.
Continue with “Which business tasks should you delegate to AI?” and “How to build an adaptable AI workflow and evaluate a pilot”.