If your managed network provider says its AI agents will handle routine tasks and configuration work, the contract should say exactly which actions the agents may take on their own, which need your approval, how every change is recorded, how it is reversed, and who pays when an agent gets it wrong. A sales deck is not a commitment. Only the services agreement, the statement of work and the referenced runbooks are.

This article is for IT, operations, finance and procurement teams comparing managed network, managed IT or SD-WAN providers. It focuses on contract and provider-selection terms, not your internal AI governance program.

Key takeaway: Do not ask whether a provider uses AI. Ask what the agents are permitted to change, what evidence they leave behind, and what the provider commits to when a change goes wrong. Get the answers in writing.

Why AI agents change the managed services conversation

Vendors are packaging agentic systems for network operations. A recent Network World listing describes a system that lets network teams automate routine tasks, manage configurations and control how much work agents perform independently. A TechTarget item makes a related point: as AI agents move from answering questions to acting, businesses face a growing challenge of keeping accountability in step with the pace of change.

That second point matters for buyers. Traditional managed services agreements assume a named engineer made a change under a documented process. When an agent makes the change, the old language can leave gaps about who authorized it, what evidence exists and who is responsible.

The feature sets are still evolving, so verify current capabilities directly with each provider rather than relying on press summaries.

Scope: define what the agents may touch

Start with a plain-language definition of agent activity in the agreement. Cover the following.

• Systems in scope: routers, switches, firewalls, SD-WAN edges, wireless controllers, DNS and DHCP, and cloud network constructs. Name them or reference an asset list.
• Systems out of scope: anything the agent must never reach, such as core firewalls, payment-network segments or healthcare device VLANs.
• Data the agents can access: configuration files, logs, traffic metadata and credentials. State whether that data is used to train or improve models, and whether it is shared with a third-party model provider.
• Subprocessors: which AI or tooling vendors sit behind the provider’s agents, and how you will be told when they change.

Tier the actions

A simple tier model keeps the negotiation concrete. This is an original framework you can adapt, not an industry standard.

Adjust the tiers for your environment. A multi-site restaurant group and a hospital will draw the Tier 1 and Tier 2 lines in different places.

Approval: who says yes, and how fast

Approval terms often fail in practice because nobody is named or the process is too slow to use. Specify:

1. Named approvers and backups on your side, with contact methods and escalation order.
2. Approval evidence: the approval must be recorded against the specific change, not given in a general email thread.
3. Emergency authority: whether the provider may act at Tier 2 during a declared outage without approval, and what post-incident review follows.
4. Change windows and freezes: agents should respect blackout periods such as retail peak season or hotel check-in hours.
5. Ability to tighten: the right to move any action to a stricter tier on written notice, without a fee or a re-papered contract.
6. Notice of expansion: the provider cannot widen agent autonomy, or add new agent capabilities, without advance written notice and your consent for Tier 2 and above.

The last point is easy to miss. Agent features can arrive as product updates, so a contract signed today can quietly cover a different level of automation next year.

Logging and rollback: the evidence trail

You cannot manage what you cannot reconstruct. Require that every agent action produce a record containing:

• the action taken, the target device and a timestamp
• the identity the agent acted under, distinct from human engineers
• the trigger, such as an alert, a ticket or a scheduled task
• the approver, if one was required
• the before-and-after configuration, or a pointer to it
• the outcome and any validation checks run

Then address access and retention. Can you export these logs on demand, in a usable format, without a professional services fee? How long are they kept, and do you keep them after termination? If you have compliance obligations, ask your counsel or auditor what retention you need rather than guessing.

Rollback terms

Rollback should be a contract commitment, not a feature claim. Ask the provider to commit to:

• a pre-change backup of configuration for every Tier 1 and above change
• a target time to revert a bad change, measured from detection or from your request
• a human on call with authority to reverse agent actions at any hour your service covers
• an emergency stop: the right to suspend all agent activity on your environment immediately, with confirmation
• a post-incident report within a stated period for any agent-involved outage

Ask to see how rollback has been tested. A demonstration in a lab or a reference customer conversation tells you more than a policy statement.

Liability and service levels for agent-caused outages

This is where finance and procurement should be present. Many managed services agreements cap liability and limit remedies to service credits. Have counsel review how those clauses apply to an outage caused by an agent, and check these points.

• No carve-out for automation: the agreement should treat an agent’s actions as the provider’s own acts, not as a tool failure outside its control.
• SLA treatment: do outages caused by provider changes count against uptime commitments, or are they excluded as planned or maintenance activity?
• Credit versus damages: if service credits are the sole remedy, consider whether they come close to your real cost of downtime.
• Third-party models: if the provider blames an upstream AI vendor, confirm that does not reduce its obligations to you.
• Insurance: ask what coverage the provider carries and whether it addresses automated actions. Request evidence rather than a verbal assurance.
• Security incidents: if an agent’s credentials are misused, confirm how notification, investigation and cost are handled. Your cybersecurity requirements should align with this clause.

This is general guidance, not legal advice. Your counsel should decide what protection is realistic for your deal size.

A common scorecard for comparing providers

When three providers all say they use AI, comparison gets muddy. Send each the same questions and score the written answers from 0 (no answer or marketing language only) to 3 (contractual commitment with evidence).

1. Which action types can agents perform without approval today? Provide the list.
2. Can we restrict agents to Tier 0 and Tier 1 only?
3. How are agent identities separated from human engineer identities?
4. What does the change log contain, and can we export it?
5. What is the committed time to roll back an agent change?
6. Can we trigger an emergency stop, and how is it confirmed?
7. Is customer data used to train models or shared with third parties?
8. How will we be notified of new agent capabilities, and can we decline them?
9. How are agent-caused outages treated for SLA credits and liability?
10. Can we opt out of agent activity entirely at renewal or mid-term?

Weight the questions by your risk. A business with a few sites and a lean IT team may value rollback and emergency-stop terms most. A regulated business may weight logging and data use more heavily.

Hypothetical example

This example is illustrative, not a real customer. A 40-location retailer receives two SD-WAN managed service proposals. Provider A promises autonomous optimization across all sites. Provider B limits agents to Tier 0 and Tier 1 by default, requires approval for routing and security policy changes, and commits in writing to a revert target. Provider A may be cheaper and faster on paper. But if the retailer cannot get Provider A to define limits, logging or liability, Provider B’s narrower scope may be the lower-risk choice until Provider A closes those gaps.

Who each approach suits

• Broad agent autonomy suits organizations with strong in-house network skills that can audit changes, simple environments and a tolerance for brief disruptions.
• Approval-gated agents suit most multi-site businesses, since they get faster ticket handling and diagnostics while a person still authorizes meaningful changes.
• Agents limited to monitoring and ticketing suit regulated, safety-sensitive or revenue-critical environments, or buyers who want to build trust before expanding scope.

Any of these can work. The risk comes from not knowing which one you have bought.

Pre-signature checklist

• Agent activity is defined in the agreement or SOW, not just in marketing material.
• Actions are tiered, and Tier 2 and above require named human approval.
• Out-of-scope systems are listed.
• Agent actions are logged, exportable and retained for a stated period.
• Rollback time, backups and an emergency stop are committed in writing.
• Provider cannot expand agent autonomy without notice and your consent.
• Data use, model training and subprocessors are disclosed.
• Agent-caused outages are covered by SLA and liability language your counsel has reviewed.
• You can restrict or opt out of agent activity without penalty.
• Every competing provider answered the same scorecard.

Next step

Before you sign or renew, put your current or proposed agreement beside this checklist and mark every item that is missing or vague. If you want a vendor-neutral comparison of providers on equal terms, you can talk with MoJo Tech Group about managed IT and network services.

Sources

• Network World: headline describing an agentic AI system for network teams to automate routine tasks, manage configurations and control agent independence.
• TechTarget: headline on keeping accountability in step as AI agents move from answering questions to acting.