Most AI projects fail at deployment, not at demo. theagency47's 5-phase methodology (Discover, Design, Build, Deploy, Optimize) exists to push the failure-prone moments forward, into the cheap-to-fix phases, so the expensive phases (build, deploy) carry only solved problems.
theagency47 deploys AI agents through five phases, Discover, Design, Build, Deploy and Optimize, each with explicit gate criteria. Discovery produces an agent shortlist, Design a signed one-page spec per agent, Build an agent that passes its eval suite, Deploy a period of human-in-the-loop monitoring, and Optimize weekly KPI tuning until performance is stable and the agent is handed over.
theagency47 · Updated November 2026These phases run inside every project. For company-wide programmes, see the AI transformation roadmap and how we run AI transformation for mid-size and large companies.
Goal: Pick the right agents to build.
We walk through your processes, what you do, where the unbillable hours live, where errors cluster. Output is a shortlist of candidate agents, each scored on three dimensions: hours reclaimed, error rate reduction, throughput gain. We do not build agents that fail any of the three. You can simulate payback period for the "hours reclaimed" dimension before the call.
Gate: Client sign-off on the shortlist.
Goal: Make the work specifiable.
A one-page spec sheet per agent: job-to-be-done, inputs, outputs, tools, autonomous decision boundaries, escalation rules, KPIs, eval cases. This is the single most under-rated phase in agent work. Specs prevent scope creep, anchor refunds, and force "vibes" into "verifiables."
Gate: Signed spec sheet per agent.
Goal: Make the agent pass evals.
Two-week sprint per agent. Daily progress updates. Eval suite of 20–30 test cases built alongside the agent itself, not after. Common-case, edge-case, adversarial and regression inputs. Agents must pass 90%+ of our internal eval (how we test) before they leave this phase.
Gate: must pass 90%+ of our internal eval (how we test) before launch; client reviews output samples.
Goal: Run safely in production.
Production cutover with human-in-the-loop monitoring for 14 days. Every agent action is reviewable by a human before downstream effects, through approvals in OS47. Team training session delivered. Monitoring and alerting wired before traffic flips.
Gate: high pass rate on real tasks during the 14-day review, with human oversight.
Goal: Hand over a stable system.
KPI dashboard tuned. Weekly optimization until performance is stable across two consecutive reporting cycles. Agent moves from human-in-loop to autonomous-with-sampling. Documentation finalized. Handover ceremony.
Gate: Two consecutive stable reporting cycles. Client signs handover.
Once handover closes, clients can opt into a monthly retainer for ongoing optimization, health checks, and new agent builds.
The default AI agency pattern is "let's try a prompt and see what happens." This produces demos but not deployments. We refuse to write a prompt until a one-page spec is signed. This shifts the riskiest decision (what is this agent actually for?) into the cheapest-to-revise phase.
"We'll improve it in production" is how AI projects fail. We build the eval suite alongside the agent and gate launch on pass rate. The eval suite then re-runs automatically after every change in production. If an update breaks an eval, the update reverts. This is what "production-grade" actually means.
The default agency pattern is "we'll deploy and then bill you forever for maintenance." We deploy, optimize until stable across two consecutive cycles, then hand over to your team with the source files. Clients can then opt into a retainer, but it is a choice, not a captivity. Full ownership transfers to you upon final payment at handover.
theagency47 standardizes on a small, opinionated stack so we ship faster and clients can take ownership without a steep learning curve. For the broader frame of how this methodology maps to the three levels of AI in a business, and to why we built theagency47 as a delivery agency rather than a consultancy, see those pages.
The full stack is documented in the engagement contract. Clients receive the full source (prompts, configs, KB schemas, eval cases, integration code) upon final payment at handover.
Methodology means what we exclude as much as what we include. Five things theagency47 does not do:
Spark = 14 days · Workforce Starter = 30 days · Workforce Pro = 60 days · Scale package = 90+ days. Each phase is gated. See pricing for tier details.
One page, eight fields: job-to-be-done, inputs, outputs, tools available, autonomous decision boundaries, escalation rules, KPIs, eval cases. Client signs before any building.
20–30 test cases per agent with known expected outputs. Split common-case (60%), edge-case (20%), adversarial (15%) and regression (5%, growing as production fixes are added). Must pass 90%+ of our internal eval (how we test) before go-live. Re-runs automatically after any production change.
For 14 days after deployment, every agent action is reviewed by a human before downstream effects. After 14 days, with a high pass rate on real tasks during the 14-day review, the agent moves to autonomous mode with sampling-based audit.
Automatic revert to the last passing version. The change that broke the eval is investigated. Production fixes go through the same eval gate as initial deployment.
Same five phases, smaller scope. New agents added under a retainer follow the full Discover → Optimize cycle, just compressed (typically 5–10 days for a task-tier agent under retainer).
Next: See our services →
Talk to our AI agent
Ask about our services, prices or how a project runs, in Greek or English. It answers out loud.
Uses your microphone. Ends automatically after 3 minutes.
Test version; the AI can make mistakes. Calls are recorded and the transcript is sent to our team; audio is processed by xAI. Privacy policy