Watch Tiberiu Socaci explain how to become an AI company, organise knowledge and permissions, and start with one team. Replay and presentation available.
What does it mean to become an AI company?
In this webinar, recorded on 6 October 2026, Tiberiu Socaci explains Makeitfuture's approach to building an AI company: give people and agents access to the knowledge they need, define who owns each process, and connect agents to the systems where work happens.
The aim is to make AI part of how teams complete work together. A sales agent might review opportunities and request updates from deal owners. A finance agent might check an invoice against the relevant records. People set the priorities, approve sensitive actions and remain accountable for the result.
This edited summary follows the session's explanations and examples. Watch the replay above for the full discussion.
Where should you use automation, and where does AI fit?
Tiberiu starts with a practical distinction. When a process follows predictable rules, conventional automation is his first choice. Moving data between systems or carrying out a fixed sequence of steps may not require an AI agent.
AI becomes useful when the task involves interpreting context or deciding how to handle an exception. For example, reviewing a sales conversation against the CRM requires understanding what was discussed and whether the relevant opportunity is already recorded.
The decision starts with the work you need to complete. Choosing an AI model comes later, once you know which parts of the process need judgement and which can follow defined rules.
The four requirements for an AI company
People own the decisions
People bring in context from customers, partners and changes in the market. They also decide what an agent may do. Each process needs a named owner, clear approval rules and a record of who requested or approved an action.
Moving work into a shared conversation helps the team see the request, review the result and contribute information that the agent does not have.
Company knowledge is accessible with the right permissions
An agent cannot use a decision that remains in someone's head. Relevant meeting notes, procedures, customer records and earlier decisions need to be available in a form it can read.
Access still depends on the person and the task. Shared company knowledge does not mean every employee or every agent should be able to read every document or inbox.
Agents can act in the systems where work happens
If an agent can only recommend a change, a person still has to carry it out. Tiberiu describes how he became a bottleneck in Makeitfuture's CRM review because the process waited for his feedback.
The response was to clarify ownership and give the agent an approved way to move the process forward. Permissions, review requirements and responsibility need to be defined alongside the connections to business systems.
Improvements feed back into the procedure
A useful feedback loop checks the result, records corrections and improves the procedure for the next run. These reviews can start from a request, a scheduled task or an error in an existing workflow.
One correction should not automatically become a company-wide rule. Tiberiu uses a proposal example: removing a clause for one customer does not mean removing it from every future proposal. Someone needs to approve whether that change becomes standard practice.
What does an AI agent need to do useful work?
The webinar describes six parts of the working environment:
- A model: the language model used to interpret the request and reason about the task.
- An agent environment: the tools, files, permissions and checks that let the model carry out work. Tiberiu discusses Claude Code and Codex as examples used in Makeitfuture's approach.
- Connections: approved access to the applications and workflows the task depends on.
- Skills: written procedures that explain how a particular job should be done.
- Memory and context: relevant company knowledge and previous decisions, available within the right access boundaries.
- Improvement loops: a way to check results and turn approved feedback into better procedures.
A model is one part of that system. The surrounding instructions, information and checks affect how the agent performs its job.
Why distinguish personal connections from shared tools?
An agent's access to someone's email should remain tied to that person's permissions. A colleague asking a question in the same channel should not gain access to that inbox.
Tiberiu also explains how a defined tool can make a multi-step job available without exposing an entire business system. A contract-creation workflow, for example, can provide one approved action instead of giving every requester unrestricted access to the applications behind it.
What should go into an agent's skills?
Some skills help an agent carry out a task, such as creating a document. Others capture the company's preferred way of doing it: proposal structure, presentation branding or contract-review steps.
Tiberiu argues that company-specific procedures remain important as models improve. Skills also need ownership and version control, so the team knows who can change a procedure and who must approve it.
Three examples from Makeitfuture's own operations
CRM review: removing the founder as the bottleneck
Tiberiu describes an agent reviewing Slack discussions and emails against HubSpot records to find sales opportunities that needed attention. The process slowed when it depended on him to provide all the feedback, even though other colleagues owned some of the deals.
He asked the agent to prepare a shared Google Sheet, moved responsibility for coordinating the review to a colleague, and had it follow up with the relevant deal owners. The lesson was about process ownership as much as the agent's ability to analyse information.
Contract changes: coordinating several departments
A change to contract terms created work for different teams: updating quotations and templates, preparing addenda for existing customers, and updating the website terms.
Tiberiu assigned those parts to the relevant colleagues and asked the agent to check progress weekly and prepare a report. The agent supported coordination; the departments remained responsible for their work.
Marketing: reviewing AI proposals before implementation
The webinar also describes a scheduled review of Makeitfuture's website performance. The agent identifies problems and proposes improvements, while Alexandra reviews the suggestions and decides which changes or articles should go ahead.
The team can see this work in the marketing channel, including the feedback and approvals that guide the next step.
How does ChannelGate bring this into the team's daily work?
Tiberiu presents ChannelGate as the layer that brings agents into shared team conversations. In the setup he describes, each channel has its own working context, approved skills and connections, with access controlled for the people using it.
Keeping requests and results visible lets colleagues contribute to the same task and learn from one another's use of AI. Running the work in the cloud also lets a person leave their laptop and continue the conversation from another device.
The session focuses on Makeitfuture's experience of this approach. Read more about ChannelGate and the AI operating system Makeitfuture uses.
Audience Q&A: what were the biggest challenges with AI agents in Slack?
Tiberiu identifies two challenges. The first was governing access to personal connections. He wanted an agent to help with his email without making his inbox accessible to everyone else in the channel.
The second was keeping up with the increased number of conversations. Shared threads made work more visible, but there were also more discussions to follow. He describes using that visibility to review how a sales opportunity was being handled and add feedback in the same thread.
This is an edited account of the audience question and answer, rather than a verbatim quotation.
How can your company start?
The approach presented in the webinar begins with choosing work that matters to the business, then putting the required connections and knowledge in place. Internal AI champions help their colleagues learn to work with the agent, and the team builds and reviews the process together.
Start with a defined job, a responsible person and a way to measure the result. Review the outcome and the cost before extending the approach to more teams. Makeitfuture's AI strategy consulting can help you choose the first process and plan its implementation.
