The industry average for a production AI agent project is roughly 7 months from kickoff to launch, according to a 2026 Deloitte survey of 340 enterprise AI initiatives. We ship ours in six weeks. Not a demo. Not a proof of concept. A monitored, tested, integrated AI agent doing real work against real production data. After more than 1,000 projects, we've learned that the six-week constraint isn't a marketing line. It's the operating model that keeps AI projects from drifting into the swamp where most of them die. Here is exactly how we run it.

Why Six Weeks Is the Right Constraint

Most AI projects don't fail because the technology is too hard. They fail because the timeline is too long. A project that takes 7 months touches three quarterly planning cycles, four executive priority shifts, and at least one reorg. By the time it ships, the workflow it was supposed to automate has changed, the team that asked for it has moved on, and the budget has new owners with new questions.

Six weeks is short enough to ship before the political weather changes. It's also long enough to do the work properly if you treat every week as a hard constraint, not a soft target. The shape of the sprint matters more than the number of weeks. Each phase has a single named owner, a single deliverable, and a single decision that has to be made before the next phase starts.

We don't compress timelines by skipping steps. We compress them by removing the gaps between steps where most agencies and internal teams quietly leak weeks.

Week 1: Discovery and Scoping

Week one is not about building anything. It's about deciding what we will and will not build. We run a structured discovery process that we've refined across hundreds of AI automation projects. The week ends with a one-page scope document that the client's executive sponsor signs off on before any code is written.

The week breaks down into four blocks. Days one and two are workflow mapping. We sit with the people doing the work today and document every step, every system, every exception. Days three and four are constraint discovery. We pin down the data sources, access requirements, security review process, and integration surface area. Day five is scope lock. We pick one workflow, one set of success metrics, and one definition of done.

The most common mistake here is scoping the version of the workflow people wish existed instead of the one that actually runs. Real workflows have exceptions, manual overrides, and undocumented rules. If the scope assumes those don't exist, the agent ships and immediately breaks on real input.

Week 2: Architecture and Prototype

By the end of week two, we have a working prototype that hits the real data. Not synthetic data. Not a sanitized export. The actual production data the agent will operate on once it ships.

The first two days are architecture decisions. Which model. Which orchestration framework. Which tool integrations. Which retrieval pattern, if any. We document the choices in a one-page architecture brief that names every dependency and every failure mode we expect.

Days three through five are the prototype build. The prototype only has to do one thing well: prove that the riskiest part of the workflow is solvable. If the riskiest part is data extraction from messy PDFs, we build that piece first. If it's tool calling against a finicky internal API, we build that. We do not build the easy parts in week two. The easy parts are easy in week four.

We end week two with a 30-minute demo to the client's technical lead. If the prototype doesn't survive that demo, we replan the project rather than push forward with a known weakness. That has happened on roughly 8% of projects. The replan always saves the engagement.

Weeks 3 and 4: Production Build

These are the heads-down weeks. The prototype expands into the full agent. We wire up the remaining tools, build the orchestration logic, add the retrieval layer if the architecture calls for one, and connect to every system the agent needs to touch in production.

We run a daily 15-minute standup with the client's named point of contact. The standup has three questions. What shipped yesterday. What ships today. What is blocked. The point of contact is not a project manager. It is the person on the client side who owns the workflow being automated, because that is the person who can answer scope questions in minutes instead of days.

The single biggest accelerator in this phase is parallel work on logging and observability. Most teams add monitoring at the end. We instrument every tool call, every model invocation, and every decision point from day one of the build. That instrumentation is what makes week five possible. We covered the full approach in our post on monitoring AI agents in production.

Week 5: Testing and Refinement

Week five is when most internal AI projects discover they need another month. We don't, because the test plan was written in week one and the agent has been instrumented since week three. We run our four-layer test process: unit tests on individual tool calls, integration tests on tool chains, scenario tests on real workflow examples, and adversarial tests on edge cases and known failure modes.

The testing isn't done by us alone. The client team runs the agent against their own backlog of real examples, including the messy ones they didn't share during discovery. Almost every project surfaces at least one workflow variant that nobody mentioned in week one. That is not a failure. It is the test catching exactly what it is supposed to catch.

We fix the gaps, retest, and lock the agent's behavior. By the end of week five, the agent is passing the test suite at the threshold defined in the scope document. If it isn't, week six does not start. We extend the project rather than ship something that will quietly degrade in production.

Week 6: Deployment and Handoff

Week six is two things in parallel. Production deployment and team handoff.

Deployment is staged. Day one of week six, the agent goes live on a narrow slice of real traffic, usually 10% of the workflow volume. Days two and three, we expand to 50% while monitoring drift, error rates, and the metrics defined in the scope. By day four, the agent is handling 100% of the workflow with humans reviewing flagged exceptions.

The handoff is the part most agencies skip. We document the agent's prompts, tool schemas, decision logic, and known limitations. We train two named people on the client side to triage alerts, roll back to the previous version if needed, and request future changes. The team that owns the agent post-launch should be able to keep it healthy without us. An agent that only one external team can maintain is a future hostage situation, not a deliverable.

What Lets Us Move This Fast

The six-week timeline works because of a small number of operating choices we made early and never relaxed.

We run one project per pod per sprint, so context switching is zero. The pod is the same three people start to finish: an AI engineer, a workflow lead, and a designer who owns the human-facing surfaces. We build on a shared internal framework that handles the parts every project needs the same way, so 100% of the build effort goes to the parts that are actually client-specific.

We also say no to projects that don't fit the constraint. Workflows that require novel research, brand-new model fine-tuning, or rebuilding an existing system from scratch don't belong in a six-week sprint. They get a longer engagement or we point the client to a different approach. The discipline of refusing the wrong projects is what keeps us shipping the right ones on time.

Where to Go From Here

If you're sitting on a workflow that's been "in scoping" for two quarters, the problem usually isn't the workflow. It's the process. A well-scoped AI agent project should leave discovery in days, not months. If you want to see what a six-week sprint would look like against one of your workflows, start a project with us and we will run the discovery the same way we just described.

The teams winning with AI in 2026 aren't the ones running the biggest budgets. They're the ones shipping production systems before the meeting that scoped them gets a follow-up.