The hybrid organisation

At the corporations MING Labs talks to about their agentic enterprise programmes, employees work with Copilot, and hundreds of assistants run in Copilot Studio alongside it; we have yet to meet a client with more than ten AI agents that hold a role of their own.[S1] The step companies now face leads from the assistant that helps one person to the agent that takes on a role in a team.

Salesforce uses the term agentic enterprise for companies in which people, AI agents and platforms work together.[S2] Our approach to that is the hybrid organisation: people and AI agents hold clearly agreed roles in the same team. A named person is accountable for every result, and each new team builds on what earlier teams have learned.[S3]

Clear roles, shared knowledge and human accountability belong together in our approach: without the accountable person you get a risk; without shared knowledge, a silo; without the clear role, you stay with assistants.[S3] The shared knowledge lives in a knowledge graph, the company’s brain; section 04 shows how it is built.

The hybrid organisation shapes how each team works and, in the end, the shape of the whole company. It moves the line of what is worth doing: work that used to be left undone or was not even conceivable becomes feasible. That is why this is about competitiveness and not only about cost.

MING Labs has designed and built digital products and systems for enterprises since 2011, and has worked with autonomous agents in its own team since January 2026: within the limits of their roles, they take on tasks from start to finish and deliver results, rather than assisting one person. For our approach to roles, the practical question is whether the system completes an agreed assignment without continuous input from a person and delivers the result. Such an agent keeps working while the person is in the car. The product name alone does not establish that. Our approach distinguishes three levels: leadership with its target picture, rules and bar at the top, teams that give each agent a role and an accountable person in the middle, and the shared foundation of clear roles, shared knowledge and connected systems beneath them.[S3]

This article shows each level in a different organisation: leadership through conversations with corporations, teams through the agent Tabea at the mid-sized company STREIT, and the foundation through our own operation.

For a quick overview: Our short answer for decision-makers.

01 Most companies are still testing

According to a Lünendonk study published in September 2026, surveying 180 IT and business leaders at mid-sized companies and corporations, only one percent of the surveyed companies in Germany, Austria and Switzerland use AI agents end to end across several core processes; most are piloting them.[S4]

At the same time, 66 percent say that fewer than a quarter of their previous AI pilot projects have reached productive use.[S4] Broad rollout is still at an early stage; 19 percent already use agents in individual areas of work. One industrial group we have worked with for years estimates that most of its AI use cases are still assistants, not agents.[S1]

Assistants help individuals; agents with a role help the team.

In McKinsey’s global survey on AI as a whole, 89 percent of organisations use AI regularly in at least one function, yet only about 6 percent of respondents attribute at least 5 percent of their operating earnings to it and call its impact significant; that share has stayed flat.[S10]

80 percent of respondents report higher personal productivity, while only 37 percent report a contribution to operating earnings.[S5] That gap matches what we see in companies: many people get faster personally because they discover the tools for themselves; the team does not get faster, because the context is lost at every handover. An agent with a role in the team shares its context with everyone who works on the task.

At one consumer goods group we exchange views with, Copilot agents and self-built agents handle small tasks for individual employees; the company has found that the big lever only appears once it questions entire processes.[S1] Behind this, we see a pattern: agents are built as one-offs, each with its own knowledge and its own rules; a correction improves only that one agent, and the maintenance effort grows with every agent added.[S3]

Lünendonk lists the obstacles: integration with systems and data, security and data protection, a shortage of skilled staff for implementation and operation, and poorly coordinated AI initiatives.[S4] Without access to those specialists, even a good pilot can stall. In June 2025, Gartner predicted that over 40 percent of agentic AI projects will be cancelled by the end of 2027, because of escalating costs, unclear business value or inadequate risk controls.[S6] Each of Gartner’s three reasons can be tackled before the first pilot: cost with a calculation per correctly completed case, unclear value with a bar agreed in advance, and risk with a role that has limits and an accountable person.

Back to contents

02 Every agent needs a role

A role includes a purpose, a result that someone needs, limits and a person who is accountable for that result.[S3] A mid-sized company shows what that looks like day to day: with the office supplier STREIT, we developed the agent Tabea for its customer service. According to STREIT, she already prepares up to 40 to 50 percent of orders automatically, with up to 120 hours of capacity gained per month; as an outlook, the company names up to 80 percent automation.[S7]

STREIT supplies office products, office technology, office furniture and services from sites in Gengenbach, Freiburg and Villingen-Schwenningen. In its customer service, Tabea reads order emails, matches customers and products, and creates order proposals for staff.[S7] The case study Tabea at STREIT shows the workflow.

Salesforce recommends starting with a high-impact use case.[S2] MING Labs starts with the role.[S3] A use case describes what a system should be able to do; a role describes which work someone takes on in the team, who needs the result and who decides. STREIT, too, describes an AI agent as a fixed part of operational workflows, with a clear role, defined permissions and traceable handover points.[S7]

One industrial group we have worked with for years describes its dilemma like this: either you think too small, the agent as a better chat that automates a little, or too big, the agent as a replacement for a person.[S1] The role sits in between: a clearly bounded part of a team’s work, assigned to one person. How seriously that is meant showed at one client, where the manager intended for the agent was not allowed to have reports at her pay grade; the agent was given a different manager.[S3] Only with a place in the org chart does a tool become a colleague.

A role starts with the question of who needs the result.

We have seen what happens without that question: our agent Major Tom was meant to coordinate the work of our other agents. In March 2026, after 13 days and 764 messages, we shut him down; he had a capability but no accountability for an outcome.[S8]

Back to contents

03 What leadership decides

Leadership decides three things: the target picture, the rules of the shared foundation and the bar, meaning the result a pilot must reach to go further. It sets the bar before the first pilot, and the target picture and foundation before rollout at the latest.[S3] STREIT is a mid-sized company; in a corporation, more teams and more approvals come into play. There, as part of the foundation, leadership sets out once for all teams what STREIT provides for AI agents from the outset, such as access rights, logging, human control and escalation paths.[S7]

Before the pilot, leadership and team agree how quality, time spent and cost will be assessed. The company’s own experts rate quality without knowing whether a person or the agent produced the result. After the test and again after the pilot, the accountable person decides: continue, revise or stop.[S3] Our guide to testing before production explains how real cases and expert acceptance support the test.

At one consumer goods group, two managers took opposite sides in the same conversation with us: one wanted to make the new way of working tangible in a single team first, the other wanted a target picture and a repeatable process from the outset.[S1] For us, the two belong together. The first agents emerge in teams, what they need as a foundation becomes the foundation for everyone, and the target picture and the first teams develop in parallel.[S3]

Leadership draws the line: the role standard, review procedures, logging and approval paths apply across the company; each business unit decides its own tasks, roles and accountable people.[S3] The works council and the data protection officer belong at the table before the first pilot: people hand over tasks and take on new ones, and an agent’s log also shows who in the team corrected what.

Leadership itself takes on new tasks along the way: describing and filling roles for agents, accepting agents’ results, approving the next stage of autonomy and leading teams through the change in their work. Whoever writes these tasks into managers’ goals and job descriptions makes the change a matter of leadership rather than an IT project.[S3]

McKinsey reached the same conclusion in September 2026: the gap between AI’s promise and its value is also a leadership gap, and the decisions on ambition, organisation and culture cannot be delegated by the CEO. Operating models need to be redesigned around human-agent teams; managers will spend less time gathering and passing on information and more time on judgement, coaching and exceptions. McKinsey calls role modelling by senior leadership the single strongest predictor of a successful introduction: 70 percent of employees say they feel ready to use AI, but only 27 percent of leaders say their organisation is ready to make the changes it requires.[S10]

Back to contents

04 One team learns for all

For that to work, the lesson has to be stored in a shared foundation. We build it with the first team, from three parts: clear roles (who does what, and who decides), shared knowledge (what the company knows) and connected systems (what an agent may see and change, with at most the rights of a colleague in the same role).[S3] The foundation can be built on existing platforms. Shared knowledge also holds the teams’ reviewed corrections.

We build shared knowledge as a knowledge graph: the company’s roles, rules, workflows and terms are linked in it, so that an agent can look up what applies and a person can see where a rule came from. In conversations, this layer is often called the company’s brain. Defining and building it well is the part of our work that holds the most experience, because it decides whether the second team is faster than the first.[S3] Where this process and context knowledge lives is the company’s decision: our own foundation runs on servers in Germany, and we can run it for clients on servers in Germany or the EU, including models that run there. Location alone is not enough; who has access and who can replace the service are part of the question, as our concept of digital sovereignty in AI explains. McKinsey advises boards the same way: a company’s data, knowledge, decision logic and semantic layer should be permanent assets, the models beneath them interchangeable; the advantage comes from the learning loops built around the models, and it cannot be bought, only built.[S10]

A knowledge graph now partly builds itself; a person still has to curate it and decide what is right and what can go. For that, agents show where data is missing or contradictory, instead of someone scanning spreadsheets with tens of thousands of rows. Data maintenance thus becomes a hybrid task for data owners and agents, and poor master data is no longer a reason to wait.[S3]

Reviewed corrections help the next agent too.

Because agents need their knowledge in writing, what used to live only in people’s heads gets written down along the way, and it stays in the company when people change jobs.[S3] We explain how to tell whether corrections are working in our answer on whether an AI agent is improving.

Back to contents

05 Who has the final say

An agent has to deliver where people already work.

Every result an agent produces needs a person who receives it and is accountable for it; in our approach, that is a named person in the team.[S3] For STREIT, introducing an AI agent “is not an IT project in the classic sense”, and an agent’s success depends on “whether people trust it”.[S7]

One simple decision helped us: our agents bring their results to the people accountable for them, for example into their email inbox.[S9] Clients confirm it: a sales leader at one chemicals group wanted no further interface for his sales staff to learn; without adoption, he said, the best ideas are wasted time and money.[S1]

The conversations with corporations yield a further observation: people forgive colleagues their mistakes, agents less so.[S1] Anyone introducing agents plans for that lower tolerance in how they communicate the change, and gives agents their autonomy in stages.[S3]

The accountable person must be able to recognise when the agent’s work is wrong, and to explain what it does and why. MING Labs calls this duty the Comprehension Obligation. Review therefore needs time in the plan and someone to cover for that person.[S3] It also sets a limit on the span of control: a person is accountable only for as many agents as she still understands, and no agent commissions other agents across several levels without a person being able to explain the chain.[S3]

Back to contents

06 Where the value shows up

Time gained only counts once the team uses it for other work.

MING Labs recommends measuring an agent’s value by the cost per correctly completed case, with simple cases and exceptions assessed separately.[S3] The figure is compared before and after the agent starts work, and the cost includes development, system integration, operation, model calls and human review. If an employee pieces every case together again to check it, person and agent are doing the same preparatory work. A pilot that mostly sees easy cases looks better than everyday operation will.

The public debate is currently about tokens, the unit language models are billed in. McKinsey now calls token costs a board-level topic: the same AI task can consume up to 30 times more tokens at one company than at another, depending on how it is orchestrated.[S10]

To better control model costs, we do not run every step through a language model. Most of an agent, as we build it, is fixed program code: matching customers, checking article numbers, calculating deadlines. The language model is used where language has to be understood or written. That makes model calls and their costs easier to trace. Total cost per case also depends on the mix of cases and the review effort.[S3] Not every role needs the strongest model: a rare specialist role may cost more tokens than an agent that reliably repeats the same work.[S3]

Tokens are only part of the cost. There is also human in the loop: people reviewing and approving results. In our worked example, one accepted result costs €18.90, so 100 results cost €1,890 a month: €200 for model calls and other services, €120 for fixed software, €1,250 for human review and approval, and €320 for ongoing support. Every value is an assumption.[S11]

How the agent comes to need less human review

Our hybrid organisation approach addresses this review effort. People train the agent for its role: they correct recurring errors, have the resulting rules checked by subject experts and record them in shared knowledge. That knowledge is available to the next agent too. The team should not have to explain the same error again on every case.[S3]

Trust grows from demonstrated quality. When the agent proves reliable on real cases, the accountable person approves more autonomy for clearly bounded tasks. Those tasks no longer need every result approved individually; exceptions and high-risk cases stay with people. Sampling and fresh tests after changes check that quality holds. When errors occur, approval is restricted again.[S3]

This connects the learning loop in section 04 to the team agent in section 02: less repeated review means less duplicate work and less waiting for approval. Our aim is to reduce human review effort per correctly completed case as the agent learns its role. We measure whether that happens alongside result quality.[S3] Our answer on the running costs of an AI agent works through every cost type.

At STREIT, by the company’s own account, the team puts the capacity it has gained into customers and special cases.[S7] In any company, whether that turns into more revenue or lower costs only becomes clear once the whole workflow is compared.

Comparing effort per correctly completed case shows whether onboarding actually reduces it. If a task then costs less than its expected value, the team can take on work that used to be left undone.

Back to contents

07 Work that used to be left undone

Agents can make work possible that used to be left undone or was not in the plan.

Many valuable tasks are left undone because doing them costs more than they are expected to bring in. Typical examples are analysing only a sample of sales calls, checking only the largest tenders and looking after small customers on the side.

Training agents for their roles can move this boundary. Reviewed shared knowledge improves their foundation; reliable results allow more autonomy and less human review. When that brings the effort below the value of a task, work previously left undone becomes worthwhile. We call this latent work: the value was already there, but doing it did not pay.[S3]

Further teams can build on knowledge that experts have checked. Their agents must also prove themselves on their own cases before the accountable person approves greater autonomy. Whether review effort falls further is measured again. When ongoing tasks such as scanning every tender each day become affordable, a team can also offer services it had not previously planned.[S3]

The line of the possible moves

What used to be left undone gets done, in two ways. In breadth, agents take over: they check every tender, every call, every small customer. According to McKinsey, agents at a European insurer review 95 percent of sales calls, up from 3 percent.[S14] In depth, people gain: they get time for the work that goes first under time pressure, namely advice, special cases and relationships. According to STREIT, its customer service team uses the capacity gained for customers and special cases.[S7]

This makes good on the sentence from section 06: time gained only counts when the team uses it for other work.

The line of the conceivable moves

Teams can also consider work they had previously ruled out as too costly. At MING Labs, one agent goes through the tender portals every day, checks each tender for fit and prepares the bid. Another summarises each morning what has changed with customers and in utilisation (section 08).

Without agents, these tasks regularly take time for screening, comparison and preparation. Surveys point the same way: in Microsoft’s survey of 20,000 AI-using knowledge workers across ten countries, 58 percent say they now deliver work they could not have delivered a year ago.[S12] In a study analysing 331 responses from software developers at Microsoft, 12.4 percent describe entirely new work.[S13] Both are self-reports that do not measure the value of the work.

This is where competitiveness comes from

Making this work economically viable creates an opportunity to do more for customers. The following effects are goals for the pilot, not measured results:

Additional work often has no direct previous baseline. Cost per case remains relevant but does not capture the additional value on its own. So a value assumption belongs before the pilot, for instance for the tender agent:

additional tenders checked × bid rate × win rate × order value

The formula estimates additional revenue over a defined period. Whether the work pays off depends on the expected contribution margin after additional bid and agent costs. That becomes the yardstick from section 03. This work also needs a role and an accountable person, because more results are not yet value.

What further teams can reuse

New tasks produce experience and corrections. Rules that have been checked by subject experts and apply beyond the original case enter the shared knowledge (section 04). The next agent starts with this reviewed foundation. It must also prove itself before gaining more autonomy. The aim is less onboarding, rework and ongoing review. Each team must check whether this reduces cost and makes further work economically viable. McKinsey sees the lead in these learning loops: it cannot be bought, only built.[S10]

Whoever only speeds up existing workflows today sees small gains, according to McKinsey,[S15] and efficiency alone, it says, is not a survival strategy.[S10] That is a reason to test additional work in the pilot, alongside improvements to existing workflows.

What new work the company takes on

New work does not appear by itself. Acemoglu and Restrepo show for the USA over the three decades before their 2019 publication that automation displaced tasks more strongly while new tasks emerged more weakly.[S16] Using agents only to save costs risks overlooking opportunities to create new work.

Which neglected and which previously unthinkable work a team takes on first therefore belongs in leadership’s target picture. The pilot must show which additional work is worthwhile for the team.

Back to contents

08 From one team to the whole company

For individual teams to form a hybrid organisation, every new agent builds on what already exists. In our conversations with corporations, the discussion turns to a register of agents for this, much like the org chart for people: name, role, rights, accountable person. It is to be kept by a central unit, such as a staff unit or a programme. Anyone who wants a new agent checks there first whether an existing one can take on the task; one corporation creates template roles from which its business units derive their agents. The central unit thus becomes a kind of HR department for agents: it knows how agents are onboarded and evaluated, and when one is shut down.[S1]

How big a role becomes is decided, in our approach, by two questions. How strictly must the workflow be followed: an agent for a fixed process gets a tight frame in which the language model has little room; an agent for open tasks such as marketing may work more broadly. And which access rights does it need: an agent that may read fifty systems and write to ten is too big; then it becomes two or three with clear limits.[S3]

At the same time, an avalanche of tools arrives in corporations, from the company itself and from vendors; if nobody defines what applies company-wide and what stays with the business unit, every department uses something different and none of it scales, as one chemicals group has found.[S1] Existing assistants and pilots, including those built in Copilot Studio, go into the register with a role and an accountable person, or they are switched off.[S3] In a mid-sized company, one person can keep the register; after the pilot, the accountable person in the team looks after the agent’s work.

MING Labs has already set up agents for further teams: for sales, tenders and proposals, marketing and operations. One of them scans tender portals every day, checks whether a tender fits, gives the team a recommendation and prepares the proposal. Another keeps the CRM up to date, tracks utilisation and writes a morning brief.

Many separate pilots and a hybrid organisation end up differing in four places: the agents, the result, the knowledge and the next team.

Back to contents

09 Where you can start

Who would you hire tomorrow if you could?

The answer to that question points to work that is being left undone today, and with it to the first role. The second question reaches further: what would you do for your customers if it cost next to nothing? To start, one team, one recurring task and one person who will be accountable for the result are enough.[S3] This is where the step from assistant to agent with a role of its own begins.

Only one percent of the companies Lünendonk surveyed work with agents end to end.[S4] Corporations, too, are at the start, so anyone who begins now is early. STREIT describes how new roles emerge and how employees have to build trust.[S7] Waiting postpones that work as well. Whoever sets the target picture, foundation and bar now decides for themselves which pilots move forward and which neglected work comes first.

Lünendonk counts a shortage of specialists for implementation and operation among the obstacles.[S4] Where those specialists are missing, MING Labs builds the agents together with your teams. We talk to corporations about their programmes, run several agents on one shared foundation in our own operation and developed the agent Tabea with STREIT. Our hybrid organisation page describes how we define roles for people and agents together.

At MING Labs, the way to a hybrid organisation begins with an orientation call, an assessment workshop and a pilot within a few weeks.[S7] STREIT and MING Labs offer this entry jointly to mid-sized companies. The team takes part with the person who will be accountable for the result; in a corporation, leadership joins as well, and we set the target picture, foundation and bar with it in parallel. After the workshop, you have the basis for deciding on the pilot: the biggest lever, a business case, the role description, the split of the work and the bar. After the pilot, a developed and tested first agent is in place, and the accountable person decides what happens next.

Back to contents

About this article: The Lünendonk and McKinsey figures are self-reports from different surveys; McKinsey asks about AI as a whole. Gartner’s figure is a prediction from June 2025. STREIT’s figures are client-reported, without a published measurement period and not independently verified; time gained is not yet a cost saving, and STREIT offers such agents together with MING Labs. The observations from corporations come from confidential, anonymised conversations in autumn 2026; we name industries, not companies. The approach and our own operation describe our practice and are not a measurement for other companies. The cost example uses assumptions, not measured running costs. The figure on latent work is a schematic. Microsoft’s figures are a vendor’s self-reports, the developer study is a preprint on one group, and McKinsey’s insurer is a case that has not been independently verified.