Agentic AI Methodology: how we take an AI agent to production
The Sevways Agentic AI Methodology organises the work in six blocks, from the real cases of the operation to the agent in production. The agent suggests, the person approves, and autonomy only grows when the numbers justify it.
- sends
- nothing
- answers reviewed
- all
- sends
- after approval
- answers reviewed
- all
- sends
- with a record
- answers reviewed
- only low-confidence ones
- 01Discover●
- 02Design●
- 03Evaluate●
- 04Build and protect●
- 05Pilot in shadow mode●
- 06Production and improvement◐
- 01
Real cases from the start
We start with the emails, requests and documents the operation receives every day.
- 02
Continuous evaluation
Each agent is tested with real cases before every change. Without evaluation there is no predictability.
- 03
A person decides
The agent reads the ERP on its own. Writing to the ERP or sending an email always goes through a person.
In summary
- Blocks
- Six, from Discover to Production and improvement
- Duration
- To be agreed in the assessment
- Autonomy
- Shadow mode, full approval and gradual reduction
- Logging
- What the agent read, proposed and wrote, in MAJISTI
Who it is for
Administration
Wants to know what AI will do, with what risk and who answers for each agent.
Operations management
Has teams answering the same questions from customers, suppliers and colleagues every day.
Service teams
Review and approve answers prepared by the agent instead of writing them from scratch.
IT team
Needs to connect the agents to the ERP with access rules, logging and a way to switch them off.
How an agent built with this method works
Every request goes through the same steps. Each one records what the agent read and decided.
- Customer or supplier email
- Internal request
- ERP dataread by dedicated tools, with the permissions of whoever asks
- Route
- Extract entities
- Plan
- Run tools
- Validate
- Respond
- Respond
- Clarify
- Escalate
- Record
- Proposed answera person approves or corrects it before it goes out
- Clarifying questionwhen a piece of data is missing
- Case handed to a personwhen in doubt, the agent escalates
Six principles
Real data first
We design with the real emails, requests and documents of each topic, including the difficult cases and the sector's vocabulary.
Semantic tools
The agent uses tools with business meaning and defined contracts, instead of loose queries to the ERP.
Continuous evaluation
A set of real cases per topic, regression tests and adversarial tests run before every change.
Human in the loop
The agent suggests and the person approves. Writing to the ERP or sending an email always requires a person.
Incremental delivery
First the most common path, then the difficult cases, then the other topics and the email channel.
Full observability
What the agent read, what it proposed, what it wrote and who approved is recorded.
The agent suggests, the person approves
Reading the ERP is automatic. Writing to the ERP or sending an email always needs a person. Autonomy grows step by step.
1 · Shadow mode
The agent processes real requests and sends nothing. We compare what it would have done with what the team did.
2 · Full approval
A person approves or edits every answer, and each one is classified as correct, partial or incorrect.
3 · Gradual reduction
Only low-confidence answers are reviewed. The rest goes out with a record.
Every failure becomes a test
An error found goes into the evaluation set and is tested in every following version.
When in doubt, escalate
When confidence is low or the request is out of scope, the case goes to a person with the context already gathered.
One job per agent
Each agent does one thing. We aim to cover most requests, not all of them: the rest stays with people.
Four habits we teach teams
Inspired by Anthropic's AI Fluency Framework (the 4Ds: Delegation, Description, Discernment, Diligence). They apply to anyone using an AI assistant and to anyone approving an agent's work.
Delegate
Decide what to give the AI and what to keep. Example: the AI summarises the contract; the decision to sign stays with you.
Describe
Explain the task, the context and the expected result clearly. Example: "short reply to the customer, in English, with the delivery date confirmed in the ERP".
Evaluate
Check what the AI produced before using it. Example: confirm the amount and the date before approving the answer.
Take care
Answer for the data you use and for what you send, and say when AI was involved. Example: do not paste personal data into an assistant without usage rules.
What we measure
The same indicators from pilot to operation. The targets are agreed with each company in the assessment.
Correct intent
Whether the agent understood what the request wants.
Extracted data
Whether it found the right customer, order, shipment or date in the request.
Tools
Whether the ERP queries ran and returned what they should.
Correct answers
How many answers the person approved without changes.
Handover to a person
Whether the agent escalated the cases it should, and only those.
Response time
From the request received to the answer ready for approval.
Cost per task
The model and service consumption of each request handled.
Human review
The share of answers that still go through a person, and how it changes.
Six blocks, from the first workshop to operation
- 1 · Discoverworkshops with the operators, each topic treated as a capability, indicators, operating model and security policy; we follow the real work, read the emails of each topic and map the intents and where each piece of data sits in the ERP
- 2 · Designsemantic tools with MCP contracts, limits and a confidence policy (respond, clarify or escalate); agent architecture, MCP server and a record of everything it does
- 3 · Evaluateset of real cases per topic, automatic evaluation by a second model, adversarial and regression tests; runs before every change
- 4 · Build and protectstep-by-step development; role-based access, masked personal data, protection against instructions hidden in requests, automatic cut-offs and handover to a person when something fails
- 5 · Pilot in shadow modethe agent works on real requests without sending anything; at the end, the decision to go ahead or not is the company's
- 6 · Production and improvementgo-live and team training; then weekly, monthly and quarterly improvement cycles
The duration of each block is agreed in the assessment, depending on the number of topics and the state of the data.
The agent suggests, the person approves
Writing to the ERP or sending an email always goes through a person. Review only decreases when real cases show the agent gets it right, and every failure becomes a permanent test.
Connected to
Governance
Who decides during the project?
A steering committee and a project committee, with a weekly status update and a monthly report. Moving from pilot to production is the company's decision.
How does a use of AI move from personal to critical?
There are three levels of use: personal, team and critical. Each move up a level has its own rules for approval, recording and evaluation.
Who looks after AI in each area?
Each area has a lead who knows the process, collects the difficult cases and helps colleagues use the agents.
How do you know what the agent did?
What the agent read and wrote, what it proposed and who approved is recorded. That record lives in MAJISTI, the Sevways AI control tower, and feeds the monthly review of Agent operations.
Want to apply the methodology to one of your processes?
We start with an assessment: we choose the topics, look at the data and agree the blocks and durations.
Got a project? Let us talk.
Answer within 24 working hours.