Resources, Governance
Rolling out AI agents: your first two weeks and beyond
Most agent rollouts go wrong in the same way: too many processes at once and no figures to judge them by. This guide sets out a two-week pilot for one process, the decision at the end of it, and how to grow from there to company-wide use.
Updated 20 November 2026. 9 minute read.
In short
- Start with one recurring, measurable process that has a clear owner. Time the current way of working first, so you have a baseline to compare against.
- Give the pilot agent a role, a manager, a short list of apps, approval rules for anything hard to reverse and a budget. Agree all of this before day one.
- For ten days, the employee who owns the process approves or corrects the agent's work. Track how often work is approved unchanged, changed or rejected.
- On days 11 to 14, compare quality, time and cost per completed task with the baseline, then decide to stop, adjust or expand.
- Scale one process at a time, with a champion in each team, trained approvers, approvals loosened only on evidence, a fixed review cadence and a plan for incidents.
Before you start
A pilot answers one question: does an agent do this process well enough, at a cost that makes sense? You can only answer it if the process is right and the groundwork is done. Allow a week or two for the steps below.
Pick one process
- Recurring: it happens often enough to give you dozens of examples in ten working days.
- Measurable: you can say what done looks like and check it.
- Owned: one employee is responsible for the result today and will review the agent's work.
- Contained: a mistake is annoying, not serious. Leave decisions about individuals, such as hiring, for later.
Hypothetical example: a wholesaler with 200 employees picks drafting replies to order-status emails in its customer service inbox. The customer service lead owns the process. Around 40 such emails arrive each day.
Set a baseline
Time a sample of tasks done the current way and score their quality against a short rubric. Use the method in our guide to measuring AI agent performance: minutes per task, quality score and cost per completed task. Without a baseline, the decision on day 14 is a matter of opinion.
Agree the rules and get sign-off
- Approval rules: which actions the agent may take alone, and which need an employee to approve first.
- A budget for the pilot, with a warning before the limit.
- The AI provider key: get the account and key approved through your normal purchasing and security checks.
- A register entry: record the use case, its owner, the data it touches and a short risk check. Confirm it is not a high-risk use under the EU AI Act.
Days 1 to 3: set up one agent
Set up a single agent. Resist adding a second one, even if it looks easy. Each item below should be written down, not only configured.
- A role: one sentence on what the agent does and what it does not do. For the example: drafts replies to order-status questions; does not handle complaints, refunds or address changes.
- A manager: the process owner, who answers for the agent's work.
- Apps: only what the task needs. Here, read access to the order system and the ability to draft, not send, emails.
- Approval rules for anything hard to reverse: sending to a customer, deleting records, spending money, changing data in another system.
- A budget: a monthly limit with a warning at 80%, so you hear about rising costs before the agent pauses.
Test on a handful of past examples before going live. Check that a request for approval reaches the process owner where they already work, and that pausing the agent works.
Days 4 to 10: run with review
Now the agent does real work, and the process owner reviews every item that needs approval. The aim is to learn where the agent is reliable and where it is not.
- Approve or correct: the owner approves good work, edits work that is nearly right and rejects the rest, with a short reason.
- Turn corrections into lessons: when the same fix comes up twice, write it down as a rule the agent should follow, and approve it so the agent uses it from then on.
- Check the activity log daily: ten minutes on what the agent did, what it cost and anything unexpected.
- Track three rates: approved unchanged, changed before approval and rejected. Note the reason for each change.
Keep the owner's review time in the figures. If reviewing a draft takes as long as writing the reply, the agent is not yet saving time.
Days 11 to 14: decide
Compare the pilot with the baseline on three measures. Use the same rubric and the same way of timing as before.
- Quality: is the approved work at least as good as the baseline sample?
- Time: employee minutes per task, including review and correction, against the baseline.
- Cost per completed task: AI usage, tools and employee time, divided by tasks completed successfully.
Make one of three decisions
- Stop: quality is below the baseline and the corrections show no clear pattern. Record why, so the next pilot avoids the same trap.
- Adjust: the agent is close but makes repeated, fixable errors. Change the role, instructions or apps, and run another two weeks.
- Expand: quality holds, time and cost per task are better, and the change rate is falling. Keep the agent in use and move to the next process.
Write the decision and the figures on one page and share them with the sponsor. That page is also the template for every later pilot.
Scaling up
The second and third processes
Repeat the pilot for each new process: baseline, set-up, review, decision. Choose the next processes from teams that asked for them, not only from the one that ran the first pilot. Run no more than two pilots at once until the routine is familiar.
Champions per team
Name one champion in each team that adopts agents. A champion collects ideas for new processes, helps colleagues with approvals, spots problems early and reports back to the rollout lead. Give them a few hours a week for it, agreed with their manager.
Approver training
Every approver needs to know what the agent does, where it tends to fail, how to check its work and when to pause it. Training also covers automation bias, the habit of approving without reading. Article 4 of the EU AI Act requires companies that use AI systems to take measures to support AI literacy among their staff. Our guide to AI literacy explains what that means after the Digital Omnibus.
Loosening approvals as trust grows
Loosen one action at a time, and only on evidence: a low change rate over several weeks, confirmed by reading a sample. Move that action from approval to act and report, and keep reviewing the log. In the example, order-status replies might go out alone after a month, while complaints still need approval.
Review cadence
- Weekly for each agent during its first month.
- Monthly after that: a sample of work, change rates, costs and incidents.
- Straight away when the model, instructions, apps or approval rules change.
- At least yearly for each register entry.
When an agent does something wrong
- Contain: pause the agent or remove its access to the app concerned.
- Fix the effect: correct the record, or contact the customer or colleague affected.
- Find the cause in the activity log: the instructions, the data, a missing approval rule or the review.
- Correct it: update the role or instructions, add a lesson, and tighten the approval rule for that action until it has a clean record again.
- Record it in the register entry and share what you learned with other approvers.
If the action was approved, treat it as a shared failure, not the approver's alone. Ask what the approver saw, whether it was enough to spot the problem, and whether they had time. Blaming approvers teaches them to reject everything or to stop reporting.
Change management in a company of 100 to 500 employees
At this size, most employees know the colleagues whose work changes first. Rumours spread faster than announcements, so communicate early and plainly.
Communication
- Say which process is changing, why, and what the agent will and will not do.
- Say who stays in charge: employees approve the agent's work and can stop it.
- Share the pilot results, including what did not work.
- Answer questions about jobs honestly. Do not promise what leadership has not decided.
Roles
- Sponsor: a member of leadership who owns the decision to expand.
- Rollout lead: runs the pilots, the register and the review cadence.
- Process owners and approvers: manage the agents and review their work.
- Champions: one per team.
- IT, security and the data protection officer: sign off apps, keys and data access.
Works council and employee information
If you have a works council or other employee representatives, involve them before the pilot, not after. For high-risk AI systems used at the workplace, Article 26(7) of the EU AI Act requires employers to inform workers' representatives and the affected workers before the system is put into service. For Annex III uses, these rules apply from 2 December 2027. Recital 92 adds that the Act does not replace duties to inform or consult workers under EU or national law, so check those rules with HR or legal advisers even for agents that are not high-risk.
Keep employees in the tools they already use
Adoption is easier when little changes in an employee's day. Send approval requests to the chat tool the team already uses, keep the agent's output in the systems it already works in, and give interested employees read access so they can see what agents do.
The checklist
- One recurring, measurable process with a named owner.
- Baseline: minutes per task, quality score and cost per completed task.
- Approval rules and a pilot budget agreed in writing.
- AI provider key approved by IT and purchasing.
- Register entry and short risk check completed.
- One agent with a role, a manager, limited apps, approval rules and a budget.
- Daily log check and change-rate tracking during the pilot.
- A written stop, adjust or expand decision with figures.
- A champion per adopting team and trained approvers.
- Rules for loosening approvals, a review cadence and an incident procedure.
- Works council or employee representatives informed where relevant.
How Blentera helps
Blentera is a platform where companies run AI agents with employees in charge. Each agent has a role and a manager in the org chart, and employees have the role of owner, admin, operator or viewer. Approval rules can be set per agent, project, app or tool, and employees answer them in Slack or Microsoft Teams. Budgets per company, agent and project warn at 80% and pause at the limit. Corrections become lessons in the company memory once approved, the activity log shows what agents did, and the EU AI Act use-case register records each use case. Routines run on a schedule or from a webhook. The Free plan has unlimited users and 5 agents working at once, enough for a pilot. Business, at €199 a month, allows 100 agents at once and adds single sign-on and an onboarding call.
Questions
- Two weeks is enough for one recurring process: three days to set up, a week of real work with review, and a few days to compare the results with your baseline and decide. Allow time beforehand to measure the baseline and agree the rules.
- Pick a process that happens often, has a clear definition of done, has one owner and causes little harm if a step goes wrong. Avoid decisions about individuals, such as hiring, in the first pilot.
- Compare it with a baseline measured before the pilot: quality against the same rubric, employee minutes per task including review, and cost per completed task. Also look at whether the share of work changed before approval is falling.
- For high-risk AI systems at the workplace, Article 26(7) of the EU AI Act requires employers to inform workers' representatives and affected workers beforehand. National law may require information or consultation in other cases too, so check with HR or legal advisers early.
- When an action has a clean record over several weeks, with few changes before approval, confirmed by reading a sample. Loosen one action at a time, keep reviewing the activity log, and tighten the rule again after an incident.
Sources
Related
This guide is general information, not legal advice. Check how the rules apply to your company with your own advisers.
Get started
Put your first agent to work in minutes.
Three steps, and your first agent is working for your company.
Step 1
Create an account
Sign up with your work email and invite your team. No sales call needed.
Step 2
Add your first agent
Create an agent, give it a role, and connect your own AI key and the apps it may use.
Step 3
Give it a task
Ask in Blentera or your chat app, and follow the work until it is done.