Polish AI harness. How to speed up work and keep the team’s skills
Series · part 2 of 3: people
Author: Andrzej Okoński, Head of Software Development
Deploying AI changes the division of work, the way people learn and relationships within the team. A good harness should give employees influence over these changes: the ability to check answers, question recommendations and develop their own skills. Its quality is also shown by how the organisation treats the people whose tasks are taken over by technology.
In the first part of the series, we described a salesperson preparing an offer with the help of AI. Let us return to this hypothetical example. The assistant finds the price list, organises the arrangements and creates a draft. The salesperson finishes the task faster. After a few weeks, however, the manager raises the expected number of offers, the time for checking the content disappears, and a junior member of the team starts approving proposals they cannot explain on their own.
A technical measurement may still show success. The organisation should also notice the growing dependence on the system and the burden on the people who are responsible for its results. That is why the design of a harness also covers the way work is managed.
Leave assessment and decisions to people
The harness sets the boundaries of what technology may do: access to data, independence and the ability to perform operations. The assistant helps to process information, while employees talk, compare materials and direct the work. People keep room for their own judgement and for making decisions.
Agency has a very practical meaning here. Employees know when they are using AI, understand their role and can refuse to approve a wrong proposal. They can also report a problem without fear of being seen as opponents of progress. These conditions need to be built into procedures, training and managers’ objectives.
Anxiety about employment, frustration with technology, withdrawal from relationships or reduced cognitive effort are different phenomena, and their link with technology also depends on working conditions. The WHO lists, among other things, job insecurity, excessive workload and low control over work as risks to mental health. A well-designed AI deployment can reduce these risks.
Talk honestly about the future of work
The promise that “AI will free you from routine” leaves the most important question unanswered: what will happen to the people whose work consisted mainly of that routine? The management board should explain which tasks it wants to change, which decisions it has already taken and which issues remain open.
A joint study by the ILO and NASK describes the exposure of occupations to generative AI, that is, the potential for tasks to change. Exposure alone is not a forecast of the loss of a specific number of jobs. The outcome also depends on how AI is deployed and on the possibilities of adapting skills. This is an important argument for analysing the real process in the company instead of applying labour market headlines to the fate of a particular team.
We propose that every pilot be accompanied by a work change plan. It should set out the tasks taken over by the system, people’s new duties, the time set aside for learning and how the workload will be assessed. In our example, part of the time saved can be spent on conversations with clients and on preparing for more difficult negotiations. Such a decision, however, has to be taken and included in the team’s plan.
If job reductions are a real scenario, this must be said honestly, and the people affected by the change must be given concrete help in moving to a new role.
Reduce frustration through good work design
A poorly matched assistant that sometimes answers accurately and at other times fails to recognise a document takes up attention instead of saving time. Add to that unclear blocks, repeated questions and the need to retype data between applications.
It is worth starting by observing a few ordinary tasks. Where does the employee lose the thread? Which message do they not understand? Can they recover work they have started? How do they report missing knowledge or an error? Such information should reach the person responsible for improving the process.
In our SAIE ecosystem, the environment for working with assistants and company documents is TWIN:DESK. A quotation from a specific fragment of the source makes it possible to check the answer straight away. PROXY:AI brings order to access to models and the application of policies. During deployment, we make sure that the user receives an understandable reason for a limitation and knows how to continue the task. New users are guided through TWIN:DESK by a walkthrough tailored to their role.
A way of working in the event of a failure is also needed. The salesperson should know where to find the price list and who will help them prepare an offer when the assistant is unavailable. This skill is worth keeping even when most tasks run automatically.
Keep the ability to think independently
A ready-made answer encourages quick acceptance, especially under time pressure. The risk is that employees gradually stop practising the activities needed to recognise an error. This applies in particular to people who are just learning the profession.
In a study published at CHI 2025, covering 319 knowledge workers, greater trust in generative AI was associated with less reported critical thinking effort. It was a survey, so it does not prove a lasting decline in cognitive abilities or a causal link. It does, however, point to a problem that is worth taking into account when designing work.
In the offer process, the salesperson can be required to check the price and the terms of any deviation from the standard. In a more difficult case, it is worth first asking them for their own assessment and then comparing it with the assistant’s proposal. In training, examples containing deliberate errors are useful: an outdated discount, a missing attachment or the wrong recipient.
The scope of checking should match the consequences of the decision. Rewriting every paragraph or taking a test before every minor action would quickly turn oversight into a nuisance. For critical decisions, on the other hand, it is necessary to check whether the approving person understands the key premises and can give a reason for rejecting a recommendation.
In our CDF methodology, this intention is expressed by the human competence gate. We design its specific form for a given process, and approval works in the application itself: in TWIN:DESK, the agent presents a plan for approval before it does anything.
Maintain relationships and knowledge sharing
In companies, many skills develop while solving problems together. A junior employee hears why an experienced colleague rejected a seemingly attractive option. The team learns about exceptions that have not yet been written into the instructions.
If every question goes only to the assistant, it is worth checking whether there are still opportunities for such conversations. We propose short reviews of difficult cases, working in pairs on new tasks and clearly naming the people to whom doubts can be referred. AI can prepare material for the meeting and help record the arrangements.
A company assistant should also communicate its role clearly. In matters concerning conflict, performance appraisal or overload, contact with an authorised person is needed. Access to a conversation with a model should not be the only form of support available.
Measure wellbeing without creating more surveillance
The result of a pilot is worth assessing together with the users. Alongside the time needed to prepare an offer and the number of corrections, we propose checking whether employees understand their responsibility, can point out an error and know how to interrupt a task. It also needs to be established whether the time saved remains once verification is included.
Information about the experience of work can be collected in conversations and short, voluntary surveys. Team results should be aggregated so as not to reveal the answers of individuals, especially in small groups. Before the measurement starts, its purpose, access to the data and the way the results will be used should be explained.
The number of conversations with AI does not describe the quality of a person’s work. Likewise, reporting many errors may be a sign of careful checking. When designing indicators, it is worth rewarding the detection of problems and knowledge sharing rather than the mere intensity of tool use.
An AI system should not diagnose employees on its own on the basis of their work conversations. In serious matters, the organisation’s role is to provide support from people and specialist help.
Turn the principles into a SAIE deployment
Our CDF methodology provides a place for these arrangements throughout the deployment cycle: from assessing readiness and needs, through preparing people, to evaluating the pilot and later reviews. The process owner should agree with the team what benefits it wants to achieve and by what signals it will know that the way of working needs adjusting.
CDF Platform supports running the stages, collecting documents and recording decisions. The methodology defines what needs to be agreed and assessed; the platform helps to organise this work. Day to day, people are supported by managers, HR, those responsible for health and safety at work, and the teams themselves.
TWIN:DESK and PROXY:AI provide the environment and the technical mechanisms, which we configure in line with these arrangements. An example is a process in which the assistant prepares a draft, the employee opens the sources and corrects the content, and sending requires their conscious decision. We test how the whole solution works together with the team in a pilot.
Start with a conversation about one task
Before extending a deployment, it is worth sitting down with the team to look at a specific process and agree: what AI will take over, what people will continue to learn, what they will spend the time saved on and who will respond to the difficulties reported. These arrangements should have an owner and a review date.
In our example, success will mean that an employee who understands the terms of a reliable offer and has time to check them prepares it more efficiently. If, at the same time, they can deal with an exception and draw on the support of the team, the organisation keeps the skills it needs for further development.
In the third part, we will move on to legal obligations: classifying AI applications, data protection and documented accountability for decisions.