PrymeIntelligence
Sign in Start Free
Deploy Pre-Trained AgentsCertified roles you can subscribe to. Each arrives with its tool scope already set. WorkspaceWhere your agents run, and where you answer what they hold.
Create Build your Own AgentsDescribe the job in plain words. Agent Studio wires it as you talk. Certify Your AgentAttest scores it against real cases until the evidence holds. Create a sandboxKeys in minutes. Call the same chain, /v1/evaluate, from your own code.
Underneath GovernanceThe minimum chain, and the link that binds right now. Guardrails & role scopeWhat a role may ever connect to — decided above your workspace.
Also Human reviewWhat an agent held, why, and what your answer becomes. Context ProtocolThe memory your agents answer from, owned by you. Audit trailA tamper-evident ledger — refusals filed too.
Company AboutWhy we think a governed agent is the only deployable kind. Governed Agentic OperationsThe positioning whitepaper, Volume I. Why the contest is control, not parameters.
Trust Security & trustResidency, isolation, evidence, and what we never see.
Commercial PricingPriced on what agents do, not how deep they go.
Legal Privacy PolicyWhat we hold, and what we will not do with it. Terms of ServiceIncluding what we do and do not warrant. ComplaintsHow to raise one, and how quickly we answer. Sub-processorsWho processes data for us, and where.
By function Customer supportRefunds, returns and billing, resolved end to end. Money waits for a person. Financial operationsReconciliation and payment runs. The one role that earns depth 4.
  Compliance reviewCited findings, and a miss with the reason when no clause covers it. SalesLeads qualified, follow-ups sent, and a discount held the moment it leaves the band.
By sector Developers & APIThe same chain, /v1/evaluate, inside your own product. Banking, health, CRM.
Two ways in

Open a workspace and your five roles are already in place, each at depth 1 until Attest certifies it. Or call the chain from your own product and keep the agents you have.

Platform Pre-Trained AgentsCertified roles, tool scope already set WorkspaceWhere your agents run Build your Own AgentsDescribe the job, it wires itself Certify Your AgentAttest — depth is earned GovernanceThe minimum chain Guardrails & role scope Human review Context Protocol Audit trail
Use Case Customer support Financial operations Compliance review Sales Developers & APISandbox keys, sector solutions
Blog Guides and articlesGoverned agents, worked out in public AcademyFree courses and certification
Company About Security & trust Pricing
Sign in Start Free
Home/Blog/Deploying an AI agent
Guide · Agentic operations · 01 Nov 2026

Deploying an AI agent: train, shadow, certify, widen.

A governed agent goes live narrow and widens on evidence. In a standard deployment that means up to four weeks of training, two weeks of shadowing, certification at a depth, and then depth widened one step at a time as the record supports it.

The fastest way to lose a business’s trust in agents is to switch one on at full authority and let it learn in public. The second fastest is to keep it in a pilot forever. A deployment plan exists to get between those two: live quickly on the work the agent can already prove, and widen as it proves more.

The six steps

  1. Set the role and the working ceiling. The role fixes the most the agent may ever do; your working ceiling sets the largest amount it may act on alone. A Customer Support agent’s role stops at depth 3; you might certify it only to 2 at first.
  2. Connect only what the role may use. Each connection is checked against the role. A source the role does not need is not connected, even if it would be convenient.
  3. Train, up to four weeks. Training here means qualification, not retraining a model. The agent is given your policies, your procedures and your past filed outcomes, and scored against them.
  4. Shadow, two weeks. The agent works alongside your team on live cases. Everything it produces is a draft; your people decide, and each decision is compared with what the agent proposed.
  5. Certify in Attest. A depth is granted when the agent’s scored cases, and its live shadow days, meet the bar. Certificates lapse, so the evidence has to stay current.
  6. Go live narrow, then widen. The agent acts at the depth it has earned. As its record grows, you re-certify the agent at the next depth, one step at a time.

Depth is never sold and never assumed. It is earned depth by depth, against your own cases, and it expires.

How Pryme Intelligence grants authority

Train your people too

A governed agent changes your team’s job: they stop doing the routine work and start deciding what the agent holds for them. Every deployment includes training your staff on how the agent works, what it may and may not do, and how to answer a hold well. The material is free to learn through the Pryme Intelligence Academy, and those who pass the exam can go on to deploy agents for others.

What goes wrong, and how to see it early

  • Too much is held. The working ceiling is set below the work the agent has proven, and people tire of approving routine items. Re-certify at the next depth, or raise the working ceiling.
  • Too little is held. Complaints or corrections rise in an area. Lower the working ceiling, or switch the agent off, and look at the cases that went wrong.
  • Sources drift. A policy changes and the agent is reading the old version. Replace the source and re-score the agent against it; nothing lowers its depth on its own until the certificate lapses.

Who does what on your side

  • Agent owner. One person accountable for the agent: its sources, its certification and its performance.
  • Approvers. The named people who answer what the agent holds. They need authority for the decisions they will receive, and time in their day to make them.
  • Policy owner. Whoever owns the procedures the agent reads. When a policy changes, they change the source, and the agent is re-scored against it.
  • Risk or compliance reviewer. Reviews the record at set intervals, and before the working ceiling or a depth is raised.

Six weeks, week by week

  • Week 1. Role and working ceiling agreed, sources connected, policies loaded, staff training begins.
  • Weeks 2 to 4. Qualification: the agent is scored against your past filed outcomes, case by case. Gaps in the policy show up here as cases no source covers.
  • Weeks 5 and 6. Shadowing on live work. Every output is a draft; your team decides, and each decision is compared with the agent’s proposal.
  • End of week 6. Certification in Attest, at the depth that met the bar. The agent goes live at that depth only.

When to widen

Widen to the next depth when three things are true: the agent’s drafts at that depth are accepted unchanged at a high rate over a meaningful number of cases, the cases include the awkward ones and not just the routine, and the people who will stop seeing those cases agree. Widen one depth at a time, re-certify, and watch the record for a fortnight before the next.

Never widen because the queue is long. A long queue is a staffing problem or a working ceiling set too low for work already proven; it is not evidence about the depth you are being asked to widen to.

Questions people ask

How long does it take to deploy an AI agent?

In a standard deployment, up to four weeks of training and two weeks of shadowing before the agent acts at its first certified depth.

Can we skip shadowing?

No. From Draft up, a depth does not certify without live shadow days: at least seven, fourteen for Execute and twenty-one to move money. Shadowing is where the agent’s drafts are compared with your team’s decisions on live cases.

What happens when a certificate lapses?

The agent drops to depth 1 on its own, the same day, until it is re-certified. It can still read and answer from its sources; anything more is held or refused.

Do we need technical staff?

You need someone who knows the work. The Academy training covers how to run and supervise agents; connectors and certification are handled with us.