EU Firms Are Adopting AI Agents. Most Stop Too Early.

By Lazar Milicevic · Published October 15, 2026 · 7 min read

You gave your team ChatGPT seats, and the workflow is still: paste an email in, copy a draft out, paste it into Gmail. That saves two minutes and feels like progress. OpenAI's new EU report points at the same gap, and it has less to do with adoption percentages than with how deep each team's use goes.

What the OpenAI report says, and what it doesn't

The report is a vendor describing its own customers. OpenAI Economic Research published "The EU's next chapter in AI is agentic" on October 14, 2026. Read it as a directional signal, not an independent audit.

What it does say:

• Weekly active users of Codex and ChatGPT Work among EU small and medium businesses reached 8.1 times their May level by October 5, 2026, with growth in every member state. That is growth relative to a May 1-7 baseline, not the share of EU SMBs using agents. Reported here.
• OpenAI defines an SMB as 200 or fewer associated users. That is a count of users in its data, not the EU's official employee-and-turnover definition.
• Growth ranged widely: Belgium 12.1x, Italy 10.8x, Luxembourg 10x. Every country grew more than 5x, and the report ties the acceleration to the introduction of ChatGPT Work in July (source).
• In OpenAI's EU sample of ChatGPT Enterprise customers, Codex and ChatGPT Work now account for more than half of output tokens. Codex is a coding tool, so that share may not reflect non-technical workflows.
• The depth gap: in September, EU "frontier firms" (the top 10% by output tokens per active user) generated more than four times as many output tokens per active user as typical EU firms (source). Frontier firms outside the EU generated more than twice the output tokens per active user of EU frontier firms.

What it does not say: that small businesses use agents "in the shallowest way possible." The depth data covers Enterprise and Business accounts, not SMBs specifically. Tokens per user is also a rough proxy for depth. A big user can be burning tokens on long chats.

My claim is narrower and comes from building these systems. Breadth is easy. Eight-fold growth in weekly users means more people opened the tool. It doesn't mean more workflows got wired into real systems. That second part is where the hours come back.

Shallow vs. deep use, with a concrete example

Shallow use is a human acting as the integration layer between a chat window and the rest of the business. Deep use moves the human out of the loop for routine cases and into an approval role for the rest.

Here is the same task, an inbound customer email asking about an invoice, done both ways.

Same model in both columns. The difference is the plumbing. A model in a browser tab can't see your billing system, your calendar, or your lead list, so it can only help with what you paste in. That is why the common complaint "we tried ChatGPT, but it didn't integrate with anything" is accurate. It's also why generic SaaS often disappoints small teams: it assumes a workflow you don't have.

The product shift behind the growth: scheduled, acting agents

The report's growth numbers line up with a product change. OpenAI launched ChatGPT Work on July 9, 2026, an agent that executes tasks across apps and files (it launched on GPT-5.6; check the current model, since that may have changed). Two features matter for the depth question:

• Scheduled Tasks. Per the ChatGPT Business release notes, tasks can run once, repeat on a schedule or trigger, or monitor for changes. That is the difference between a one-off prompt and a recurring workflow.
• Approval for sensitive actions. ChatGPT Work asks a human before taking sensitive actions, and enterprise admins can preauthorize which actions it may take and set usage limits (SiliconANGLE).

On cost, Business Standard seats are listed at $25 per user per month billed monthly or $20 billed annually (USD, may vary by country), and include both ChatGPT Work and Codex. So the agent itself is not a separate purchase. What it costs you is the integration work, which is the part nobody budgets for.

Here is a sketch of what a scheduled, monitoring task looks like in practice. This is illustrative pseudo-configuration, not a real ChatGPT Work file format:

# Illustrative only: the shape of a deep workflow, not a vendor schema
workflow: inbound_lead_followup
trigger:
  type: new_message
  source: contact_form_or_shared_inbox
steps:
  - extract: [name, company, request_summary]
  - lookup:
      system: crm
      on_missing: create_record
  - draft_reply:
      must_reference: request_summary
  - route:
      default: human_approval
      auto_send_when: routine_and_approved_template
logging:
  record: [input, crm_record_id, draft, decision, actor, timestamp]

Notice what takes most of the engineering: the lookup, the route rule, and the logging block, not the draft. That is the connection work.

Pick one workflow using three tests

Don't try to automate your company. Pick one workflow that passes all three:

1. It happens at least ten times a week.
2. It touches at least two systems.
3. A mistake is cheap to catch.

Fewer than ten runs a week and you won't get enough volume to measure improvement. One system and you're just reformatting text, which a chat window already does. An expensive mistake means you'll never trust it enough to loosen the approval step.

Inbound lead follow-up passes all three for most small teams. Here's the exact shape:

• Trigger: a form submission or email lands.
• Step 1: the agent extracts name, company, and what they asked for.
• Step 2: it checks your CRM for an existing record and creates one if there isn't.
• Step 3: it drafts a reply that references what they actually asked.
• Step 4: the draft arrives in a chat message with approve and edit buttons.

You approve in ten seconds instead of writing for ten minutes. Other workflows that tend to pass the three tests: invoice reminders, appointment rescheduling, and routing support emails to the right person.

Measure, then loosen the leash

Run it for two weeks with human approval on everything. Track exactly two numbers:

unedited_rate  = drafts_sent_unchanged / total_drafts
minutes_saved  = avg(manual_minutes_per_lead) - avg(approval_minutes_per_lead)

When the unedited rate is high, let the routine cases send on their own and keep approval only for the unusual ones: a large deal, an angry customer, anything that mentions a refund or a legal term. If the rate stays low, the draft prompt or the CRM data is the problem, and you've learned that cheaply.

An illustrative example with round numbers (not a benchmark): if you handle 20 leads a week and writing each reply takes 10 minutes, that's about 200 minutes. If approval takes 10 seconds per lead, you spend about 3 minutes. The saving only shows up when the agent is connected to the CRM and inbox. Pasting into a chat window gets you the 2-minute version of this.

Build compliance in from day one

If you sell into Europe, your customers ask where their data goes, and a wrong entry in a billing or customer record is an audit problem, not an annoyance. The deeper-use path therefore has to include three things from the start:

• Logging. Every run records the input, the systems touched, the output, and what decision was made.
• Approval steps. The agent proposes and a human approves anything that touches money, contracts, or personal data, until the numbers justify loosening it.
• A clear record of why. Store the reasoning or the rule that triggered each action, not just the result.

Do this up front and compliance becomes a feature you can show a customer. Bolt it on later and it becomes the reason the project gets paused. For specific obligations such as data protection and e-invoicing, check the official guidance for the countries you operate in rather than relying on a blog post, including this one.

Why bizflowai.io helps with this

At bizflowai.io the work is the unglamorous middle: wiring an AI model into the inbox, CRM, accounting tool, and chat approvals a small team already uses, with logging and an approval queue built in from the start. The approach is the one described above: one workflow, two weeks of human approval, two numbers tracked, then selective automation. I'm a senior engineer who builds AI automations for small teams, and I start by checking whether a workflow passes the three tests before writing any code.

Want more like this?

I publish practical AI automation, GenAI engineering, and faceless content workflows on YouTube every week.

Subscribe to bizflowai.io on YouTube — never miss a new tutorial.

Planning an AI automation project or need a second opinion on your architecture?

Connect with me on LinkedIn — Lazar Milicevic, GenAI Engineer & bizflowai.io Founder.

Visit bizflowai.io for our services, case studies, and AI consulting.