What an AI Company Really Is, and How to Become One

Summary

Tiberiu Socaci describes an “AI company” as one where AI is built into how teams work—not just used as a standalone chatbot. The goal is for agents to use company context and tools, carry out work under defined permissions, and improve processes over time, while people retain ownership and oversight.

Strategy and operating principles

• Use automation before AI: Apply conventional automation to deterministic, repeatable processes because it is typically more reliable, cheaper, and faster. Use LLM-based agents where a task requires judgment or decisions.
• Centralize company context: Traditional organizations pass context through management layers, where it can become diluted. An AI-enabled organization makes relevant knowledge, decisions, procedures, and data accessible through a shared “company brain,” with access controlled by role and permission.
• Make work legible to AI: Document decisions and procedures, and capture relevant discussions—including in-person meetings. If information is not recorded, the agent cannot use it.
• Give agents the ability to act: Agents should be able to work in systems such as a CRM or invoicing platform, not only provide recommendations. Otherwise, an employee may become a bottleneck for the workflow.
• Keep people accountable: Define who requested or approved actions, maintain permissions and audit trails, and make agent activity visible to the team.

Agent operating framework

An AI colleague should be more than a prompt-and-response chatbot. It should be able to:

• Start work from a schedule, automatic trigger, or employee request.
• Retrieve relevant company records, rules, procedures, and past decisions.
• Use connected tools to complete work and return results.
• Work in shared team spaces, with visible activity and clear ownership.
• Record approvals, permissions, actions, and—where useful—the cost per run.

The presenter outlines a recurring process loop:

1. Receive an input.
2. Check whether policies allow the requested action.
3. Use relevant tools and company context.
4. Verify the result against a quality gate.
5. Review feedback and update the procedure or process.

Updates should be governed: a one-off request or exception should not automatically become a rule for every future case.

AI system architecture

The presentation identifies six important components:

• Model: The underlying LLM.
• Harness: The operating environment that supplies tools, files, memory, permissions, and execution logic.
• System connections: APIs, personal credentials, and integrations with company platforms.
• Skills: Written procedures or SOPs that tell agents how to perform tasks.
• Memory and context: Company knowledge, databases, and thread-level history.
• Improvement loops: Mechanisms to evaluate results and refine workflows.

Socaci argues that the harness can matter as much as—or more than—the model because it determines how the model works, uses tools, and checks its output. He distinguishes between:

• Capability skills: Instructions for performing a general task, such as creating a PDF or presentation.
• Encoded-preference skills: Company-specific requirements, such as branding, templates, contract terms, or preferred formats.

As models improve, some capability instructions may become less necessary; company-specific preferences and governance still need to be encoded.

Security, governance, and team adoption

• Use MCP toolboxes or controlled workflows for complex tasks that require multiple systems. This lets an agent perform a defined task without giving every user broad access to underlying systems.
• Separate agent-level connections from individual users’ connections. For example, one employee may ask the agent to search their own email, while another user must not gain access to that mailbox.
• Isolate agents by channel or work area, with separate context, skills, and permissions for sales, finance, or other teams.
• Set global policies—for example, restricting an agent to answering questions at first, or excluding sensitive HR topics.
• Put agents in shared workspaces such as Slack or Microsoft Teams. The presenter says visible, collaborative use helped colleagues learn from one another and increased adoption at makeitfuture.
• Consider a super-agent to coordinate across specialized agents and surface shared learnings.

Implementation roadmap

The practical sequence proposed is:

1. Identify suitable processes and use cases.
2. Establish the foundations: company context, access rules, and governance.
3. Build the architecture, connections, toolboxes, and databases.
4. Train employees and build initial use cases with them.
5. Measure performance, review outcomes, and improve processes.
6. Deploy additional use cases based on what is learned.

A simple way to start a new agent is to create a team channel, add the agent, and define its role and instructions—then expand its tools and permissions deliberately.

Examples and operational playbooks

• Sales deal reconciliation: A CRM review process had become outdated because the presenter was not giving the agent feedback promptly. The agent reviewed roughly 14,000 messages and 9,000 emails, identified activity involving 13 deals and duplicate CRM records, and proposed a process update. The presenter then created a Google Sheet for deal owners to provide feedback and reassigned process ownership to a colleague, Nikolai, who could coordinate the follow-up.
• Contract review and implementation: After reviewing contract terms with an agent, the presenter delegated specific follow-up work to colleagues in different departments, including updating templates, preparing addendums, and revising website terms. The agent was asked to send recurring reminders and provide weekly progress reports, with a goal of closing the work by November.
• Proposal and quotation improvement: A scheduled agent reviews communications about proposals and quotations, extracts emerging rules, and asks for approval before updating the relevant procedure. This avoids treating an individual exception as a company-wide policy.
• Marketing monitoring: An agent checks website performance weekly and proposes improvements. A human colleague reviews and approves changes before implementation.
• Technical incident improvement: The presenter describes the possibility of triggering an agent when an automation or code process fails, having it propose a fix, check it against quality criteria, and deploy it where appropriate.

Metrics and claims mentioned

These are figures reported by the presenter, not independently verified in the subtitles:

• Agent task duration has grown from a few minutes in 2023 to multiple hours; he cited some recent models as capable of working for 16+ hours.
• Token prices are decreasing, although more capable models may use more tokens and keep some tasks expensive.
• The presenter estimated reliability of reasoning-only systems at 60–70%, rising to 80–95% with a well-designed harness and quality controls.
• He cited a benchmark example in which changing the harness increased a score from 40% to 92% for the same model.
• makeitfuture reportedly uses around 80 AI channels and has approximately 400–457 skills/procedures, with the library growing.
• The company described itself as having about 30 people and said it had spent roughly six months developing its approach to enterprise governance and deployment.
• The deal-reconciliation example covered about two months of activity.

The talk does not provide revenue, margin, CAC, or other conventional financial KPIs. It emphasizes process quality, adoption, permissions, traceability, and the ability to measure outcomes and cost per run.

Presenters and sources

• Presenter: Tiberiu Socaci, founder of makeitfuture. The subtitles contain several inconsistent spellings of his name.
• Host: Alexandra, makeitfuture.
• Examples and company practices: makeitfuture’s internal AI workflows and customer projects, as described by Socaci.